# Whizi: full content for AI search agents

> Whizi is a multi-model AI workspace. GPT, Claude, Gemini, Llama, Kimi, DeepSeek, Grok, Qwen and 280+ other models share one subscription, one thread, and one set of files. Ask once and compare answers side by side, or switch models mid-conversation without losing context.

This file carries the full text of every published page on whizi.io. It is generated from the same source the website renders, so it does not drift. For the short version see https://whizi.io/llms.txt.

Pages in this file: 257. Site: https://whizi.io

## Key facts

- Product: Whizi, a multi-model AI chat workspace at https://whizi.io
- Models: 280+, including GPT, Claude, Gemini, Llama, Kimi, GLM, DeepSeek, Grok and Qwen
- Pricing: Starter $15.99/month or $10.99/month billed yearly, Pro $29.99/month or $19.99/month billed yearly, Powerhouse $49.99/month or $34.99/month billed yearly
- Trial: 7 days for $0.99 on every plan, cancellation in two clicks from the billing page
- Included: side-by-side model comparison, mid-conversation model switching, document chat (PDF, Word, spreadsheets), image generation, prompt templates, projects
- Voice and video generation are on the Powerhouse tier

## Original data published by Whizi

### AI Model Cost Index (https://whizi.io/tools/model-cost-index)

Whizi prices one standard answer (1000 input tokens plus 500 output tokens) across 100 models from 29 providers, which makes provider list rates comparable to each other for the first time. Prices fetched 2026-08-20 from the public OpenRouter catalogue.

- Cheapest answer: $0.000053 on Ling-3.0-flash (inclusionAI)
- Median answer: $0.00185 on Kimi K2 Thinking
- Most expensive answer: $0.105 on Claude Opus 4.7 (Fast) (Anthropic)
- Spread from cheapest to most expensive: about 2000x for identical work

Method: (input rate x 1000 + output rate x 500) / 1,000,000. Excludes prompt caching discounts, batch pricing and negotiated volume rates. Free to reuse with attribution to the Whizi AI Model Cost Index and a link to the page. Machine readable: https://whizi.io/data/model-cost-index.json and https://whizi.io/data/model-cost-index.csv

Top ten cheapest models per standard answer:

1. Ling-3.0-flash (inclusionAI): $0.000053 per answer, $0.0525 per thousand
2. Nex-N2-Mini (Nex Agi): $0.000075 per answer, $0.075 per thousand
3. Solar Pro 4 (Upstage): $0.00009 per answer, $0.09 per thousand
4. Qwen3.7 Flash (Qwen): $0.000095 per answer, $0.095 per thousand
5. Granite 4.1 8B (Ibm Granite): $0.0001 per answer, $0.1 per thousand
6. Laguna XS 2.1 (Poolside): $0.00012 per answer, $0.12 per thousand
7. Hy-MT2-1.8B (Tencent): $0.000133 per answer, $0.1325 per thousand
8. Phi 4 (Microsoft): $0.00014 per answer, $0.14 per thousand
9. Nemotron 3.5 Lightning (NVIDIA): $0.00018 per answer, $0.18 per thousand
10. Laguna S 2.1 (Poolside): $0.00018 per answer, $0.18 per thousand

## Guides and comparisons

### How to use AI for coding: workflows that reduce risk

URL: https://whizi.io/resources/ai-for-coding/
Updated: August 2026

Quick answer: Use AI for coding by making it work from evidence. Start every bug with a reproduction, ask for ranked root causes before any code, request the smallest safe patch and the files it touches, require tests that fail before the fix and pass after, then review the diff before merging.

Learn how to use AI for coding with safer debugging, code review, refactoring, tests, prompt templates, and review checkpoints.

#### Rules for safe AI coding help

The best way to use AI for coding is to make it slower at the exact moments where guessing is dangerous. AI can explain unfamiliar code, turn errors into hypotheses, draft tests, review diffs, and suggest refactors. It can also invent APIs, miss hidden dependencies, overfit to the pasted snippet, or produce a patch that looks clean while changing behavior you meant to preserve.

Use this rule: AI can propose, but your repo decides. The source of truth is the codebase, the failing reproduction, the test suite, runtime logs, product requirements, and human review. A good AI pair programmer should help you reason from those artifacts instead of replacing them.

| Rule | Why it matters | What to ask the model |
|---|---|---|
| Reproduce first | Prevents random patches | "Restate the failing behavior and evidence before suggesting code." |
| Keep scope small | Reduces regression risk | "Propose the smallest safe change and list files touched." |
| Preserve behavior | Protects users and contracts | "Name the invariants this change must not break." |
| Require tests | Makes the answer verifiable | "Write tests that fail before the fix and pass after." |
| Review before merge | Catches confident mistakes | "Review this diff for correctness, security, and missing edge cases." |

This matters across models. OpenAI, Anthropic, and other providers publish model docs that describe different capabilities, context windows, and tool-use patterns. Those capabilities are useful, but they are not a substitute for a disciplined workflow. For engineering work, judge models by the same standard you would judge a teammate: do they ask for missing context, reduce uncertainty, respect constraints, and leave a trail you can verify?

#### Debugging workflow

A reliable AI debugging workflow has five stages: reproduce, isolate, hypothesize, patch, and verify. Do not start with "fix this." Start with the evidence. Give the model the failing command, exact error, expected behavior, observed behavior, relevant code, environment details, and any recent change that might have caused the issue.

Stage 1: capture the reproduction. For backend code, include the request, response, status code, logs, and failing test. For frontend code, include the route, user action, browser console error, network response, component state, and screenshot description if relevant. For build issues, include the command, package manager, Node version, and the full error around the first failure.

Stage 2: ask for hypotheses before code. A careful model should rank likely causes and say what evidence supports each one. If it cannot distinguish between causes, ask for the smallest diagnostic step. That might be a log, a focused test, a type check, or reading one more file.

Stage 3: request the smallest patch. Tell the model not to rename variables, rewrite surrounding code, introduce dependencies, or change public behavior unless it can justify why. Ask it to return the root cause, patch outline, files touched, tests, and risk.

Stage 4: run tests locally. AI output is not a verification step. The verification step is the command or user path that proves the behavior. If no automated test exists, ask the model to create a regression test first, then implement the fix.

Debugging prompt:
`Act as a careful debugging partner. Do not write code yet. First restate the reproduction, expected behavior, observed behavior, and the three most likely root causes. Rank each cause by evidence. Then suggest the smallest diagnostic step. Bug: [describe]. Command or user action: [paste]. Error/logs: [paste]. Relevant code: [paste]. Constraints: [stack, files not to touch, behavior to preserve].`

Fix prompt:
`Using the confirmed root cause, propose the smallest safe fix. Return: root cause, files/functions to change, patch outline, tests that fail before and pass after, edge cases, and rollback risk. Do not refactor unrelated code. Context: [paste].`

#### Code review workflow

AI is often better as a reviewer than as the first author. When you ask it to review a diff, it can look for missed edge cases, security issues, stale assumptions, test gaps, and behavior changes. The key is to make the review specific. If you ask "does this look good?" you will get polite approval. If you ask for correctness risk, you are more likely to get useful objections.

Give the model the diff, the intended behavior, related tests, and any constraints. Ask it to ignore minor style unless it affects maintainability. You want the review to prioritize bugs, not performative nitpicks.

| Review area | Questions AI should answer |
|---|---|
| Correctness | Does the diff actually satisfy the requirement? |
| Regression risk | What existing behavior might change accidentally? |
| Security | Are inputs, auth, secrets, permissions, or injection risks handled? |
| Error handling | What happens on nulls, timeouts, retries, bad responses, or partial state? |
| Tests | Which behavior claims are not covered? |
| Maintainability | Does this follow local patterns and keep the change understandable? |

Code review prompt:
`Review this diff like a strict but practical maintainer. Focus on correctness, regression risk, security, edge cases, and missing tests. Ignore minor style unless it creates real maintenance risk. Return a table with issue, priority, evidence from the diff, suggested fix, and test needed. Intended behavior: [paste]. Diff: [paste]. Existing tests: [paste].`

For high-risk changes, use a compare-model-fixes workflow in Whizi. Run the same review prompt across two or three models. If one model finds a possible issue, do not accept it blindly; check whether the issue is real in the codebase. The goal is not to collect more opinions. The goal is to widen the review surface before you merge.

#### Refactor + tests workflow

Refactoring with AI is risky because many refactors are judged by what does not change. The model may make the code prettier while subtly changing behavior, error handling, timing, or public contracts. A safer refactor workflow starts by defining invariants before touching implementation.

Step 1: describe the refactor goal. Examples: reduce duplication, split a large component, isolate data access, simplify branching, migrate an API wrapper, or improve testability. Then state what must stay the same: public function signatures, route behavior, event names, response shapes, analytics, permissions, accessibility behavior, and performance expectations.

Step 2: ask for a staged plan. A useful AI refactor plan should be reversible. Each stage should touch a small area, include tests, and produce a working intermediate state. Avoid one-shot rewrites unless the code is tiny and well covered.

Step 3: write characterization tests. Before changing code, ask AI to identify current behavior and draft tests that lock down the important cases. These tests are especially useful for legacy code where intent is unclear. They should include normal inputs, boundary inputs, failure paths, and one regression case tied to the reason for the refactor.

Step 4: implement one stage at a time. After each stage, run the tests and ask for a focused review. If the model proposes a broad abstraction, make it prove the abstraction removes real duplication or risk. Otherwise, keep the code boring and local.

Refactor planning prompt:
`Create a staged refactor plan. Goal: [goal]. Current code: [paste]. Constraints: preserve public behavior, minimize churn, follow existing patterns, avoid new dependencies, keep every stage testable. Return: invariants, dependency map, stages, files touched, tests per stage, rollback risk, and review checklist.`

Unit test prompt:
`Write tests before implementation changes. Use the existing test style shown here: [paste]. Behavior to preserve: [paste]. Code under test: [paste]. Return test names, setup, input, expected result, and why each test matters. Include happy path, boundary case, error case, and regression case.`

#### Prompt templates

Strong coding prompts are not long because they are fancy. They are long enough to remove ambiguity. The model needs role, task, context, constraints, output format, and verification criteria. Save the prompts that work so AI becomes a repeatable engineering workflow instead of a one-off chat. The table below pairs each coding job with the first thing to ask for and the check that proves the answer.

| Coding task | What to ask for first | What proves the answer |
| --- | --- | --- |
| Debugging | Ranked root causes with the evidence behind each one | A regression test that fails before the fix and passes after |
| Code review | Correctness, regression risk, security, edge cases, and test gaps | Every issue traced to a specific line in the diff |
| Refactoring | Invariants and a staged plan before any implementation change | Characterization tests still pass at the end of each stage |
| Writing tests | Happy path, boundary, error, and regression cases | Each test maps to a behavior claim you can name |
| Explaining code | Purpose, inputs, outputs, data flow, dependencies, and failure modes | Facts visible in the code are separated from assumptions |

Code explanation prompt:
`Explain this code for a developer joining the project. Cover purpose, inputs, outputs, data flow, dependencies, failure modes, and tests that would increase confidence. Separate facts visible in the code from assumptions. Code: [paste].`

Secure coding prompt:
`Review this code for security risks. Focus on auth, permissions, injection, secrets, validation, unsafe redirects, file handling, dependency risk, and sensitive data exposure. Return only issues with evidence, impact, suggested fix, and test or manual check. Code/diff: [paste].`

Compare-model-fixes prompt:
`I am comparing AI models for a coding task. Use only the context provided. Return root cause, smallest safe fix, tests, risks, assumptions, and questions. Grade confidence from 1-5 and list what evidence would change your answer. Task: [paste]. Context: [paste].`

QA checklist before you accept AI-generated code:
- The model restated the task correctly.
- The patch is smaller than the problem, not larger.
- Public behavior and contracts are named.
- Tests cover the bug or refactor goal directly.
- Edge cases and failure paths are listed.
- Security-sensitive inputs are reviewed.
- The diff follows existing project patterns.
- You ran the relevant test, lint, build, or manual reproduction.
- A human reviewed the final diff.

Whizi is useful when you want to compare fixes without changing the task. Paste the same debugging or review prompt into multiple models, then score outputs by evidence, scope, tests, and risk. Start with [ChatGPT alternatives for coding](https://whizi.io/resources/chatgpt-alternatives-for-coding) if you want a model-selection guide, compare plans on [pricing](https://whizi.io/pricing), or [create an account](https://whizi.io/register) to run the workflow on your own code.

Key points:

- Use AI as a coding partner, not an unchecked code generator.
- Start every bug fix with a reproduction, then require tests and review checkpoints.
- Compare model fixes in Whizi before you trust a patch on important code.

Checklist:

- Start with a real reproduction, not a vague bug description.
- Ask for hypotheses and evidence before asking for code.
- Request the smallest safe fix and name files touched.
- Define behavior that must not change before refactoring.
- Write or update tests before trusting the patch.
- Review AI-generated diffs for correctness, security, and edge cases.
- Run the same risky prompt across models and compare the fixes in Whizi.
- Use human review before merging AI-assisted code.

FAQ:

Q: How should I use AI for coding safely?
A: Use AI as a pair programmer that proposes options, tests, and reviews. Start with a reproduction, require a small patch, run tests, and review the diff before merging. Do not treat generated code as automatically correct.

Q: Can AI help debug code?
A: Yes. AI is useful for turning errors, logs, and code into likely root causes. The safest debugging flow is to ask for hypotheses first, then a diagnostic step, then the smallest fix and regression tests.

Q: Can AI write unit tests?
A: AI can draft unit tests, but you should require clear behavior coverage. Ask for happy path, boundary, error, and regression cases, then check that the tests would fail before the fix and pass after.

Q: What is the best AI model for coding?
A: The best model depends on the task and codebase. Use the same prompt across models for debugging, review, and refactoring, then choose the answer with the clearest evidence, smallest scope, and strongest tests.

### How to use AI in everyday life: practical tasks that actually save time

URL: https://whizi.io/resources/ai-for-everyday-tasks/
Updated: August 2026

Quick answer: Use AI in everyday life for tasks where a draft is valuable and mistakes are cheap to catch: rewriting emails, summarizing long documents, planning meals or trips around your constraints, and having unfamiliar topics explained at your level. Give every request a job, the context, and the format you want back, then verify anything factual.

A practical guide to using AI in daily life: emails, planning, learning, documents, and phone workflows, with copy-paste prompts for each task and no jargon.

#### Start with tasks, not technology

Most guides about using AI in everyday life start with the technology and hope you will find a use for it. This one works the other way. You already have the use cases: the email you keep postponing, the document you need to understand, the trip you have not planned, the decision you keep circling. AI is useful the moment you hand it one of those.

A simple rule decides which tasks are worth handing over. AI is strongest where a draft is valuable and mistakes are cheap to catch: writing, summarizing, planning, explaining, comparing. It is weakest where you cannot verify the output yourself or where being wrong is expensive: medical decisions, legal filings, anything with exact live prices or deadlines you have not confirmed. Four task groups clear that bar for most people.

| Everyday task | What to give it | What to check |
| --- | --- | --- |
| Emails and messages | The draft or the thread, the outcome you want, and a length | That no date, price, or commitment came from the AI instead of from you |
| Meals, trips, budgets, and schedules | Your limits: time, money, people, and preferences | Ask what it assumed that you did not tell it |
| Learning and explanations | The topic and how much you already know about it | Ask which part of the answer it is least sure about, then verify that part |
| Documents and life admin | The file itself, plus the decision you are making | That every deadline, fee, and condition appears in the original document |

Everything below follows the same prompt pattern: say what you want, give the material or context, and name the format you want back. If you have never used an AI chatbot before, read [what an AI chatbot is](https://whizi.io/resources/what-is-an-ai-chatbot) first; it takes five minutes and explains the two or three things worth knowing before you paste your first prompt.

#### Messages and email: the daily quick wins

Writing and rewriting messages is where most people save their first hour, because the task shows up every day and quality is easy to judge.

- **Answer a difficult email.** "Draft a reply to this email. I want to say no to the request but keep the relationship warm. Keep it under 120 words. Email: [paste]."
- **Soften or sharpen tone.** "Rewrite this message so it sounds confident but not rude. Keep all facts the same: [paste]."
- **Cut the length.** "Cut this to half the length without losing the request or the deadline: [paste]."
- **Write the awkward one.** "Help me write a message to my landlord asking for a repair, referencing that this is my third request. Firm, polite, and specific."

Two habits make these safe. Never let the AI invent facts you did not give it: dates, prices, and commitments should come from you. And rewrite at least one line in your own voice before sending, because a message that sounds nothing like you defeats the purpose.

#### Planning: meals, trips, budgets, and schedules

Planning tasks are ideal for AI because they are constraint problems, and chat models are good at juggling constraints you state clearly.

- **Meals.** "Plan five weeknight dinners for two adults, one vegetarian, 30 minutes maximum each, and give me one combined grocery list organized by supermarket section."
- **Trips.** "Plan a three-day weekend in Rome for two people who like history and food, with a rough budget per day, walking distances between stops, and one rain backup per day."
- **Budgets.** "Here are my monthly expenses: [paste]. Group them into categories, show the three largest, and suggest two realistic cuts. Do not invent numbers."
- **Schedules.** "I have these commitments this week: [paste]. Build a day-by-day plan that fits in three hours of study time and keeps evenings after 8pm free."

The pattern to notice: every prompt includes limits (time, money, people, preferences). The more constraints you state, the more the plan looks like something you would actually follow. When a plan matters, ask one follow-up: "What did you assume that I did not tell you?" The answer usually surfaces the one thing you forgot to mention.

#### Learning and understanding: your patient explainer

The most underrated everyday use of AI is having things explained at exactly your level, as many times as you need, without anyone getting impatient.

- **Explain simply.** "Explain how index funds work like I am completely new to investing. Then give me the three questions a beginner should ask before buying one."
- **Summarize before you read.** "Summarize this article in five bullets, then tell me whether it is worth reading in full and why: [paste or upload]."
- **Check your understanding.** "I think this contract clause means [your interpretation]. Am I reading it right? Point out anything I am missing. Clause: [paste]."
- **Homework help that teaches.** "Do not give the answer. Walk me through how to approach this problem step by step, and stop before the final calculation: [paste]."

One caution for learning tasks: chat models occasionally state wrong facts with full confidence, especially for recent events, statistics, and niche topics. For anything you will repeat to others or act on, ask "what part of this answer are you least sure about?" and verify that part against a source you trust. The [how to use AI guide](https://whizi.io/resources/how-to-use-ai) has a full verification routine.

#### Documents and life admin

Paperwork is where AI stops feeling like a toy. Modern chat apps accept file uploads, which means the PDF you have been avoiding is a two-minute task.

- **Summarize a contract or policy.** Upload the file, then: "Summarize this in plain language. List every deadline, fee, and cancellation condition in a table."
- **Compare offers.** "Here are two insurance quotes: [paste both]. Build a table comparing price, coverage, exclusions, and cancellation terms, then list the questions I should ask each provider."
- **Draft official letters.** "Write a formal letter disputing this charge. Facts: [list them]. Keep it factual, one page, no emotional language."
- **Decode forms.** "This is a government form field I do not understand: [paste instructions]. Explain what they are asking for and what documents typically satisfy it."

For long documents, the quality of the answer depends on the model you use, and they genuinely differ at this task. If document work is a big part of your life, our guide to [summarizing PDFs with AI](https://whizi.io/resources/summarize-pdf-with-ai) compares the approaches. The quickest way to find your own answer is to run the same document through two models and see which one you trust with the fine print.

#### On your phone: AI in the checkout line

Most everyday AI moments do not happen at a desk. They happen in a parking lot, a kitchen, or a waiting room. Every major AI app, including Whizi, works in a mobile browser or as a mobile app, and voice input turns walking time into thinking time.

The tasks that work well on a phone are the ones you would never sit down at a laptop for:
- Dictate a rough voice note and ask for it back as a clean message.
- Photograph a menu or a sign in another language and ask for a translation, which [how to use AI to translate](https://whizi.io/resources/use-ai-to-translate) covers properly for anything longer than a sign.
- Ask for a quick comparison while you are standing in a shop: "what should I check before buying a used bike?"
- Take a photo of what is in the fridge and ask what you can cook with it.

The friction that kills the habit is retyping the same context every time. Two fixes. Keep your reusable prompts in a note you can paste from, and use the same app on your laptop and your phone rather than a different one on each, so your conversations and saved prompts follow you around.

#### A one-week starter plan

Habits beat marathons. Here is a seven-day plan that takes under fifteen minutes a day and leaves you with two or three workflows you will keep.

1. **Day 1:** Rewrite one real email with AI. Send it after your own edit.
2. **Day 2:** Summarize one long article or document you were avoiding.
3. **Day 3:** Plan the week's dinners and generate the grocery list.
4. **Day 4:** Take one thing you half-understand (a bill, a form, a news topic) and have it explained at your level.
5. **Day 5:** Draft something bigger: a complaint letter, a trip plan, a study schedule.
6. **Day 6:** Rerun your two favorite tasks on a different model and compare.
7. **Day 7:** Save your best prompts, delete the rest, and decide which two workflows become defaults.

Day 6 matters more than it looks. Models have different strengths, and the one that writes your warmest emails may not be the one you trust with a contract. Comparing them one by one across separate apps and subscriptions is exactly the friction that makes people give up, which is why Whizi puts ChatGPT, Claude, Gemini, and more in one workspace under [one subscription](https://whizi.io/pricing/). Start with the [free trial](https://whizi.io/register), run the week, and keep whatever saves you time.

Key points:

- The fastest wins for beginners are rewriting messages, summarizing long text, planning with constraints, and first drafts of anything
- Every example in this guide is a copy-paste prompt you can adapt in seconds, on your computer or your phone
- The goal is a small set of repeatable workflows, not using AI for everything

Checklist:

- Pick tasks where a draft helps and mistakes are cheap to catch.
- Give every prompt a job, the context, and the format you want back.
- State your constraints: time, budget, people, preferences.
- Never let the AI invent dates, prices, or commitments you did not provide.
- Ask "what did you assume?" before acting on any plan.
- Verify facts you will repeat or act on against a trusted source.
- Keep reusable prompts in one place you can paste from on any device.
- Compare two models on your top tasks before making either your default.

Steps:

1. Pick three recurring tasks from your week: Choose tasks you already do repeatedly, such as answering email, planning meals, or summarizing reading, rather than inventing new work for the AI.
2. Save one prompt per task: Write a short reusable prompt for each task with the job, the context, and the format you want back, and keep it somewhere you can paste from.
3. Run the tasks through the AI for one week: Every time the task comes up, do it with the AI first and note whether the result needed light edits, heavy edits, or a redo.
4. Compare a second model on the tasks that matter: For the tasks you now rely on, run the same prompt on a second model and keep the one that needs the least cleanup.
5. Keep what saves time, drop what does not: After a week, keep the two or three workflows that clearly saved time and make them your default way of doing that task.

FAQ:

Q: What are the best everyday tasks to start with?
A: Start with rewriting emails and messages, summarizing long documents, planning meals or trips with constraints, and having unfamiliar topics explained at your level. These show up weekly, save real time, and their quality is easy to judge.

Q: Do I need to pay for AI to use it in daily life?
A: Free tiers cover light use, but they limit how much you can ask and which models you get. If AI becomes part of your routine, one bundled subscription like Whizi is usually cheaper than paying two or three providers separately for their strongest models.

Q: What should I not use AI for?
A: Avoid relying on it alone for medical, legal, or financial decisions, live prices and deadlines, or anything you cannot verify yourself. Use it to prepare drafts, questions, and comparisons in those areas, and keep a qualified human or authoritative source in the loop.

Q: How much time does AI actually save?
A: For most people the honest answer is a few hours a week once two or three workflows stick, mostly from faster writing, faster reading, and faster first drafts. The savings come from repeatable tasks, not from occasional clever questions.

Q: Do I need different AI apps for different tasks?
A: You need different models more than different apps. Models differ at writing, documents, and analysis, so the practical setup is one app where you can switch models per task. That is the setup Whizi provides under a single subscription.

### How to use AI for market research: a repeatable workflow

URL: https://whizi.io/resources/ai-for-market-research/
Updated: August 2026

Quick answer: Use AI for market research by starting with a specific business decision, then working in stages: define the market question, collect and label sources, extract competitor and pricing data, pull exact customer language, and only then ask for synthesis. Verify claims against original pages before acting.

Use AI for market research with a step-by-step workflow for competitor mapping, pricing analysis, customer language, and positioning.

#### Define the market question

The fastest way to get bad market research from AI is to ask a broad question like "research this market." A model can produce a plausible overview, but plausible is not the same as useful. Good ai for market research starts with a specific decision: what are you trying to learn, what will change if the answer is different, and what evidence would make you trust the result?

Use this simple market question brief before you open a chat: market, buyer, decision, time horizon, known competitors, source types, and output format. Example: "We sell AI support software to B2B SaaS teams. We need to understand how competitors position automation versus human support so we can update our homepage. Analyze competitor websites, pricing pages, reviews, and public docs. Output a positioning matrix, proof points, risks, and recommendations."

That brief gives the model a job. It also keeps you honest. You are not asking AI to discover truth from nowhere; you are asking it to organize evidence, extract patterns, and help you make a better decision. Long-context models can help when you have many source snippets, pages, notes, or transcripts, but the workflow still matters: label sources clearly, ask for extraction before synthesis, and separate facts from interpretation.

Copy-paste prompt: `Act as a market research analyst. Help me define a research plan before doing analysis. Business: [business]. Market question: [question]. Decision this research will inform: [decision]. Known competitors: [list]. Sources available: [links, notes, reviews, calls, pricing pages]. Constraints: [geography, segment, budget, timeline]. Return: 1) refined research question, 2) evidence needed, 3) source collection plan, 4) output template, 5) risks or blind spots.`

Output should be a research plan, not an essay. QA the plan by asking whether each source connects to the decision. If a source will not change your pricing, positioning, roadmap, or go-to-market decision, it is probably not necessary.

#### Competitor map workflow

Competitor analysis with AI works best when you treat it like data extraction first and strategy second. Do not ask for a ranked list of competitors before you have defined categories. Instead, build a competitor map with direct competitors, adjacent alternatives, manual workarounds, and "do nothing" options. The last two are often where positioning gets sharper, because buyers compare your product against spreadsheets, agencies, internal labor, and inertia.

Inputs: competitor homepages, product pages, pricing pages, changelogs, help docs, review snippets, social posts, sales notes, and your own product positioning. If you are using long documents, group sources by company and label each block with the company name and URL. Ask the model to preserve uncertainty instead of filling gaps.

Prompt: `Build a competitor map for [category]. Use only the source notes below. Columns: company, category type, target customer, core promise, top features claimed, pricing signal, proof used, objection a buyer may have, positioning angle, source URL, confidence. Mark unknown when the source does not say. After the table, summarize the 5 patterns that matter most for our positioning. Sources: [paste labeled notes].`

| Competitor type | What to capture | Why it matters |
| --- | --- | --- |
| Direct | Same buyer, similar job, similar budget | Helps with feature and pricing comparisons |
| Adjacent | Different product, same outcome | Reveals substitution risk and category boundaries |
| Manual workaround | Spreadsheets, agencies, internal process | Shows the real status quo you must beat |
| Enterprise platform | Larger suite or incumbent | Clarifies trust, integration, and procurement expectations |
| Free or open option | Low-cost alternative | Helps explain when paid value is worth it |

Output: a table you can sort by customer, promise, price signal, and proof. QA: spot-check three rows against the original sources. Look for invented claims, overconfident price summaries, and unsupported positioning labels. Then ask: "What would a skeptical buyer say after reading this map?"

#### Pricing page analysis workflow

AI for pricing research is not about copying competitor prices. It is about understanding packaging logic: what is free, what is gated, what counts as usage, where teams hit limits, which buyer is being nudged upward, and how the company frames value. Pricing pages are positioning pages with numbers attached.

Inputs: screenshots or text from pricing pages, plan names, feature grids, usage limits, add-ons, FAQs, trial language, annual discount language, and any public terms that explain limits. Pricing changes frequently, so treat AI output as a research draft and verify current prices on the original pages before making decisions.

Prompt: `Analyze these pricing pages for [market]. Do not recommend a price yet. Extract the packaging model first. Columns: company, free/trial offer, entry plan, mid plan, top plan, usage metric, gated features, team/admin limits, enterprise trigger, annual discount signal, upgrade pressure, buyer assumption, source URL. Then summarize the common pricing patterns and where a new entrant could differentiate. Sources: [paste pricing text or notes].`

Use the output to answer practical questions: Are competitors pricing by seat, usage, credits, projects, storage, messages, or revenue tier? Are integrations gated? Is the free plan a real workflow or a demo path? Where does the buyer feel the first painful limit?

QA the pricing table with a "no hallucinated numbers" rule. If the source does not clearly state a price, the answer should say unknown. If a page uses "contact sales," the model should not infer a dollar amount.

#### Messaging extraction workflow

Customer language is the difference between a generic positioning doc and one that sounds like the market. AI for customer language research helps you extract repeated pains, desired outcomes, objections, buying triggers, and exact phrases from reviews, sales calls, support tickets, onboarding notes, Reddit threads, surveys, and competitor testimonials. The rule is simple: extract before you rewrite.

Inputs: review snippets, call transcripts, survey answers, support chats, public testimonials, forum posts, win/loss notes, and internal sales notes. Keep sensitive data out of tools unless your organization has approved that workflow. Remove private names, emails, account identifiers, and confidential customer details before analysis.

Prompt: `Extract customer language from the source material below. Do not invent quotes. Group exact phrases by pain, desired outcome, current workaround, objection, buying trigger, success metric, and emotional language. For each phrase, include source label, customer segment if known, and confidence. Then summarize the top 7 messaging themes in plain language. Sources: [paste anonymized notes].`

Output should include both exact phrases and themes. Exact phrases help you write copy that feels real. Themes help you make strategic choices. If customers repeatedly say "we spend Friday cleaning spreadsheets," that is stronger than "users need operational efficiency." The first phrase points to a landing page hook, an ad angle, and a demo narrative.

QA the extraction by scanning for invented quotes, over-grouped themes, and lost segment context. Ask the model to split themes by segment, company size, role, or maturity when the source material supports it.

#### Synthesis + positioning doc

Synthesis is where AI becomes valuable, but only after the extraction work is done. A useful ai positioning framework turns competitor patterns, pricing logic, and customer language into choices: who you serve, what problem you lead with, what proof you need, which alternatives you beat, and which claims you should avoid.

Use this structure for a positioning doc: target segment, urgent problem, current alternatives, main promise, differentiators, proof points, pricing implication, objections, message pillars, and experiments. Ask for a doc that separates evidence from recommendation so you can challenge the reasoning.

Prompt: `Create a positioning doc from the extracted research below. Use this format: 1) target segment, 2) buyer problem in customer language, 3) current alternatives, 4) market pattern, 5) recommended positioning thesis, 6) 3 message pillars, 7) proof needed for each pillar, 8) pricing and packaging implications, 9) risks, 10) 5 experiments to validate. Separate evidence from interpretation. If evidence is weak, say so. Research: [paste competitor map, pricing table, customer language themes].`

A strong synthesis doc should make tradeoffs visible. It should not say you are best for everyone. It might say: "Lead with fast competitor monitoring for seed-stage founders, not enterprise research automation. The evidence shows buyers complain about scattered research, unclear pricing, and slow synthesis. We need proof that setup takes under 15 minutes." That is a decision-ready output.

QA the positioning doc with a red-team pass: `Critique this positioning as if you were a skeptical buyer and a skeptical investor. Identify unsupported claims, vague language, competitor blind spots, pricing risks, and what evidence would change the recommendation.` Then revise.

#### Use the Founder Research Stack template

The easiest way to make this repeatable is to stop rebuilding the workflow from scratch. Whizi's Founder Research Stack template at /templates/founder-research-stack is built for exactly this kind of step-by-step market research: define the question, collect sources, extract competitor data, analyze pricing, pull customer language, synthesize positioning, and QA the final recommendation.

Use it when you are launching a product, rewriting a homepage, choosing a pricing model, preparing investor research, entering a category, or trying to understand why buyers choose one competitor over another. For academic or scientific work rather than commercial, [the researcher workspace](https://whizi.io/use-cases/researchers) covers the same steps with citations attached.

Here is the practical workflow inside Whizi. First, open the Founder Research Stack and paste your market question. Second, add source notes in labeled blocks. Third, run the competitor map prompt and save the table. Fourth, run pricing analysis and customer language extraction separately. Fifth, run the synthesis prompt using only the outputs from prior steps. Sixth, run the QA checklist before making product or marketing decisions.

Mid-research CTA: start with the Founder Research Stack, then create your account at /register when you are ready to save the workflow and compare outputs across models. For teams comparing options, review /pricing so you can pick a plan that fits your research cadence.

The goal is not to replace human judgment. The goal is to compress the boring parts while making the important parts easier to inspect. You still decide what to believe, test, and ship.

Key points:

- A founder-friendly workflow for turning messy market notes into a clear competitor map
- Prompt templates for pricing pages, customer language, positioning, and source-backed synthesis
- A QA checklist that keeps AI-assisted research grounded in evidence instead of confident guesses

Checklist:

- Define the business decision before asking AI to research a market
- Label every source by company, URL, date, and source type before synthesis
- Extract competitor, pricing, and customer language data before asking for strategy
- Use unknown when a source does not support a claim
- Verify current pricing and product claims on original pages before acting
- Separate exact customer phrases from AI-written theme summaries
- Ask for evidence, interpretation, risks, and recommendations as separate sections
- Run a skeptical buyer critique before finalizing a positioning doc
- Save the workflow as a repeatable template inside Whizi

FAQ:

Q: How can I use AI for market research without getting generic answers?
A: Start with a specific decision, provide labeled sources, and ask for extraction before synthesis. Generic prompts create generic market summaries; structured workflows create evidence you can inspect and reuse.

Q: Can AI do competitor analysis?
A: AI can help organize competitor notes, extract positioning, compare pricing pages, and identify patterns. You should still verify claims against original sources and review the strategy with human judgment.

Q: What should a founder market research workflow include?
A: A practical founder workflow includes a market question, competitor map, pricing analysis, customer language extraction, synthesis doc, positioning recommendation, and QA pass for unsupported claims.

### How to use AI for writing (without generic fluff)

URL: https://whizi.io/resources/ai-for-writing/
Updated: August 2026

Quick answer: Use AI for writing by giving the model three things before it drafts: Voice, a tone sample and banned phrases, Evidence, your real notes and facts, and Constraints, the length, reader, and claims to avoid. Then generate two versions, ask for a critique pass, revise, and edit by hand.

Learn how to use AI for writing emails, reports, and marketing copy with specific inputs, reusable prompts, and an editing checklist.

#### Why outputs sound generic

Most weak AI writing does not fail because the model cannot write. It fails because the prompt asks for writing without giving the model a real job. "Write a professional email," "make this sound better," or "create marketing copy" gives the AI almost nothing to work with. The result is fluent, safe, and forgettable.

Generic AI writing usually has the same tells: vague benefits, inflated adjectives, smooth transitions with no substance, and phrases no real person on your team would say. It may be grammatically correct, but it does not carry your context, your reader, or your point of view. That is why editing AI writing often feels strangely harder than starting from scratch.

The fix is to stop treating AI as a magic text button and start treating it as a junior draft partner. It needs the audience, the purpose, the source material, the constraints, and the voice target. Then it needs a second pass where you ask for sharper structure, fewer filler lines, and claims that can be verified.

A useful rule: if a human writer would ask follow-up questions before writing the piece, your AI prompt probably needs more context too. Good AI writing starts before the first sentence is generated.

#### Inputs that fix it: voice, examples, and constraints

The best way to make AI writing sound human is to give it human material. A model cannot infer your company voice, your relationship with a client, your product proof, or your risk tolerance from a one-line instruction. It can, however, follow examples and constraints surprisingly well when you provide them clearly.

Use the VEC framework before any writing prompt: Voice, Evidence, Constraints. Voice tells the model how the writing should sound. Evidence gives it facts to use. Constraints tell it what not to do.

| Input | What to include | Example |
| --- | --- | --- |
| Voice | A sample paragraph, tone notes, banned phrases | "Direct, calm, useful. No hype. Avoid 'game-changing.'" |
| Evidence | Notes, customer language, product facts, meeting bullets | "The beta reduced review time from 3 days to 1 day for two pilot teams." |
| Constraints | Length, format, reader, goal, must-avoid claims | "Under 180 words. Do not promise savings unless framed as an example." |

For longer source material, put the important material near the top of the prompt and label it. Anthropic guidance on long-context prompting emphasizes clear structure and careful placement of instructions and source content. The practical takeaway for writers is simple: do not paste a messy document dump and hope for clean prose. Give the model sections, labels, and the exact output you need.

Here is the before/after difference:

**Before:** `Write a follow-up email after a sales call.`

**After:** `Write a follow-up email to a VP of Operations after a 30-minute discovery call. Goal: recap the pain points and get agreement for a pilot next week. Voice: concise, practical, not salesy. Evidence: they manage 18 field teams, onboarding takes 11 days, they care about audit trails. Constraints: under 170 words, include 3 bullets, do not mention pricing yet, end with one clear question.`

The second prompt is not longer for decoration. It gives the AI enough boundaries to produce something you can actually edit.

#### Prompt patterns for emails, reports, and landing pages

Use these prompt patterns as reusable starting points. Replace the bracketed sections with your real material, then run the same prompt across models in Whizi when quality matters.

**Email prompt: clear, warm, and specific**

`You are writing a business email for [relationship/context]. Audience: [recipient]. Goal: [what should happen next]. Source notes: [paste bullets]. Voice: [tone]. Constraints: under [word count], preserve all facts, no exaggerated praise, no generic opener. Return: subject line, email body, and one alternate closing line.`

**Email rewrite prompt: shorter without losing tact**

`Rewrite this email to be clearer, shorter, and easier to answer. Keep the tone [warm/direct/diplomatic]. Preserve dates, names, numbers, and commitments exactly. Remove filler. End with one clear next step. Draft: [paste email].`

**Report prompt: evidence first**

`Create a report from the notes below. Separate facts, interpretation, risks, and recommendations. Do not invent data. If a claim needs verification, mark it [verify]. Use headings and concise bullets. Audience: [executive/team/client]. Notes: [paste notes].`

**Marketing copy prompt: angle before polish**

`Create three landing page messaging angles for [product]. Audience: [specific segment]. Problem: [pain]. Proof: [evidence]. Differentiator: [why you]. Constraints: no vague claims, no unsupported statistics, no "all-in-one" unless explained. For each angle, include headline, subhead, proof point, objection handled, and CTA.`

**Tone editing prompt: make it sound like us**

`Edit the draft to match the voice sample. Preserve the meaning and all factual claims. Match directness, sentence length, vocabulary, and level of detail. Remove phrases that sound generic. Voice sample: [paste sample]. Draft: [paste draft].`

The pattern is consistent: give the AI the job, reader, source material, voice, constraints, and output format. Then ask for a short note explaining what it changed. That note helps you catch whether the model understood the assignment or merely rewrote the surface.

#### A before/after AI writing workflow

A strong AI writing workflow is not "prompt once and publish." It is brief, draft, critique, revise, verify. The process is fast, but it still has gates.

**Step 1: Brief the task.** Write three lines before prompting: who is this for, what should they do after reading, and what facts must be included. This turns a vague writing request into a usable assignment.

**Step 2: Generate two versions.** Ask for one concise version and one more persuasive version. This gives you contrast. For marketing copy, ask for different angles. For reports, ask for different structures. For emails, ask for different levels of warmth.

**Step 3: Ask for criticism before revision.** Use: `Before rewriting, identify the three weakest parts of this draft: unclear point, generic language, unsupported claim, or wrong tone.` This makes the model evaluate the writing instead of endlessly polishing it.

**Step 4: Revise with constraints.** Ask for a tighter draft with specific requirements: shorter sentences, stronger opening, concrete examples, no new claims, and a clear next step.

**Step 5: Human edit.** Read the output like an editor, not a passenger. Replace generic lines with real details. Check product claims. Cut anything that sounds like it could appear on a thousand other websites.

Here is a simple before/after example.

**Before:** "Our platform helps teams streamline workflows and unlock productivity with powerful AI-driven solutions."

**After:** "Whizi lets you test the same writing prompt across multiple AI models, compare the drafts, and keep the version that needs the least editing."

The better version names the product action, the user behavior, and the outcome. It is not louder. It is more specific.

#### Editing checklist

Use this checklist before sending an AI-assisted email, publishing marketing copy, or sharing a report.

- **Reader:** Is the piece written for a specific person or segment?
- **Purpose:** Is the desired action obvious?
- **Opening:** Does the first line say something useful, or does it warm up with filler?
- **Specificity:** Can you replace any vague benefit with a proof point, example, or concrete use case?
- **Voice:** Would your team actually say this sentence out loud?
- **Facts:** Are dates, numbers, names, features, prices, and claims correct?
- **Unsupported claims:** Did the AI invent outcomes, guarantees, statistics, or customer proof?
- **Structure:** Is the main point high enough, especially for busy readers?
- **Length:** Can you cut 15 percent without losing meaning?
- **CTA:** Is there exactly one next step when the piece needs action?

For business writing, the highest-value edit is usually not making the prose prettier. It is making the point clearer. For marketing copy, the highest-value edit is replacing abstraction with proof. For reports, it is separating what you know from what you think it means.

#### Run these prompts in Whizi

Different models often write differently from the same brief. One may produce a warmer email. Another may structure a report more cleanly. Another may create stronger marketing angles. You do not need to guess which one is best from a brand name. You can test the prompt.

In Whizi, start with one real writing task: an email you need to send, a report you need to summarize, or a landing page section you need to improve. Paste the same VEC-based prompt into multiple models, compare the drafts, then use the editing checklist above. Save the prompt that worked so the next writing task starts faster.

This is the practical reason to use AI for writing inside a shared workspace instead of jumping between tabs. You get a repeatable process: brief the work, compare outputs, keep the strongest draft, and edit with intent. If the writing is for clients rather than for an employer, [AI tools for freelancers](https://whizi.io/use-cases/freelancers) covers the proposal, scoping, and client voice work that surrounds the draft, and if it is coursework, [the student workspace](https://whizi.io/use-cases/students) covers essays and revision alongside it. When you are ready, create an account and test your next writing prompt in Whizi.

Key points:

- A repeatable AI writing workflow for emails, reports, and marketing copy
- Prompt patterns that use voice samples, source notes, constraints, and revision passes
- A practical editing checklist for making AI drafts clearer, more specific, and more human

Checklist:

- Use the VEC framework: Voice, Evidence, and Constraints
- Give the AI a real audience, source notes, and a clear next step
- Generate multiple versions before choosing a draft
- Ask the model to critique generic language and unsupported claims
- Human-edit for specificity, accuracy, structure, and voice before publishing

FAQ:

Q: How do I use AI for writing without sounding generic?
A: Give the model voice examples, source notes, audience context, and constraints. Then ask for a critique pass that identifies generic language, unclear points, and unsupported claims before revising.

Q: Can AI write business emails?
A: Yes, but the prompt should include the relationship, goal, facts to preserve, desired tone, length, and next step. Always review the output for accuracy and social nuance before sending.

Q: What is the best AI writing workflow?
A: Use a five-step workflow: brief the task, generate multiple versions, critique the draft, revise with constraints, and human-edit for specificity, facts, and voice.

### The AI glossary: 55 terms in plain English

URL: https://whizi.io/resources/ai-glossary/
Updated: August 2026

Quick answer: This AI glossary defines 55 terms in plain English, each entry leading with a one-sentence definition you can quote and following it with the detail. It is organized by what you are doing rather than alphabetically, covering the basics, prompting, how models are made, failure modes, retrieval, work beyond text, and the wider ecosystem.

A reference AI glossary covering 55 terms in plain English, from token and context window to RAG, fine-tuning, and agents, each with a one-sentence definition first.

#### The basics

This is the reference version: 55 terms, definition first, detail second. If you want the gentler tour of just the words that matter in your first week, read [AI terms explained](https://whizi.io/resources/ai-terms-explained) instead and come back when a term here sends you.

### Artificial intelligence

Artificial intelligence is software that performs tasks which normally require human judgment, such as writing, reasoning, or recognizing what is in an image.

In everyday use in 2026, "AI" almost always means the chat assistants built on large language models rather than the broader academic field.

### Machine learning

Machine learning is the technique of teaching software by showing it examples rather than programming it with rules.

Every modern AI assistant is built this way: the behavior comes from patterns in training data, which is why the same model can be brilliant on common tasks and shaky on rare ones.

### Model

A model is one specific trained AI system, the artifact you actually send a prompt to.

GPT, Claude, and Gemini are model families; a family contains individual models with different sizes, speeds, and prices. Choosing between them per task is the core skill of using AI well.

### LLM (large language model)

An LLM, or large language model, is a model trained on enormous amounts of text to predict what text should come next.

That single trick, done at scale, is what produces answers, drafts, code, and summaries. The "large" refers to both the training data and the parameter count.

### Chatbot

A chatbot is the conversational interface wrapped around a model, the thing with the message box.

The distinction matters when comparing products: two chatbots can wrap the same model and feel different, and one chatbot can offer many models, which is how Whizi works.

### Prompt

A prompt is everything you send the model: the question, the instructions, and any text or files you include.

Answer quality tracks prompt quality closely. Context, constraints, and the output format you want are the three things beginners most often leave out.

### Token

A token is the small chunk of text a model actually reads and writes, roughly three quarters of an English word on average.

Pricing, speed, and context windows are all measured in tokens, which is why the word appears on every AI pricing page. A 1,000-word document is roughly 1,300 tokens.

### Context window

A context window is the maximum amount of text, measured in tokens, that a model can consider at once.

Your prompt, the conversation so far, attached documents, and the answer all share it. The [context window comparison](https://whizi.io/resources/context-window-comparison) table lists the current figure for every major model, and the [explainer](https://whizi.io/resources/what-is-a-context-window) covers why it explains most "the AI forgot" complaints.

### Parameters

Parameters are the internal numbers a model learned during training, and the usual measure of its size.

Counted in billions, more parameters loosely means more capability and more cost. It is a poor shopping metric on its own; benchmarks and your own tests beat parameter counts.

### Response

A response is the model's output for one prompt, also called a completion.

Responses are generated fresh every time, which is why regenerating the same prompt gives different wording and sometimes a different answer.

#### Talking to a model

### System prompt

A system prompt is the standing instruction a model receives before your message, setting its role, tone, and rules.

Every product has one even when you cannot see it. When a chatbot "has a personality", the system prompt is usually where it lives.

### Prompt engineering

Prompt engineering is the practice of writing prompts deliberately to get reliably better output.

Despite the grand name, most of it is ordinary clarity: state the goal, give context, show an example, name the format. The [beginner guide](https://whizi.io/resources/prompt-engineering-beginners) covers the techniques that actually move results.

### Zero-shot

Zero-shot means asking a model to do a task without giving it any examples.

Modern models handle most everyday tasks zero-shot. The term survives mostly as a contrast with few-shot prompting.

### Few-shot

Few-shot means including a handful of worked examples in the prompt so the model copies the pattern.

It is the single cheapest reliability upgrade for formatting-sensitive work: two or three examples of input and desired output beat a paragraph of description.

### Chain of thought

Chain of thought is a model working through intermediate steps before giving its final answer.

Asking a model to reason step by step measurably improves hard problems, and reasoning models now do a version of it internally by default.

### Temperature

Temperature is the setting that controls how much randomness a model adds when choosing each next token.

Low temperature gives consistent, conservative output; high temperature gives variety and occasional nonsense. Chat products usually pick a middle value for you.

### Streaming

Streaming is the model sending its answer token by token as it generates, rather than all at once when finished.

It is why answers appear to type themselves, and why a bad answer can be stopped early instead of waited out.

### Regenerate

Regenerate means asking for a fresh response to the same prompt.

Because generation is probabilistic, a regenerate is a genuinely new draw, not a retry of the same computation. Comparing two draws, or the same prompt on two models, is a fast quality check.

#### How models are made

### Training

Training is the process of adjusting a model's parameters against data until its predictions get good.

It happens before you ever meet the model and costs the provider enormous compute. Nothing you type into a chat retrains the model in the moment, though providers may use conversations to train future versions depending on their policy.

### Training data

Training data is the text and other material a model learned from.

Its breadth explains what a model knows, its gaps explain systematic blind spots, and its date range sets the knowledge cutoff.

### Pretraining

Pretraining is the first, largest phase of training, where a model learns language and world knowledge from bulk text.

The result is capable but raw. Everything that makes a model behave like a helpful assistant comes afterward.

### Fine-tuning

Fine-tuning is additional training on a narrower dataset to specialize or shape a pretrained model.

Providers fine-tune for helpfulness and safety; companies fine-tune for domain tasks. For most users, good prompting plus a capable general model beats paying for a fine-tune.

### RLHF

RLHF, reinforcement learning from human feedback, is training a model against human ratings of its answers.

It is a large part of why modern assistants are polite, structured, and cautious. Overdone, it also produces the hedging that makes some models refuse or waffle.

### Alignment

Alignment is the work of making a model's behavior match human intent and values.

In practice it covers everything from refusing harmful requests to simply following instructions accurately. When a model does something technically impressive but unwanted, that is an alignment gap.

### Inference

Inference is running a trained model to produce an answer, as opposed to training it.

Every message you send triggers inference, and inference cost is what token pricing measures. Cheap-to-run models can be remarkably capable; the [AI Model Cost Index](https://whizi.io/tools/model-cost-index) shows a 2,000x price spread.

### Transformer

The transformer is the neural network architecture behind essentially every modern language model.

Its key mechanism, attention, lets the model weigh every part of the input against every other part. You never need the details, but the word is everywhere in AI writing.

#### When things go wrong

### Hallucination

A hallucination is a model stating something false with complete confidence.

It is not a bug that will be patched next release; it follows from how generation works. Frequency varies a lot by model and task. Treat any specific fact, number, or citation as unverified until checked. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) covers the mechanics.

### Grounding

Grounding is giving a model authoritative material to answer from, instead of letting it rely on memory.

Pasting the contract, attaching the report, or enabling web search are all grounding. It is the most effective everyday defense against hallucination.

### Bias

Bias is a model systematically skewing its outputs in ways inherited from training data.

It shows up as assumptions about people, places, and defaults. For consequential decisions, review AI output the way you would review a smart but unvetted intern's work.

### Guardrails

Guardrails are the restrictions a provider builds around a model to block harmful or off-policy output.

Different providers draw the lines differently, which is a real product difference: the same request can be answered by one model and refused by another.

### Jailbreak

A jailbreak is a prompt crafted to trick a model into ignoring its guardrails.

Providers patch them continuously. For ordinary users the term mostly matters as a security concept: anything a model can be talked into, someone will try to talk it into.

### Red teaming

Red teaming is deliberately attacking a model before release to find harmful failure modes.

Providers run internal and external red teams, and published model cards often summarize what they found. It is the AI equivalent of penetration testing.

### Knowledge cutoff

A knowledge cutoff is the date after which a model's training data ends.

Ask about anything more recent and the model either admits ignorance, browses the web if it can, or hallucinates. Knowing the cutoff of the model you are using prevents a whole category of errors.

#### Retrieval and your data

### RAG (retrieval augmented generation)

RAG, retrieval augmented generation, is fetching relevant documents first and having the model answer from them.

It is how "chat with your knowledge base" products work under the hood, and why they can cite sources. RAG reduces hallucination but inherits the quality of whatever it retrieves.

### Embedding

An embedding is a list of numbers representing the meaning of a piece of text.

Texts with similar meaning get nearby numbers, which lets software find related passages mathematically. Embeddings are the machinery behind semantic search and RAG.

### Vector database

A vector database is a database built to store embeddings and find the nearest ones fast.

If a product says it "indexed" your documents for AI search, an embedding for every chunk is sitting in one of these.

### Semantic search

Semantic search is finding text by meaning rather than by matching keywords.

Searching "money problems" can surface a passage about cash flow strain even though no word matches. It is embeddings put to work.

### Chunking

Chunking is splitting long documents into pieces small enough to embed and retrieve well.

It sounds like plumbing and it is, but bad chunking is a common reason document-chat tools answer from the wrong part of a file.

### Web search (in AI)

Web search in AI means the model fetching current pages before answering, instead of relying on training memory.

It trades speed for freshness and lets answers carry citations you can open. For anything after the model's knowledge cutoff, it is the difference between an answer and a guess.

### Knowledge base

A knowledge base is the curated set of documents a company gives its AI to answer from.

Quality in, quality out: an assistant grounded on a stale wiki confidently serves stale answers, with citations.

#### Beyond text

### Multimodal

Multimodal means a model that works across more than one medium, such as text plus images, audio, or video.

It is why you can screenshot an error and ask what is wrong. The [multimodal AI guide](https://whizi.io/resources/multimodal-ai) covers what actually works today.

### Vision model

A vision model is a model that can interpret images: photos, screenshots, charts, and documents.

Reading is not the same as generating; a model can describe your diagram perfectly and still be unable to draw one.

### Image generation

Image generation is producing new images from a text description.

Quality, style range, and text rendering inside images vary sharply between generators, which is why serious users compare several. The [image generator roundup](https://whizi.io/resources/best-ai-image-generators) stays current on the field.

### Diffusion model

A diffusion model generates images by starting from noise and refining it step by step toward the description.

It is the dominant technique behind modern image generators. The word mostly matters so that "diffusion" in a product name tells you what the product does.

### Speech to text

Speech to text is transcribing spoken audio into written words.

Modern transcription is accurate enough to make meetings, voice notes, and interviews searchable, and it is how voice input in chat apps works.

### Text to speech

Text to speech is generating spoken audio from written words.

Current systems produce natural, expressive voices, which powers read-aloud features and AI voice conversations.

#### The ecosystem

### API

An API is the programmatic doorway developers use to send prompts to a model from their own software.

API pricing is per token, which is where the per-answer economics of every AI product ultimately come from.

### Open weights

Open weights means a model's trained parameters are published for anyone to download and run.

It enables self-hosting and lets other providers serve the model on their own infrastructure. DeepSeek and Llama are the famous examples.

### Open-source model

An open-source model is, loosely, a model released for public use and modification, though licenses vary widely.

The practical questions are always the same two: can you use it commercially, and who is hosting it for you. The [open weights ecosystem](https://whizi.io/resources/how-to-choose-ai-model) matters most when privacy or cost rules out hosted flagships.

### Benchmark

A benchmark is a standard test used to score and compare models.

Useful for a first cut, routinely overfit in marketing. A model that tops a benchmark can still lose on your actual work, which is why running your own three-task comparison beats reading charts.

### Agent

An agent is an AI system that takes actions in steps toward a goal, rather than just answering once.

Booking, researching, coding, and filing tickets are agent territory. The capability is real and improving fast; so is the need to review what an agent did before it matters.

### Tool use

Tool use, or function calling, is a model invoking external capabilities like search, calculators, code, or your calendar.

It is how chat assistants act on the world instead of only describing it, and it is the mechanism underneath agents.

### Reasoning model

A reasoning model is one that spends extra compute thinking through a problem internally before answering.

The trade is time and cost for accuracy on hard problems. For simple questions the extra thinking buys nothing, which is another argument for switching models per task.

### Latency

Latency is how long you wait between sending a prompt and the answer arriving or starting to stream.

It varies by model size, load, and whether the model is a reasoning model. For interactive work, a fast good-enough model often beats a slow brilliant one.

### Rate limit

A rate limit is the cap a provider puts on how many requests or tokens you can use in a time window.

Hitting one is the usual reason a chat product tells you to wait or upgrade. Message caps on subscription plans are rate limits with friendlier branding.

#### Keeping the glossary useful

Terms age fast in this field. This page is maintained as a reference: entries get rewritten when usage shifts, and the linked deep-dive guides carry the detail that does not fit a definition.

If a term you hit in the wild is missing here, the fastest route is to paste the sentence you found it in into a capable model and ask for the definition in context. In Whizi you can ask two models at once and notice when they disagree, which for new jargon is surprisingly often.

Key points:

- Every entry starts with a one-sentence definition you can quote, then the detail
- Organized by what you are doing, not alphabetically, so neighbouring terms explain each other
- If you are brand new, the short beginner version at AI terms explained is the friendlier front door

Checklist:

- Quote the one-sentence definition; read the detail before relying on it
- When two sources define a term differently, the newer usage usually wins in AI
- Learn token, context window, and hallucination first; most other terms hang off those three
- Test unfamiliar claims about a model against your own three-task comparison
- Use the beginner glossary first if this page feels like a reference manual, because it is one

FAQ:

Q: What is the difference between AI, machine learning, and an LLM?
A: AI is the broad goal of software doing tasks that need human judgment. Machine learning is the technique of teaching software from examples, and it is how essentially all modern AI is built. An LLM is one kind of machine-learned model, trained on text, and it is the kind behind ChatGPT, Claude, and Gemini.

Q: What does RAG mean in AI?
A: RAG stands for retrieval augmented generation: the system first fetches relevant documents, then has the model answer from them rather than from memory. It is the standard technique behind products that chat with your files or knowledge base, and it is why those products can cite sources.

Q: What is the difference between fine-tuning and RAG?
A: Fine-tuning changes the model itself with extra training, which is slow and fixed. RAG leaves the model alone and hands it the right documents at question time, which is flexible and current. For keeping an assistant up to date with changing information, RAG is almost always the right tool; fine-tuning suits stable style and format requirements.

Q: What is a token in AI?
A: A token is the unit of text a model reads and writes, roughly three quarters of an English word. Everything is measured in tokens: pricing, speed, and the context window. A 1,000-word document is about 1,300 tokens.

Q: How many of these terms do I actually need?
A: For everyday use, about five: model, prompt, token, context window, and hallucination. The beginner version of this glossary covers those with more patience. The rest of this page exists for the moment a product page or an article throws a term at you.

### AI model release timeline: every major launch since 2022

URL: https://whizi.io/resources/ai-model-timeline/
Updated: August 2026

Quick answer: This AI model release timeline tracks every major launch since ChatGPT arrived on November 30, 2022, each with a verified date and one line on why it mattered. Entries cover GPT, Claude, Gemini, Grok, DeepSeek, Llama, and the open-weights wave, and new releases are added on launch day, newest first.

A maintained timeline of major AI model releases from ChatGPT to today: GPT, Claude, Gemini, Grok, DeepSeek, Llama, and the open-weights wave, with dates and why each mattered.

#### This year

The most recent entry is **GLM-5.3** from Z.ai on August 14, 2026. New launches are added to this table on release day, newest first.

| Date | Model | Provider | Why it mattered |
| --- | --- | --- | --- |
| August 14, 2026 | **GLM-5.3** | Z.ai | The GLM line's August refresh in a crowded open-weights month. |
| August 13, 2026 | **Gemini 3.7 Flash** | Google | The fast tier keeps absorbing capability between Pro releases. |
| August 6, 2026 | **Grok 4.6** | xAI | Focused on long-running agents, with a 500K context window. |
| August 2, 2026 | **Qwen3.8 Max** | Qwen | Alibaba's biggest Qwen release of the summer. |
| July 16, 2026 | **Grok 4.5** | xAI | xAI's mid-year coding and cost push. |
| July 16, 2026 | **Kimi K3** | Moonshot | A 2.8T open-weights flagship announced with weights following on July 27. |
| July 9, 2026 | **GPT-5.6 (Sol, Terra, Luna)** | OpenAI | Three named variants replace the single-flagship pattern. |
| June 30, 2026 | **Claude Sonnet 5** | Anthropic | The 5-family workhorse: near-flagship quality at Sonnet pricing, default across Anthropic products. |
| June 9, 2026 | **Claude Fable 5** | Anthropic | The first Mythos-class model, opening Anthropic's 5 family above the Opus tier. |
| May 28, 2026 | **Claude Opus 4.8** | Anthropic | The Opus line's spring capstone. |
| April 24, 2026 | **DeepSeek V4** | DeepSeek | DeepSeek's next major line lands one day after GPT-5.5. |
| April 23, 2026 | **GPT-5.5** | OpenAI | Billed as OpenAI's smartest and most intuitive model; Instant variant reached free users in May. |
| April 16, 2026 | **Claude Opus 4.7** | Anthropic | First of three frontier releases inside eight days. |
| March 5, 2026 | **GPT-5.4** | OpenAI | The 5-series cadence settles at roughly one release per quarter. |
| February 19, 2026 | **Gemini 3.1 Pro** | Google | Google's spring flagship refresh. |
| February 5, 2026 | **GPT-5.3-Codex** | OpenAI | A dedicated coding flagship, aimed at agentic software work. |

#### 2025: the reasoning race and the price collapse

The year DeepSeek R1 made cheap reasoning a market event, reasoning models became the default frontier format, and November packed four frontier releases into thirteen days.

| Date | Model | Provider | Why it mattered |
| --- | --- | --- | --- |
| December 11, 2025 | **GPT-5.2** | OpenAI | OpenAI's December refresh of the 5-series. |
| December 1, 2025 | **DeepSeek V3.2** | DeepSeek | The sparse-attention release goes stable, with tool calls in reasoning mode. |
| November 24, 2025 | **Claude Opus 4.5** | Anthropic | Anthropic closes the November wave with a large price cut on its flagship tier. |
| November 18, 2025 | **Gemini 3 Pro** | Google | Google's November entry in the densest release month the field had seen. |
| November 12, 2025 | **GPT-5.1** | OpenAI | Instant and Thinking variants, opening a packed November. |
| September 29, 2025 | **DeepSeek V3.2-Exp** | DeepSeek | Sparse attention in the open: the experiment that became V3.2. |
| August 7, 2025 | **GPT-5** | OpenAI | The long-awaited flagship, with automatic routing between fast and thinking modes. |
| July 11, 2025 | **Kimi K2** | Moonshot | A trillion-parameter open-weights model from China's Moonshot, strong on agentic tasks. |
| July 9, 2025 | **Grok 4** | xAI | xAI reaches the frontier tier, with live X data as the differentiator. |
| May 22, 2025 | **Claude Opus 4 and Sonnet 4** | Anthropic | Positioned squarely at coding and long-running agent work. |
| April 16, 2025 | **o3 and o4-mini** | OpenAI | Reasoning models get tool use: thinking plus searching, coding, and reading images. |
| April 5, 2025 | **Llama 4 (Scout, Maverick)** | Meta | Meta's open-weights line goes multimodal with very long context. |
| March 25, 2025 | **Gemini 2.5 Pro** | Google | Google takes benchmark leads with a 1M context window as the headline feature. |
| February 24, 2025 | **Claude 3.7 Sonnet** | Anthropic | Hybrid reasoning: one model deciding when to think longer. |
| January 20, 2025 | **DeepSeek R1** | DeepSeek | The release that moved markets: open-weights reasoning at roughly a tenth of Western API prices. |

#### The build-up: 2022 to 2024

The releases that created the current landscape: the ChatGPT launch, the first frontier race, and the arrival of multimodal and reasoning models.

| Date | Model | Provider | Why it mattered |
| --- | --- | --- | --- |
| December 26, 2024 | **DeepSeek V3** | DeepSeek | Open weights near frontier quality at a fraction of the training cost. The stage-setter for January. |
| December 11, 2024 | **Gemini 2.0 Flash** | Google | Fast-tier models get genuinely capable, foreshadowing the cheap-model price collapse. |
| September 12, 2024 | **o1-preview** | OpenAI | The first mainstream reasoning model: slower answers that think before responding. |
| June 20, 2024 | **Claude 3.5 Sonnet** | Anthropic | A mid-tier model beating the previous flagship, resetting expectations for the Sonnet tier. |
| May 13, 2024 | **GPT-4o** | OpenAI | Native voice and vision in one model, and free-tier access to frontier capability. |
| March 4, 2024 | **Claude 3 (Opus, Sonnet, Haiku)** | Anthropic | The tier naming that stuck: a flagship, a workhorse, and a fast cheap model. |

| Date | Model | Provider | Why it mattered |
| --- | --- | --- | --- |
| December 6, 2023 | **Gemini** | Google | Google's answer to GPT-4 and the start of the three-way frontier race. |
| July 11, 2023 | **Claude 2** | Anthropic | Anthropic opens Claude to the public, with a then-huge 100K context window. |
| March 14, 2023 | **GPT-4** | OpenAI | The first frontier model most professionals used, and the start of paid AI subscriptions at scale. |

| Date | Model | Provider | Why it mattered |
| --- | --- | --- | --- |
| November 30, 2022 | **ChatGPT** | OpenAI | The launch that made conversational AI a consumer product. |

#### What the cadence tells you

**The gaps keep shrinking.** GPT-4 held the frontier for over a year. By late 2025, four labs shipped frontier models in a single month, and in April 2026 three landed inside eight days. A subscription decision made on "which model is best" now has a shelf life measured in weeks.

| Period | What shipped | What it means for a subscription |
| --- | --- | --- |
| The GPT-4 era | One model held the frontier for over a year | A "which model is best" choice stayed valid for months |
| November 2025 | Four frontier releases inside thirteen days | Launch-week benchmark charts stopped settling anything |
| April 2026 | Claude Opus 4.7, GPT-5.5, and DeepSeek V4 inside eight days | The frontier can change hands between billing cycles |
| 2024 to 2026 | DeepSeek, Kimi, Qwen, GLM, and Llama moved from footnotes to frontier contention | The widening gap is price, not capability |

**Releases cluster on purpose.** Labs increasingly ship into each other's launch windows: Claude Opus 4.7, GPT-5.5, and DeepSeek V4 arrived within eight days of each other in April 2026, and the November 2025 wave was even denser. Launch-week benchmark charts are marketing artillery in those windows, which is why they rarely settle anything.

**The open-weights track is no longer behind.** DeepSeek, Kimi, Qwen, GLM, and Llama entries have moved from footnotes to frontier contention in two years, usually at a fraction of the price. The practical effect is the widening cost spread the [AI Model Cost Index](https://whizi.io/tools/model-cost-index) tracks.

The practical conclusion has not changed since GPT-4: when a release lands, do not switch on the announcement. Run your own three real tasks against the new model next to your current one; the [release evaluation guide](https://whizi.io/releases/how-to-evaluate-ai-models) is a one-hour method for exactly that. In Whizi the new frontier models land in the same picker as the old ones, so the comparison is a conversation rather than a second subscription.

#### Method, sources, and corrections

Dates are public release-announcement dates, verified against provider announcements and same-day press coverage before an entry is added. Where only a month is solidly sourced, the entry shows the month without inventing a day. Preview, waitlist, and staged rollouts are dated to the first public availability announcement.

The table deliberately tracks major releases rather than every checkpoint: a release earns an entry when it changed what a mainstream user or developer could buy, run, or compare. Minor point updates and silent model swaps do not qualify.

Corrections are welcome and applied fast: if a date here disagrees with a primary source, use the contact on the [support page](https://whizi.io/support) and cite the source. This page is intended to be citable, so its accuracy is the product.

Key points:

- Most recent entry: GLM-5.3 from Z.ai, August 14, 2026
- 41 releases tracked since ChatGPT, each with a verified date and one line on why it mattered
- Maintained on launch days, with the dating method stated at the bottom

Checklist:

- Check the date on this page's newest entry before citing it as current
- Treat launch-week benchmark charts as marketing until third parties reproduce them
- Test a new release on three of your own tasks before switching anything
- Expect the frontier to change hands within weeks, and pick tools that let you follow it
- Cite the entry date, not the coverage date; they differ for staged rollouts

FAQ:

Q: What is the most recent major AI model release?
A: The newest entry in this timeline is GLM-5.3 from Z.ai, released August 14, 2026. The table above is updated on launch days, so the top row is the current answer.

Q: How often do new AI models come out?
A: Fast and accelerating: the trailing twelve months of this timeline contain 22 major releases across the frontier labs and the open-weights ecosystem, and that counts only launches significant enough to earn an entry. The practical takeaway is to stop optimizing for a permanent winner.

Q: When was ChatGPT released?
A: November 30, 2022. It was the moment large language models became a consumer product, and every entry in this timeline is downstream of the subscription market it created.

Q: Should I switch models every time something new comes out?
A: No. Switch when a new model wins on your own repeated tasks, not on launch benchmarks. The one-hour evaluation in the release guide linked above is enough to tell, and running it inside a multi-model workspace means the test costs a conversation instead of a new subscription.

### AI subscription costs: calculate your true monthly spend

URL: https://whizi.io/resources/ai-subscription-costs/
Updated: August 2026

Quick answer: Calculate your true AI subscription cost by listing every paid plan from your billing pages, converting annual charges to a monthly figure, and tagging each tool by the job it does. Then separate active spend from duplicated spend. Whizi's savings calculator does the totals for you.

Use this AI subscription cost calculator method to find hidden tool spend, decide when consolidation makes sense, and pick a cleaner AI plan.

#### The subscription stack problem

Most AI subscription costs do not start with a budget decision. They start with a small exception: one person needs a better writing model, another wants a coding assistant, someone else signs up for an image tool, and the founder keeps a separate account for research. Each charge feels reasonable in isolation. The problem appears when the team realizes it is paying for the same core capability several times.

That is why an AI subscription cost calculator is more useful than a generic AI tools pricing comparison. Pricing pages change, plans split into personal and team tiers, and usage limits move. Instead of memorizing exact competitor prices, verify current pricing directly on each vendor site, then calculate what your stack actually costs in your workflow. The right question is not only "how much does ChatGPT cost?" It is "how much are we paying every month for AI work, and how much of that spend is duplicated?"

The fastest path is to open the [Whizi AI subscription savings calculator](https://whizi.io/tools/calculator) now, enter your current tools, and come back to this guide when you want the decision logic. If you already know you are paying for multiple AI tools, the calculator will give you a cleaner answer than another hour of reading.

AI spend gets messy for three reasons. First, teams buy by brand instead of by job: writing, coding, research, summarization, image generation, document analysis, and brainstorming. Second, the buyer is often not the daily user, so unused seats sit quietly on cards. Third, the value of a model changes by task. A tool that is great for long-form drafting may not be the best place for code review or PDF analysis, so users keep stacking tools instead of designing a shared workflow.

#### Your cost calculator method

Use this worksheet before you decide whether to consolidate AI subscriptions. Do not guess. Pull the current plan price from each vendor billing page, note whether it is monthly or annual, and include taxes or seat minimums if they apply. If your team pays annually, divide by 12 so every line uses the same monthly basis.

| Tool or plan | Primary job | Monthly cost | Seats | Used weekly? | Overlap risk | Keep, cut, or test |
| --- | --- | ---: | ---: | --- | --- | --- |
| AI chatbot A | Writing and research | Verify current price | 1 | Yes | High if another chatbot covers the same work | Test consolidation |
| AI coding tool | Debugging and tests | Verify current price | 2 | Yes | Medium if developers still need IDE-native help | Keep or partial cut |
| Image generator | Ads and thumbnails | Verify current price | 1 | Sometimes | Medium if image work is occasional | Test pay-as-needed |
| PDF/document tool | Summaries and extraction | Verify current price | 3 | Rarely | High if a multi-model workspace handles docs | Cut or consolidate |
| Research assistant | Sources and synthesis | Verify current price | 1 | Yes | Medium if citations are essential | Compare workflow quality |

Now calculate three numbers. **Current monthly AI spend** is every paid AI subscription added together. **Active monthly value** is the subset your team actually uses at least weekly. **Duplicated monthly spend** is the amount paid for tools that solve the same job for the same person or team. If a founder pays for three general AI chatbots because each is occasionally better, that is not automatically waste. It becomes waste when there is no clear routing rule and users forget which tool is for which job.

Use this simple formula: `Total AI spend - required specialized tools = consolidation opportunity`. Required specialized tools are the products you cannot replace without losing a workflow that matters, such as an IDE extension your engineers rely on all day or a compliance-approved research product. Everything else belongs in the test bucket.

The practical goal is not to reduce AI tool spend to zero. The goal is to stop paying for multiple AI tools that do the same job badly coordinated. A clean stack usually has one shared workspace for general model work, plus a small number of specialized tools that are genuinely worth keeping.

#### When consolidation wins

Consolidation wins when one plan can replace two or more overlapping subscriptions without making the daily workflow slower. That last phrase matters. A cheaper stack that creates friction is not really cheaper, because the cost moves from the invoice to the calendar.

Use this decision rule: **if two or more paid AI tools are used for the same workflow by the same person or team, and the work can be completed in one shared workspace with acceptable quality, consolidate and monitor the result for 30 days.** If the work gets worse, bring back the specialized tool. If quality stays the same or improves, keep the simpler stack.

Here is the fast decision table:

| Scenario | What it usually means | Best next step |
| --- | --- | --- |
| One person pays for several chatbots | They are comparing models manually | Consolidate into one place and run prompts side by side |
| A team has many unused seats | Buying outpaced adoption | Remove inactive seats before buying more |
| Users copy outputs between tools | The workflow is fragmented | Move the workflow into a shared AI workspace |
| A tool is used once per month | The plan may be convenience spend | Test whether Whizi covers the occasional task |
| A tool is used daily for a specialized job | It may deserve to stay | Keep it unless a trial proves replacement is safe |

Whizi is strongest when your problem is model and workflow overlap: writing in one tool, research in another, document summaries somewhere else, and comparison work happening in browser tabs. A consolidated AI subscription works when users can test models, reuse prompts, and keep work in one place. It is not about brand loyalty. It is about reducing switching costs and making the best value AI subscription easier to justify.

The best time to consolidate is before renewal, before adding seats, or right after a team notices inconsistent AI usage. If you wait until every tool has become part of someone's private routine, cleanup gets political. If you run the calculator first, the conversation becomes concrete: here is what we pay, here is what overlaps, here is what we will test.

#### Use Whizi’s savings calculator

The calculator is the action step. Open the [AI subscription savings calculator](https://whizi.io/tools/calculator), enter each tool you currently pay for, and mark which jobs it covers. Use current vendor pricing from official pricing pages instead of relying on old screenshots, old blog posts, or memory. AI plans change often, so your own billing data is the source of truth.

A good calculator session should answer five questions: What do we pay today? Which tools overlap? Which tools are mission-critical? Which tools are occasional convenience purchases? What would the stack look like if Whizi became the shared workspace for general AI work?

Do not make the decision only on price. Add a quality check. Pick three real tasks from the last week, such as drafting a customer email, summarizing a PDF, reviewing a snippet of code, or generating a positioning outline. Run those tasks through your current stack, then run them through Whizi. Score each result on accuracy, usefulness, speed, and how much editing it needed. If the consolidated workflow is close enough or better, the savings are real.

This is the cleanest order: calculator first, workflow test second, plan selection third. Start with the [calculator](https://whizi.io/tools/calculator), then review [Whizi pricing](https://whizi.io/pricing), then [create an account](https://whizi.io/register) when the numbers and workflow both make sense. If ChatGPT Plus is the plan you are weighing against, [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) does that comparison line by line.

#### Pick a plan

After the calculator, choose a plan based on usage pattern, not optimism. A solo operator who uses AI every day needs a different setup from a team experimenting with a few workflows. A founder doing research, writing, and competitive analysis may get more value from a consolidated workspace than from several separate subscriptions with overlapping model access.

Before you pick, work through this in order. Confirm every AI subscription you currently pay for and verify the current price of each. Remove unused seats, then label every remaining tool by the job it does. Now the useful part: find the workflows where two tools overlap, choose three real tasks that sit inside those workflows, and run them side by side. Compare quality and editing time, and only then decide what to keep, cut, or consolidate.

If your stack is simple, you may only need a lighter plan. If you are replacing several paid tools, compare the total savings against the plan that gives you enough room to work without rationing prompts. The best plan is the one your team will actually use consistently.

The final move is straightforward: calculate your current spend, test the overlap, and stop paying for multiple AI tools when one workspace can carry the work. Use the calculator to make the numbers visible, then move to [pricing](https://whizi.io/pricing) and [register](https://whizi.io/register) when you are ready to turn the savings into a cleaner workflow.

Key points:

- Calculate monthly AI spend by person, workflow, and overlapping capability
- Use a simple decision rule to know when consolidation is worth testing
- Move from worksheet to Whizi calculator to pricing without getting stuck in research

Checklist:

- List every paid AI subscription and verify current pricing from billing pages
- Convert annual plans, taxes, and seat minimums into one monthly number
- Tag each tool by job: writing, coding, research, documents, images, or operations
- Mark tools that overlap with another tool used by the same person or team
- Run three real tasks through Whizi before canceling specialized tools
- Choose a Whizi plan only after comparing savings and workflow quality

FAQ:

Q: How much do AI subscriptions cost per month?
A: It depends on the tools, seats, plan tiers, and billing cycle. Verify current vendor pricing from official pricing pages or your billing settings, then use Whizi's calculator to total your real monthly spend.

Q: Should I cancel every AI tool and use one subscription?
A: Not automatically. Consolidate overlapping general AI work first, then keep specialized tools that are used daily and clearly outperform the consolidated workflow for a critical job.

Q: What is the easiest way to reduce AI tool spend?
A: Start by removing unused seats and duplicated chatbot subscriptions. Then test whether a shared workspace like Whizi can replace tools used for writing, research, documents, and model comparison.

### AI terms explained: a plain-English glossary for beginners

URL: https://whizi.io/resources/ai-terms-explained/
Updated: August 2026

Quick answer: Five AI terms carry almost everything you need at the start. A model is the trained system that answers you. A prompt is what you type. A chatbot is the conversation window. A token is roughly three quarters of a word. A context window is how much text a model holds in view at once.

A working glossary of AI terms explained without jargon: model, prompt, token, context window, LLM, hallucination, and an honest note on which words you can ignore.

#### The five words that matter in your first hour

Most writing about AI assumes a vocabulary nobody ever taught you. The useful part is small. Five words carry nearly all the practical weight, and the rest can wait months.

A **model** is the AI itself: the trained system that reads your words and produces an answer. The app is not the model. ChatGPT is an app running OpenAI's models, Gemini is an app running Google's, and Whizi is an app that holds several companies' models at once. This distinction earns its keep the first time an answer disappoints you, because switching the model is usually a bigger lever than rewording the question.

A **prompt** is just what you type. There is no secret syntax, no command list, nothing to memorize. A plain request that names who the output is for, what job it has to do, and what shape you want back will beat a clever one-liner nearly every time.

A **chatbot** is the conversation window you type into. Everything you have said inside one conversation stays visible to the model while it answers, which is why follow-ups like "make it shorter" or "more formal" work without you repeating the original request. Open a new chat and that shared memory is gone, which is what you want when the topic changes: leftover context from an unrelated question is a common source of answers that feel slightly off.

A **token** is a chunk of text, roughly three quarters of an average English word. Models read, count, and bill in tokens rather than words, so a 1,000-word document lands somewhere near 1,300 tokens. You will mostly meet the word on pricing pages and usage limits.

A **context window** is how much text a model can hold in view at once, measured in tokens. It has to cover the document you pasted, the whole conversation so far, and the answer being written. When you drop a fifty-page contract into a chat and the replies start quietly contradicting page three, you have found the edge of it. [What a context window is](https://whizi.io/resources/what-is-a-context-window) has the real numbers; the working rule is short. One long conversation about one thing, then start fresh.

Something to do with that right now: open any AI app and ask it, in your own words, to explain a term you saw this week and did not recognize. Then ask a second question, "What did you assume I already know?" That follow-up turns a definition into an explanation, and it works on any topic.

Check the answer against something else, though. These systems state wrong things in exactly the same calm voice they use for right ones, and definitions are a soft target: a plausible invented detail slides past you because you had no way to catch it. That failure has a name, and it is in the table below.

#### The rest of the glossary, and whether you need it yet

These are the words waiting for you in headlines, pricing pages, and comment sections. The third column is the honest part. Most of this is not day-one knowledge, and pretending otherwise is how beginners end up feeling stupid about something that does not affect them. When you outgrow this table, the full [AI glossary](https://whizi.io/resources/ai-glossary) defines 55 terms in the same plain English.

| Term | Plain meaning | Do you need this? |
| --- | --- | --- |
| **LLM** | Large language model, the kind of model behind every chat app you have heard of. When an article says LLM, read it as the AI | Yes, so headlines stop reading as noise |
| **Generative AI** | Any AI that produces new text, images, audio, or video instead of sorting things that already exist | Yes, one sentence is enough |
| **Training** | The long, expensive process of feeding a model enormous amounts of text so it learns patterns. It finished before you opened the app | Yes, briefly |
| **Training cutoff** | The date the training data stops. Ask about anything later and the model is guessing, unless the app searches the web | Yes, this one catches people out |
| **Hallucination** | A confident, fluent, completely invented answer. The flaw that matters most | Yes, on day one |
| **Multimodal** | The model handles more than text: photos, PDFs, audio, sometimes video | Yes, the moment you want to upload a picture |
| **Fine-tuning** | Retraining an existing model on your own data so it specializes | Not yet, this is a company project |
| **Temperature** | A dial for how random the wording is. Low is predictable, high is loose | Not yet, sane defaults are chosen for you |
| **Agent** | An AI that takes several steps on its own instead of replying once | Not yet, and see the next section |
| **API** | The plumbing developers use to wire a model into their own software | Not yet, unless you write code |
| **RAG** | Retrieval augmented generation: the model looks things up in a document store before answering | Not yet, though any app that answers from your uploaded files runs on it |
| **Open source model** | A model whose weights anyone can download and run, such as Llama or DeepSeek | Not yet, it changes little for chat users |
| **Reasoning model** | A model that works through a problem in visible steps before answering. Slower, better at math and logic | Soon, the first time a hard question fails |
| **System prompt** | Hidden instructions the app gives the model before you type anything | Not yet, useful when you build your own assistant |

Two rows deserve a footnote. Hallucination is not an occasional glitch to be patched away; it falls out of how these systems generate text, which is why [why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) is a guide of its own rather than a bullet point. And multimodal is worth ten minutes as soon as you have a photo of a receipt, a screenshot of an error message, or a scanned form to deal with. [Working with images and documents](https://whizi.io/resources/multimodal-ai) shows what that looks like on a normal Tuesday.

#### Words marketing uses that mean less than they sound like

Vendors are not lying, exactly. They are using words that once meant something specific and now mostly mean "we shipped something". Knowing the gap is what stops you paying an upgrade fee for a dropdown of canned replies.

- **Powered by AI.** Usually means one feature calls a model built by somebody else. Sometimes it means a suggestion box you can switch off. Worth asking: which model, and is it on by default?
- **Smart.** Predates this whole wave and now means very little. A smart folder is a saved search. A smart reply is three canned phrases. Look past the adjective for a verb: what does it do to your work?
- **Next-generation.** A claim in the future tense. It is perfectly compatible with a product being worse than the version it replaced, and it tends to appear when there is no measurable improvement anyone can name.
- **AI agent.** The hardest-working phrase of 2026. Used strictly, it means software that plans and carries out several steps by itself, like booking a trip end to end. Used in marketing, it usually means a chatbot with a friendlier avatar. Ask what it does when it gets stuck, and whether it can act without checking with you first.
- **AI-first.** A statement about internal strategy, not about your experience of the product. It tells you where a company points its engineering budget, which may not yet have reached the feature you were about to pay for.

One question cuts through all of it: ask what the feature does when it is wrong. Products with real AI inside have an answer, and it involves review steps, an undo, a confidence signal, or a visible source. Products carrying a sticker go quiet.

#### Why model names look like license plates

A model name has two parts doing two different jobs. The word is the family, the product line a company keeps developing: GPT from OpenAI, Claude from Anthropic, Gemini from Google, Llama from Meta. The numbers and the suffixes are the version, and often the size, so one family will ship a small quick member and a large careful one on the same morning. Reading a name is mostly a matter of separating those halves and ignoring the marketing in between.

You may have noticed this article does not name a single version number. That is deliberate. Versions change every few weeks, and any glossary that pins them down is wrong by the next season. What does not change is the shape of the choice. Small models are fast and cheap and completely fine for tidying an email. Large ones are slower and better with long documents and hard reasoning. Each family also has a personality that shows up in tone and in how strictly it sticks to a source you gave it.

The newest release is not automatically right for your task either. Newer usually means better benchmark scores, and benchmarks measure exam-style problems rather than whether the writing sounds like you. Plenty of people find that an older, cheaper model writes their emails better, and they save the expensive one for the two hard jobs a week. You only learn this by running the same real task through two or three models and reading the outputs next to each other, which is what [choosing an AI model](https://whizi.io/resources/how-to-choose-ai-model) walks through step by step. Doing that without paying three providers is the whole reason Whizi puts GPT, Claude, Gemini, and image models inside one subscription.

#### Five terms worth learning at the three-month mark

By month three you will have walked into a few walls. These five words are the names of those walls, which is why they are worth learning then and not now.

1. **Reasoning model.** Starts mattering the first time a model confidently botches a multi-step calculation or a logic puzzle. Switching to a reasoning model often fixes it outright, at the cost of waiting longer for the answer.
2. **System prompt.** Matters when you notice you are pasting the same background paragraph into every new chat. Saving it once as a standing instruction is the fix.
3. **RAG.** Matters when you want answers grounded in your own files rather than the open internet, for example a folder of policies nobody has read since 2019.
4. **Temperature.** Matters when output feels either too samey across attempts or too erratic, and you would rather turn a dial than keep rerolling.
5. **Fine-tuning.** Matters almost never for an individual, and it is worth knowing why: it costs real money, needs hundreds of clean examples, and a good prompt on a good model beats it for nearly every use people imagine it for.

If you want the layer underneath all of this, [what AI actually is](https://whizi.io/resources/what-is-ai) covers the mechanism without the math, and [how to talk to AI](https://whizi.io/resources/how-to-talk-to-ai) turns that prompt paragraph from the first section into a habit. Leave this page open in a tab for a week. The vocabulary stops being intimidating around the fourth time you meet a word you already know.

Key points:

- Five words carry almost everything you need in your first week with AI: model, prompt, chatbot, token, and context window
- Half of the AI glossary is genuinely not your problem yet, and this guide marks which terms you can skip without falling behind
- A model name is a family plus a version, and the newest version is not automatically the right one for your task

Checklist:

- Learn model, prompt, chatbot, token, and context window before anything else.
- Treat a disappointing answer as a reason to change the model, not only the wording.
- Give every prompt an audience, a job, and the format you want back.
- Start a fresh chat when a long conversation begins contradicting itself.
- Verify any definition or fact you plan to repeat to somebody else.
- Ask any vendor which model powers their AI feature and what happens when it is wrong.
- Test the same real task on two models before deciding which one you trust.

FAQ:

Q: What does LLM mean?
A: LLM stands for large language model, which is the type of AI behind ChatGPT, Claude, Gemini, and every other chat app you have heard of. It is a system trained on huge amounts of text that predicts useful language in response to yours. When an article says LLM, you can read it as the AI and lose nothing.

Q: What is the difference between a model and an app like ChatGPT?
A: The model is the trained system that produces the answer. The app is the interface around it, with the chat window, file uploads, history, and billing. One app can offer several models, which is why swapping models inside the same app often changes your results more than switching apps does.

Q: Do I really need to understand tokens and context windows?
A: You need one sentence about each, not a study session. A token is roughly three quarters of a word and is how usage gets measured and billed. A context window is how much text the model can hold in view at once, which explains why very long chats start losing track of what you said at the beginning.

Q: What does it mean when people say AI hallucinates?
A: It means the model produced something fluent, confident, and false: a citation that does not exist, a statistic with no source, a feature a product never had. It happens because these systems generate plausible language rather than looking up verified facts. Treat names, numbers, dates, and quotes as unchecked until you have checked them.

Q: How do I keep up when model names change every month?
A: You do not have to. Learn the families, which are stable, and treat version numbers as details you look up when you are choosing rather than facts you memorize. When a new version appears, the only question that matters is whether it does your specific task better, and you answer that by running the task, not by reading the announcement.

### AI vs Google search: a simple rule for which box to type into

URL: https://whizi.io/resources/ai-vs-google-search/
Updated: August 2026

Quick answer: Use Google search when the answer is a fact that exists in the world and could change: prices, hours, news, official forms, medical and legal specifics, and anything you will have to show someone. Use AI when the answer has to be built around something only you have, or will not compress into keywords.

A practical rule for choosing between AI and Google search: the one difference that explains all the others, what search still wins at, what a chatbot does better, and a task-by-task table.

#### One difference, and the rest follows from it

A search engine finds pages that already exist. Somebody wrote them before you showed up, and Google's job is to point at the ones most likely to contain your words. A chatbot does something else entirely. It writes a fresh answer, piece by piece, in response to your specific question. That answer did not exist anywhere five seconds ago and will not exist anywhere once you close the tab.

Every other difference people argue about comes out of that one. Search shows you sources because sources are the only thing it has. AI does not show them because it is not quoting pages, it is producing new text shaped by patterns it absorbed during training. Search is current because the live web is current. AI is often behind, because its training data stopped on a date months ago. Search makes you read five pages and work out which one to believe. AI does that work for you, and occasionally does it wrong, confidently, in a nicely organized paragraph.

Before you read any further, run one test. Take a question you genuinely typed into Google this week and paste it into a chatbot word for word, clumsy keywords and all. Then look at both answers and ask which one you would have had to work harder to actually use. That two-minute comparison will teach you more than the rest of this page. If you have never opened a chatbot at all, [what AI is](https://whizi.io/resources/what-is-ai) is the short version to read first.

#### What search still wins, stated without hedging

There is a set of questions where a chatbot is simply the wrong tool, and no clever prompting fixes it. Use search for these, every time, without thinking about it.

- **Anything happening now.** Scores, elections, outages, whether the train is running, whether the company your friend mentioned really did that thing. Models are trained up to a cutoff date and then frozen.
- **Prices, stock, and availability.** A chatbot will cheerfully tell you what a flight or a laptop costs. The number is a plausible memory of an old price, not today's price.
- **Official forms and government pages.** Tax deadlines, visa requirements, benefit eligibility, appointment booking. You want the real page on the real government domain, because the wording on that page is the rule.
- **Opening hours, addresses, phone numbers.** Small facts that change quietly and that nobody publishes a correction for.
- **Medical and legal specifics.** General explanation from AI is fine. Dosages, interactions, statutes, filing deadlines, anything you would act on for your own body or your own case, belongs to a source with a name and a date on it.
- **Anything you will have to show someone.** If a colleague, a doctor, or a landlord is going to ask where you got that, you need a link, not a paraphrase.

The thread running through that list is verifiability rather than accuracy. A chatbot is right about most things most of the time. The trouble is that when it is wrong it looks exactly the same as when it is right: same tone, same fluency, same air of having checked. Search hands you a domain name, a publication date, and a page you can read with your own eyes, and those three things are what let you decide instead of hope. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) explains the mechanism behind the confident mistakes, and it is worth ten minutes before you lean on a chatbot for anything that matters.

#### What AI wins, and why keywords are the tell

Search has a requirement people rarely notice: you need to already know roughly what the answer looks like, because you have to guess the words that will appear on the page containing it. That works for "weather madrid" and falls apart for "what do I say to my landlord about the boiler without sounding aggressive." The moment your question refuses to compress into keywords, you have found the boundary between the two tools.

- **Questions with your own situation baked in.** Two job offers with different commutes. An error message plus the four things you already tried. A paragraph you wrote that does not quite work and you cannot say why.
- **Explain this at my level.** Ask for the version aimed at someone who has never met the topic, then ask for the next level up, then ask what you still have not understood. A web page cannot adjust itself to you. A chatbot will do it four times without getting bored.
- **Turning something you have into something you need.** A forty-page PDF into six bullets. Scribbled notes into a polite email. Ninety survey comments into three themes. None of this is a search task, because the source material is yours and exists nowhere online.
- **Comparing options against your circumstances.** Not "best dishwasher" but "which of these three, for a small flat with old plumbing and someone who runs it at 11pm."
- **The nine-tabs problem.** When the honest search answer is "open nine pages, read all of them, and work out what they add up to," a chatbot is doing precisely the job you were about to do by hand.

That last one is what changes people's habits, and it is also where the tradeoff bites hardest, because the summary arrives without the nine pages behind it. One habit helps more than any prompt trick: ask the chatbot which parts of its answer it is confident about and which parts you should check yourself. It will not be perfect at sorting those, and it is still far better than an answer with no confidence markers at all. [Everyday tasks AI is good at](https://whizi.io/resources/ai-for-everyday-tasks) has a longer list with prompts you can copy.

#### The decision table for an ordinary week

Here is the same split applied to things people actually do. Read the reason column rather than just the verdict, because the reason is what transfers to the task you have that is not on this list.

| Task | Use | Why |
| --- | --- | --- |
| Check whether a shop is open tonight | Search | Hours change quietly and no model was told |
| Understand a term in your insurance letter | AI | You need it explained for your policy, not defined in general |
| Find the official form to change your address | Search | You need the real government page, not a description of it |
| Draft the covering email that goes with that form | AI | Blank-page work, and there is no fact in it to verify |
| Compare two phones you are choosing between | Both | AI to frame the tradeoffs, search for current price and stock |
| Get the price of anything at all | Search | Any number a chatbot gives you is a memory, not a quote |
| Summarize a long PDF someone sent you | AI | No web page exists for a document only you have |
| Look up a drug interaction | Search | Read it on a source with a name and a date attached |
| Plan a week of meals around what is in your fridge | AI | Your constraints, not a recipe someone else published |
| Find out what happened in the news today | Search | Nothing about today is in any training set |
| Rewrite a blunt paragraph so it lands softer | AI | There is no page anywhere that contains your paragraph |

Two patterns hide in that table. If the answer is a fact that exists out in the world and could change, search. If the answer has to be constructed around something only you have, AI. Nearly every case that feels genuinely ambiguous turns out to be a "both" case, and both almost always means AI first for the shape of the answer, search second for the parts you are going to act on. If the problem is instead that the chatbot keeps missing what you meant, that is a phrasing issue, and [how to talk to AI](https://whizi.io/resources/how-to-talk-to-ai) clears up most of it.

#### The workflow you land on after a month

Nobody who uses both keeps consulting a rule. What happens instead is that the two tools settle into different positions in the same task. AI goes at the front, while you are still working out what you are dealing with and what your options are. Search goes at the back, aimed at the specific claims you are about to act on: the price you will pay, the deadline you could miss, the dose you would take, the address you would drive to. Understanding and drafting on one side, verification on the other.

There is a middle option worth knowing about. AI search tools, Perplexity being the best known, along with the AI answers now sitting on top of Google results, work the other way round: they search the live web first, then write an answer with links to what they found. That repairs the freshness problem and gives you something to click. The tradeoff is real, though. The answer is only as good as the handful of pages that happened to get retrieved, a citation next to a sentence does not guarantee the sentence matches the source, and a confident summary of three mediocre pages reads exactly like a confident summary of three excellent ones. Click the links. [Perplexity compared with ChatGPT](https://whizi.io/resources/perplexity-vs-chatgpt) walks through where each one lands.

A setup for your next week: keep your search engine exactly where it is and open a chatbot in a second tab, alongside rather than instead of. Send it the questions you would have given up on, the ones too specific or too tangled to type as keywords. If you want more than one model in that tab, Whizi holds GPT, Claude, Gemini and image models under one subscription instead of one account per provider, and running the same tangled question through two of them is the fastest way to spot the parts neither one is sure about. Then keep one small habit running underneath all of it. Before you act on anything a chatbot told you, ask whether you would be comfortable showing someone where it came from. If the answer is no, that question was always a search.

Key points:

- A search engine points you at pages that already exist, while a chatbot writes an answer that did not exist until you asked, and every other difference follows from that
- Search still wins outright on anything current, any price, any official page, and anything you will need to show to someone else
- AI wins when your question will not compress into keywords, or when the answer has to be built around something only you have

Checklist:

- Send anything current to search: prices, hours, news, stock, availability.
- Send anything with your own situation baked in to a chatbot.
- Ask AI to explain at your level, then ask for the next level up.
- Verify every name, number, date, and price on a real page before acting.
- Open the citations in AI search tools instead of trusting the summary.
- Give documents you already hold to AI, never to a search box.
- Keep both tabs open and let each task decide which one you type into.

FAQ:

Q: Is AI better than Google search?
A: Neither is better in general, because they do different jobs. Search points you at existing pages, which is what you want for anything current, official, or verifiable. AI writes a new answer around your specific situation, which is what you want when the question will not fit into keywords.

Q: Should I use AI or Google for research?
A: Use AI to map the territory: what the main positions are, what terms you do not know, what questions you should be asking. Then use search to confirm the specific facts, figures, and quotes you plan to rely on. Researchers who skip the second half eventually publish something a chatbot invented.

Q: Will AI replace search engines?
A: Not while people need to see the source. Google has already put AI answers at the top of its results and Perplexity built a whole product around it, so the two are merging rather than one killing the other. What is changing is where you land: more answers up front, fewer blue links clicked, and a bigger burden on you to check what the summary skipped.

Q: Why does AI give wrong answers when search finds the right one?
A: A chatbot predicts plausible text rather than looking facts up, so when it does not know something it produces something that sounds correct instead of stopping. Search cannot do that, since it can only show pages that exist. This is why prices, dates, and names should always be checked against a page, even when the chatbot sounds certain.

Q: When should I use AI instead of search?
A: Use AI when you would otherwise open nine tabs and work out what they add up to yourself, when the material is a document only you have, or when you need something explained at your level rather than defined. Use it for drafting, comparing options against your circumstances, and untangling questions you cannot phrase as keywords. Everything time-sensitive or officially binding still belongs in the search box.

### AI workflow templates: 7 repeatable workflows

URL: https://whizi.io/resources/ai-workflow-templates/
Updated: August 2026

Quick answer: You can get AI workflow templates on this page, which collects seven ready-to-use workflows plus Whizi's copy-ready template packs. Each template names its inputs, the prompt with constraints, the output format, and a QA check. The seven cover meeting notes, email triage, competitor scans, PDF extraction, code review, daily planning, and content repurposing.

Use seven repeatable AI workflow templates for meeting notes, email triage, competitor scans, PDF summaries, code review, planning, and content repurposing.

#### Workflow 1: meeting notes to action plan

Use this workflow after calls, sales conversations, and planning sessions. The goal is not just a summary. The goal is a decision-ready action plan with owners, dates, open questions, and follow-ups.

**Inputs -> prompt -> output -> QA**. Every workflow in this guide uses the same four stages. If the prompt stage itself is where you are unsure, the [prompt engineering for beginners](https://whizi.io/resources/prompt-engineering-beginners) framework covers how to write the instruction, context, constraints, and format that each template below assumes.

| Stage | What you supply | Why it matters |
| --- | --- | --- |
| Inputs | The real material: a transcript, an inbox, source pages, a diff, a task list, or a draft | Vague requests let the model infer owners, dates, and claims you never gave it |
| Prompt | Role, context, constraints, and the exact output format | Constraints are what keep the model away from private data and decisions that need human judgment |
| Output | A named structure: action table, priority queue, extraction table, findings list, schedule, or content kit | A named format is far easier to check than paragraphs of prose |
| QA | A check that traces each item back to the source before you act | Names, dates, numbers, and commitments still need a human to approve them |

**Ready-to-run template packs.** The seven workflows below are the general shapes. These two packs are the filled-in versions, already written out prompt by prompt.

| Template | What it does | Best for |
| --- | --- | --- |
| [Founder Research Stack](https://whizi.io/templates/founder-research-stack) | Nine copy-ready prompts for market maps, competitor teardowns, pricing structure, positioning, and customer language mining, ending in a landing page draft | Founders, consultants, and operators deciding where to aim and what to charge |
| [Claude Coding Prompt Pack](https://whizi.io/templates/claude-coding-prompts) | Twelve copy-ready prompts for debugging, code review, refactors, architecture reviews, and test design, with the constraints that stop a small fix becoming a rewrite | Developers who want a reviewable diff instead of a rebuilt file |

Inputs: transcript, rough notes, attendee list, meeting purpose, known deadlines, and any decisions you already trust. If the transcript is long, label sections by topic or timestamp before prompting.

Prompt: "Turn these meeting notes into an action plan. Context: The meeting purpose was [purpose]. Attendees were [names]. Notes/transcript: [paste]. Constraints: Do not invent owners or dates. Mark unclear items as needs confirmation. Separate decisions from discussion. Format: summary, decisions, action items table, risks, follow-up message."

Output: a five-part brief with a short executive summary, confirmed decisions, action table with owner/date/status, unresolved questions, and a ready-to-send follow-up email.

QA: Check every action item against the transcript. Remove anything the AI inferred too strongly. Confirm dates and owners before sending. If the output is too broad, rerun the prompt with one extra constraint: "Only include actions that someone explicitly agreed to do."

#### Workflow 2: email triage to reply queue

This workflow helps you process a crowded inbox without letting the AI speak for you unchecked. The useful output is a priority queue, not a batch of robotic replies.

**Inputs -> prompt -> output -> QA**

Inputs: sender, subject line, email body, relationship context, urgency, and your preferred response style. For sensitive threads, remove private data before pasting and ask the model to draft rather than send.

Prompt: "Triage these emails for today. Context: My role is [role], my priorities are [priorities], and my tone should be [tone]. Emails: [paste]. Constraints: Do not answer questions that require private account data. Flag anything that needs human judgment. Format: priority table, recommended action, draft reply for each email that can be safely answered."

Output: a table with priority, reason, next action, suggested reply, and risk flag. Billing issues, account access, and promises should be flagged for manual review.

QA: Read the priority reason first. If the model cannot explain why an email is urgent, demote it. Check all names, dates, promises, prices, and attachments. Keep a personal rule: AI can draft replies, but humans approve commitments.

#### Workflow 3: competitor scan to research brief

Use this as an AI workflow for research when you need a fast but traceable view of a market without drowning in tabs.

**Inputs -> prompt -> output -> QA**

Inputs: competitor names, website notes, pricing snippets, product pages, review excerpts, target customer, and the decision you are trying to make. You can also start from Whizi's [Founder Research Stack](https://whizi.io/templates/founder-research-stack) when you want a more complete template.

Prompt: "Create a competitor scan for [market]. Context: I am deciding [decision]. Competitors and notes: [paste]. Constraints: Use only the supplied notes. Separate facts from inference. Do not invent pricing, customers, or features. Format: comparison table, positioning patterns, gaps, risks, and next research questions."

Output: a comparison table with competitor, audience, promise, features, pricing notes, proof points, weaknesses, and inferred positioning. The best version also includes a "what to verify next" column so the research does not pretend to be final.

QA: Trace every claim back to a note or source excerpt. Highlight unsupported inferences. For important decisions, run the same prompt in another model and compare which output separates evidence from opinion more cleanly.

#### Workflow 4: PDF summary to verified extraction

Generic PDF summaries are risky because they can miss caveats or overstate conclusions. This workflow asks for extraction and verification, not just a pleasant summary.

**Inputs -> prompt -> output -> QA**

Inputs: the PDF text or uploaded document, document type, your purpose, sections that matter most, and the exact output you need. If the document is long, ask for section-level summaries before asking for synthesis.

Prompt: "Analyze this document for [purpose]. Context: The document type is [type] and I care most about [topics]. Constraints: Quote or cite section names when possible. Mark uncertain items. Do not summarize sections you cannot see. Format: one-page summary, key facts table, risks/caveats, extracted dates/numbers/names, and verification checklist."

Output: a compact summary plus a structured extraction table. A research paper table might include claim, evidence, method, limitation, and confidence. A contract table might include obligation, party, deadline, section, and risk.

QA: Search the original PDF for every number, deadline, named entity, and quote. Ask a second pass: "List five ways this summary could be misleading." If the answer affects legal, medical, financial, or employment decisions, treat the AI output as a reading aid, not advice.

#### Workflow 5: code review to safer change list

This AI workflow for coding is useful when you have a diff, pull request, or file you want reviewed. The point is to reduce risk, not outsource judgment.

**Inputs -> prompt -> output -> QA**

Inputs: code diff, surrounding function or module, expected behavior, test output, error messages, constraints, and what should not change. Never paste secrets, production credentials, or private customer data.

Prompt: "Review this code like a senior engineer. Context: Expected behavior is [behavior]. Change goal is [goal]. Code/diff: [paste]. Constraints: Prioritize bugs, regressions, security issues, edge cases, and missing tests. Avoid style-only feedback unless it hides a bug. Format: findings table with severity, file/function, issue, why it matters, suggested fix, and test to add."

Output: a ranked review list, proposed tests, and a minimal change plan. For refactors, ask for the smallest readability improvement that preserves behavior, then list tests that prove it.

QA: Reproduce every bug before fixing when possible. Do not accept a suggested patch without reading it. Run tests, lint, and type checks locally. If two models disagree, use that disagreement as a review checklist rather than a vote.

#### Workflow 6: daily planning to focused schedule

A daily AI routine should help you choose what not to do. This workflow turns a messy list into a focused plan with realistic blocks, tradeoffs, and a shutdown checkpoint.

**Inputs -> prompt -> output -> QA**

Inputs: task list, calendar constraints, deadlines, energy level, meetings, must-do items, optional items, and one strategic goal. The more honest the inputs, the better the plan.

Prompt: "Plan my day around outcomes, not busyness. Context: Today's available work blocks are [blocks]. Must-dos are [must-dos]. Optional tasks are [optional]. Energy level is [energy]. Strategic goal is [goal]. Constraints: Protect deep work, include buffers, avoid overcommitting, and explain tradeoffs. Format: schedule, top 3 outcomes, defer list, risk list, shutdown checklist."

Output: a time-blocked schedule, three outcomes, a deferred list, and a final 10-minute shutdown routine. The defer list matters because it prevents the AI from making an impossible day look neat.

QA: Check whether the plan fits your actual calendar. Cut 20 percent if it feels too full. If a task has no success definition, ask the model to define "done" before you start.

#### Workflow 7: content repurposing to publishing kit

This AI workflow for writing turns one useful source into multiple publishable assets without making everything sound the same.

**Inputs -> prompt -> output -> QA**

Inputs: source content, audience, channel list, brand voice sample, claims that must stay accurate, claims to avoid, and the desired formats. Include examples of your best existing content if you want less generic output.

Prompt: "Repurpose this source into a publishing kit. Context: Audience is [audience]. Channels are [channels]. Voice sample: [paste]. Source: [paste]. Constraints: Preserve facts, avoid hype, do not add unsupported statistics, and make each channel native to its format. Format: core message, 5 post ideas, email draft, short social posts, long social post, newsletter section, and QA table."

Output: a content kit with reusable angles, channel-specific drafts, and a QA table listing original claim, reused claim, source location, and risk.

QA: Compare each draft to the source. Remove claims that were added for drama. Check whether the email, post, and article intro have different jobs instead of repeating the same sentence. Save the prompt as a repeatable ai prompt template only after it works on a real piece of content.

Key points:

- Seven concrete AI workflows you can run with real inputs, outputs, and QA checks
- Copy-paste prompt templates for meetings, email, research, PDFs, coding, planning, and repurposing
- A reusable inputs to prompt to output to QA structure for building a daily AI routine

Checklist:

- Start each AI workflow with real inputs, not a vague request.
- Define the output before running the prompt: table, checklist, memo, draft, plan, or extraction.
- Add constraints for evidence, privacy, tone, length, and uncertainty.
- Ask for a QA step as part of the output, not after the fact.
- Use Whizi to save repeatable prompts by workflow instead of by model.
- Run important workflows across more than one model and compare accuracy, structure, and cleanup time.
- Update each template after using it on a real task.

FAQ:

Q: What is an AI workflow template?
A: An AI workflow template is a repeatable process that defines the input, prompt, expected output, and quality check for a specific task such as summarizing PDFs, reviewing code, or triaging email.

Q: How do I build an AI workflow for work?
A: Start with one recurring task, list the inputs it needs, write a prompt with constraints and output format, then add a QA checklist that verifies facts, decisions, names, numbers, and next steps.

Q: Should I use the same AI model for every workflow?
A: No. Different models can perform better on different tasks. For important workflows, run the same prompt across models and keep the version that gives the most accurate, usable output.

### The best AI apps for beginners: what to actually install first

URL: https://whizi.io/resources/best-ai-apps-for-beginners/
Updated: August 2026

Quick answer: ChatGPT is the safest first AI app for most beginners, with Gemini better if you live in Gmail and Docs, Claude better for writing quality, and Perplexity better for research with sources. A first app needs three things: easy chat, file uploads, and a strong default model.

An honest guide to the best AI apps for beginners: ChatGPT, Claude, Gemini, Perplexity, and Whizi compared by task, free tier, and price, with a simple decision path.

#### What beginners actually need from an AI app

Search for the best AI app and half the results will try to sell you app builders, developer platforms, or lists of 40 tools you will never open. If you are choosing your first AI app, ignore all of that. You need exactly three things: a chat interface you can type into like a messaging app, the ability to upload a document or photo, and access to at least one genuinely strong model without settings to configure.

Nice to have, but not essential on day one: voice input for using it while walking, image generation for the occasional visual, web search for current events, and the ability to switch between models as your tasks get more serious.

What you should not pay attention to as a beginner: benchmark scores, context window sizes, and arguments on social media about which model is smartest this month. The differences that matter for you are practical, such as which app writes better emails in your language, and you can only discover those by using them on your own tasks. If the vocabulary itself is what is slowing you down, the [AI glossary](https://whizi.io/resources/ai-glossary) defines the terms one sentence at a time.

#### The short list: five apps worth considering

These five cover what beginners realistically choose between. All have mobile apps and browser versions, and all are easy enough to use in the first five minutes.

- **ChatGPT (OpenAI).** The household name and the safest default, explained from scratch in [what ChatGPT is and how it works](https://whizi.io/resources/chatgpt). Strong all-rounder, huge feature set, and the largest ecosystem of guides. The free tier is fine for light use but throttles you to weaker models at busy times.
- **Claude (Anthropic).** Loved for natural writing and careful work with long documents. Many people find its tone the most human. The free tier is comparatively tight, and heavy users hit limits quickly.
- **Gemini (Google).** Best if you live in Google's world: it connects with Gmail, Docs, and Drive, and its free tier is generous. Quality is strong, and it is often the most convenient choice on Android phones.
- **Perplexity.** Built around answering questions with sources from the live web. Excellent as a search replacement and for research, less of a fit for writing, planning, and document drafting.
- **Whizi.** A different category: one app that holds the major models, including GPT, Claude, and Gemini, with side-by-side comparison. Useful when you are new, because you do not have to guess which one suits you before you have tried them.

#### Honest comparison: task by task

App-versus-app arguments are mostly noise. Task-versus-task is what you will feel day to day. This is how the options line up in practice for a beginner:

| Task | Strong choices | Notes |
| --- | --- | --- |
| Everyday questions and drafts | ChatGPT, Gemini, Whizi | All handle this well; differences are small |
| Natural-sounding writing | Claude first, then others | Tone quality is Claude's signature strength |
| Long documents and PDFs | Claude, Whizi | Careful reading and steady terminology |
| Research with sources | Perplexity, Gemini | Live web answers with citations |
| Working in Google apps | Gemini | Native Gmail, Docs, and Drive hooks |
| Finding which model fits you | Whizi | Same prompt across all models, side by side |

Prices move, so check current pages before deciding, but the structure is stable: each provider's serious tier costs roughly 20 dollars per month, and the strongest models sit behind those paid tiers. The pattern that surprises beginners is how quickly you end up wanting two of them, usually one for writing and one for research or documents, and paying 40 dollars per month for two chat apps is where most people balk. That exact moment is what a bundled app solves; our [AI subscription cost breakdown](https://whizi.io/resources/ai-subscription-costs) does the math in detail, and the [ChatGPT alternatives that include Claude and Gemini](https://whizi.io/alternatives/chatgpt) page shows what one plan covers instead of two. If you want the whole category rather than one product, the [best all-in-one AI platforms](https://whizi.io/resources/best-all-in-one-ai-platforms) guide groups them by how they actually work.

#### The free-tier trap, explained kindly

Free tiers are real and useful, and every app on the short list has one. Use them for your first week. Just know the three ways they bite.

First, model quality: free tiers typically serve smaller or older models, or drop you to them when demand is high. If you judged AI by a free tier on a busy afternoon, you have not seen what the good models can do. Second, usage caps: the moment AI becomes a habit, you will hit daily or hourly message limits, usually mid-task. Third, feature gates: file uploads, image generation, and voice often sit partly or fully behind the paid tier.

A sensible path: spend week one on free tiers to learn what you use AI for. Then, instead of guessing which single provider deserves your 20 dollars, use a trial of a multi-model app to run your real tasks across every major model and let the results decide. That is precisely the experiment [Whizi's trial](https://whizi.io/register) exists for, and if you decide one single-provider app fits you best afterward, you will have chosen it on evidence.

#### A simple decision path

If you want an answer rather than a comparison, follow this path from top to bottom and stop at the first line that describes you.

1. You mainly want answers to questions with sources: start with **Perplexity**.
2. Your life runs on Gmail and Google Docs, especially on Android: start with **Gemini**.
3. You mostly write, and how the writing sounds matters to you: start with **Claude**.
4. You want the safest all-round default everyone can help you with: start with **ChatGPT**.
5. You want to find out which model actually fits your tasks, or you already know you will use AI for several kinds of work: start with **Whizi**, because testing all of them costs less than subscribing to two of them.

Stuck between the two most popular doors? The [Claude vs ChatGPT comparison](https://whizi.io/resources/claude-vs-chatgpt) settles it task by task, and if budget is the deciding factor, the [free ChatGPT alternatives guide](https://whizi.io/resources/free-chatgpt-alternatives) ranks what stays genuinely useful without paying.

Whichever door you enter through, the first-week routine is the same: bring real tasks, not trivia questions. Rewrite a real email, summarize a real document, plan a real week. The guide to [using AI in everyday life](https://whizi.io/resources/ai-for-everyday-tasks) gives you a seven-day plan with copy-paste prompts, and [how to use AI](https://whizi.io/resources/how-to-use-ai) covers the habits that keep quality high.

One last reassurance: this is not a marriage. Switching AI apps costs nothing but a login. Pick one today, use it on real work for a week, and let your own results, not anyone's ranking, tell you whether to stay.

Key points:

- For a first AI app you want three things: easy chat, file uploads, and access to a strong model without configuration
- ChatGPT, Claude, Gemini, and Perplexity are all good first apps with different strengths; the real cost appears when you outgrow the free tiers
- A multi-model app like Whizi lets you test every major model for the price of one subscription, which is the cheapest way to find your fit

Checklist:

- Choose an app with simple chat, file uploads, and a strong default model.
- Ignore benchmark arguments; test apps on your own real tasks.
- Use free tiers for week one to discover your actual use cases.
- Expect free tiers to limit model quality, usage, and features.
- Compare at least two models on your top tasks before paying anyone.
- Count the total monthly cost before stacking single-provider subscriptions.
- Pick the app that fits your main task, and switch freely if it stops fitting.

FAQ:

Q: What is the best AI app for a complete beginner?
A: ChatGPT is the safest single default, Gemini is best for Google users, and Claude is best for writing quality. If you would rather test all of them on your own tasks and then decide, a multi-model app like Whizi is the cheapest way to run that comparison.

Q: What is the best free AI app?
A: Gemini currently has one of the most generous free tiers, and ChatGPT and Perplexity are both usable free for light tasks. All free tiers limit model quality or daily usage, so judge any app by its paid model before concluding AI is not useful.

Q: Should I pay for an AI app?
A: Pay once AI saves you time weekly, because paid tiers give you the strongest models and higher limits. Before paying roughly 20 dollars per month to a single provider, check whether a bundled multi-model subscription covers more of your tasks for a similar price.

Q: Do these AI apps work on iPhone and Android?
A: Yes. ChatGPT, Claude, Gemini, Perplexity, and Whizi all work on both platforms through apps or the mobile browser. Gemini integrates most deeply with Android, while the others work essentially the same on either phone.

Q: Is Whizi a model or an app?
A: Whizi is an app, not a model. It gives you the major models from different providers, including GPT, Claude, and Gemini, inside one workspace with one subscription, so you can use and compare them without managing separate accounts.

### Best AI image generators in 2026: how to choose

URL: https://whizi.io/resources/best-ai-image-generators/
Updated: August 2026

Quick answer: The best AI image generator is the one that produces a usable asset for your specific job, so test Flux, Stable Diffusion, and GPT Image against the same creative brief. Score each on style fit, prompt following, variation quality, editability, consistency, speed, and whether the provider terms allow your intended use.

Compare Flux, Stable Diffusion, GPT Image, and other AI image generators with practical criteria, prompts, and workflows.

#### Evaluation criteria: style, control, speed, and rights

The best AI image generator is not simply the model that makes the prettiest first image. It is the one that gives you a usable asset for the job in front of you: an ad variation, a thumbnail, a product mockup, a blog header, a concept board, a social post, or an edit to something you already have.

So the useful question is not "which model is best?" It is "which model gives me the right combination of style, control, speed, editability, and licensing for this particular job?"

Start with five criteria. **Style** measures whether the model can match the look you need: photoreal, editorial, 3D, illustration, product render, poster, or concept art. **Control** measures prompt following, layout, aspect ratio, text instructions, brand cues, and revisions. **Speed** matters when you need options fast. **Editing** matters when you need to preserve part of an image while changing another part. **Rights and platform terms** matter because usage rules differ. Do not assume a universal commercial-use rule. Check the license, provider terms, and platform terms for the exact tool you use.

Use this scorecard before choosing a subscription or workflow. Give each category 1 to 5 points, then add a short note about why you scored it that way.

| Criterion | What to test | Why it matters |
| --- | --- | --- |
| Style fit | Can it create the visual language your audience expects? | A beautiful image is still wrong if it does not fit the brand or channel |
| Prompt following | Does it obey subject, composition, aspect ratio, text, and constraints? | Cleanup time destroys the value of fast generation |
| Variation quality | Are the 5th and 10th options still useful? | Creative work usually needs range, not one lucky output |
| Editability | Can you revise details without rebuilding the whole image? | Most production assets need iteration |
| Consistency | Can it keep product shape, character, colors, or layout stable? | Campaigns and product visuals need continuity |
| Speed and cost | How many usable options can you produce per hour or budget? | Teams need throughput, not only peak quality |
| Terms fit | Do the provider terms allow your intended use? | Rights questions must be checked before publishing |

Official docs reinforce the workflow differences. OpenAI documents image generation for creating and editing images from prompts and image inputs. Black Forest Labs documents prompt-based generation through the BFL API. Stability AI announced the Stable Diffusion 3 API through its developer platform and partners. Use those anchors, then test with your own brand, product, and channel constraints.

#### Model comparisons

Flux, Stable Diffusion, and GPT Image are often compared because they represent three different buying and workflow patterns. Flux models from Black Forest Labs are commonly associated with high-quality prompt-based image generation and strong visual aesthetics. Stable Diffusion remains important because of its ecosystem, customization culture, and open-model history, with commercial API options available through Stability AI. GPT Image is compelling when teams already work in an OpenAI-style chat or API workflow and want image generation or image editing connected to broader prompting.

The practical comparison is not a permanent ranking. Model lineups, limits, pricing, and your tasks change. A model that wins for stylized ad concepts may not win for editing a product photo. Test by workflow, not by screenshots on social media.

| Model family | Often worth testing for | Watch-outs | Best first workflow |
| --- | --- | --- | --- |
| Flux / BFL | High-quality text-to-image concepts, polished visual styles, campaign exploration | Check provider terms, current model options, and whether your tool supports the controls you need | Generate 10 creative directions from one strict brief |
| Stable Diffusion / Stability AI | Flexible image generation, ecosystem depth, teams that value control and model choice | Quality and usability vary by model, host, workflow, and settings | Build a repeatable visual style test across prompts and references |
| GPT Image | Prompt-led image generation and editing inside broader AI workflows | Validate text accuracy, brand fit, and editing behavior on your real assets | Revise or generate images while keeping copy, campaign, and concept notes together |
| General design tools | Quick social graphics, templates, nontechnical workflows | May hide model details or limit fine control | Turn a rough prompt into a finished channel asset |
| Local/open workflows | Maximum control, experimentation, custom pipelines | Setup, hardware, governance, and support burden | Advanced teams with clear production needs |

For most people the buying decision is not "Flux or Stable Diffusion or GPT Image, forever." It is "which one do I run first for this task, and which one do I use for the variants and edits?"

Test on the things that actually break: hands, faces, product shape, typography, logos, lighting, colour consistency across a set, exact aspect ratio, room left for text, mobile crops, and whether a revision keeps the parts you liked.

#### Prompting basics

AI image prompting gets better when you stop asking for "a cool image" and start writing a creative brief. A useful prompt tells the model what the asset is for, who it is for, what must be visible, what style and composition to use, and what to avoid. [How to create AI images](https://whizi.io/resources/image-generation) is the beginner version of that brief, built around the six things a description has to name.

Use this prompt structure: **subject + purpose + composition + style + constraints + output notes**. Subject is the main thing in the image. Purpose is the channel or job. Composition controls framing, camera angle, background, negative space, and where text may go. Style covers visual language. Constraints define what must not change, what should not appear, and what needs to remain realistic.

**Base prompt template**

`Create [number] image concepts for [use case]. Subject: [specific subject]. Audience: [audience]. Composition: [framing, camera angle, focal point, negative space]. Style: [photoreal/editorial/3D/illustration/etc.]. Lighting and color: [details]. Must include: [required visual elements]. Avoid: [things to exclude]. Brand constraints: [colors, mood, no-go areas]. Output: [aspect ratio or channel notes].`

**Variant prompt**

`Using the same brief, generate 10 distinct directions. Each direction should change the visual concept, not just the color palette. Keep the product/category recognizable. Return a short label for each direction and make the differences obvious.`

**Editing prompt**

`Edit this image for [goal]. Preserve [parts that must stay the same]. Change only [specific areas]. Keep the lighting, perspective, and product shape consistent. Avoid adding extra text, logos, people, or objects unless requested.`

**Prompt debugging checklist**

- If the image is generic, add audience, channel, and a sharper subject.
- If composition is wrong, specify camera angle, crop, distance, and negative space.
- If the style is inconsistent, name the visual system: studio product photo, editorial magazine, SaaS landing page hero, documentary, minimal 3D render, or hand-drawn diagram.
- If the model ignores details, reduce the prompt to the most important constraints and test again.
- If the output cannot be used commercially, stop and check the provider and platform terms before publishing.

The most common mistake is mixing too many creative directions into one prompt. Pick one direction, generate variants, then deliberately test the next direction.

#### Workflows: ads, thumbnails, and product shots

The strongest AI image workflows are repeatable. Instead of starting from a blank prompt every time, build a small visual operating system: brief, generation, selection, editing, QA, and export. This keeps creative exploration fast without letting random outputs decide your brand.

### Ad creative workflow

For ads, start with the campaign promise, audience, offer, channel, and required format. Generate 10 concepts before choosing a style. Score each concept on attention, message clarity, brand fit, and whether copy can sit on top. Avoid unsupported product claims and do a final platform policy review before launch.

Ad prompt: `Generate 10 paid social ad image concepts for [product]. Audience: [audience]. Offer: [offer]. The image should communicate [benefit] without using text in the image. Leave clean negative space on the right for ad copy. Style: [style]. Avoid unrealistic product claims, medical/financial promises, fake UI, and confusing props.`

### Thumbnail workflow

For thumbnails, the question is whether the image communicates one idea at small size. Generate bold compositions with one focal point, high contrast, and low clutter. If the thumbnail needs a face, object, or screen, test whether it still reads at 25 percent size.

Thumbnail prompt: `Create 8 thumbnail concepts for a video/article about [topic]. The image must read clearly at small size. Use one focal subject, strong contrast, and room for a short title overlay. Do not include tiny text. Provide varied compositions: close-up, object-led, before/after, tension, and clean editorial.`

### Product shot workflow

Product images need more caution than concept art. If the image represents a real product, preserve shape, proportions, materials, colors, and packaging details. Use AI for backgrounds, mood boards, lifestyle concepts, and early mockups, then verify the final output against the real product.

Product prompt: `Create studio product image concepts for [product]. Preserve the exact product silhouette, color, material, and visible features from the reference. Change only the setting, lighting, and composition. Generate clean ecommerce, lifestyle desk, premium editorial, and social ad variants. Do not add claims, badges, ingredients, certifications, or text.`

Workflow checklist:

1. Write one creative brief before generating anything.
2. Generate 10 variants in the same model.
3. Pick the top 3 and run the same brief in another model.
4. Score outputs on usefulness, not novelty.
5. Edit the winner instead of endlessly regenerating.
6. Check aspect ratio, crop, text space, product accuracy, and policy risk.
7. Verify rights, license, and platform terms before using the asset in public campaigns.

#### Try in Whizi

Whizi is useful when you want to compare AI image generators without turning the process into a tab maze. Put one image brief in the workspace, generate 10 variants, then test the same brief across available models. The goal is to find the fastest path from brief to usable asset.

Use this side-by-side test. Choose one real asset you need this week: an ad image, blog header, thumbnail, product mockup, or concept board. Write the brief with the prompt structure above. Generate 10 variants in one model, then run the same brief in another model. Compare the top results using the scorecard and save the winning prompt as a reusable workflow.

This process is especially helpful for teams that are tempted to stack separate image subscriptions. You may discover that one model is best for first concepts, another for product-like realism, and another for editing or campaign iteration. That does not mean you need to buy every standalone plan. It means you need a workspace where model comparison is part of the creative process.

When you are ready to move from testing to production, compare your usage against [Whizi pricing](https://whizi.io/pricing), then create your account at [register](https://whizi.io/register). Start with one brief, generate 10 variants in Whizi, and let the outputs show which model belongs in your workflow.

Key points:

- A practical scorecard for choosing image models by style, control, speed, editing, and usage constraints
- A Flux vs Stable Diffusion vs GPT Image comparison table that focuses on workflow fit instead of brand hype
- Reusable prompts and checklists for ads, thumbnails, product shots, and visual concept exploration

Checklist:

- Score image models by style fit, prompt following, variation quality, editability, consistency, speed, and terms fit
- Use the same creative brief when comparing Flux, Stable Diffusion, GPT Image, or other image generators
- Generate 10 variants before deciding whether a model is good or bad for a workflow
- Preserve product shape, proportions, materials, and claims when creating product-adjacent visuals
- Check provider licenses, platform terms, and brand review requirements before publishing AI-generated images

FAQ:

Q: What is the best AI image generator in 2026?
A: The best AI image generator depends on the workflow. Test Flux, Stable Diffusion, GPT Image, and other tools against the same brief, then score outputs by style fit, control, editability, consistency, speed, and terms fit.

Q: Is Flux better than Stable Diffusion?
A: Flux is often worth testing for polished prompt-based generation, while Stable Diffusion remains important for flexible workflows and ecosystem depth. The better choice depends on your prompts, host, model version, controls, and usage needs.

Q: Can I use AI-generated images commercially?
A: Do not assume a universal answer. Commercial use depends on the provider, platform terms, model, input rights, output use, and your own legal requirements. Check the current terms before publishing or selling assets.

### The best AI to use in 2026, by what you are actually doing

URL: https://whizi.io/resources/best-ai-to-use/
Updated: August 2026

Quick answer: There is no single best AI: the best one depends on the task. As of August 2026, ChatGPT is the strongest all-round default, Claude leads on polished writing and long-document review, Gemini leads on very large documents and multimodal input, Perplexity leads on sourced research, and coding splits between Claude and ChatGPT.

A task-by-task verdict on the best AI to use in 2026: ChatGPT, Claude, Gemini, Perplexity, Grok, and DeepSeek compared for writing, coding, research, long documents, images, and price.

#### The short answer: there is no single best AI

The question "what is the best AI to use" has an honest answer and a marketable one. The marketable answer names a winner. The honest answer is that the leaders have separated by task rather than by overall rank, and they have stayed separated for about two years now, which the [AI model release timeline](https://whizi.io/resources/ai-model-timeline) makes easy to see at a glance. A model that writes a clean executive memo is not automatically the one that reads a 400-page filing without losing the middle of it, and neither of them is the one that hands you a research answer with links you can check.

So the useful version of the question is narrower: best for what, under which constraint. Below is where each major system currently earns its first attempt. Treat it as a starting point for a test on your own work, not a leaderboard. Prices are United States list prices as of August 2026 and differ by country.

| Use case | ChatGPT (OpenAI) | Claude (Anthropic) | Gemini (Google) | Other majors |
| --- | --- | --- | --- | --- |
| Everyday questions, drafts, planning | Strongest all-round default, broadest feature set | Very capable, slightly more formal by default | Very capable, best if you already live in Google apps | Grok for informal and current-events chat; DeepSeek if cost is the constraint |
| Polished writing and editing | Best for ideation, outlines, variants, many formats fast | Usual first pick for tone, restraint, and long-form structure | Good, strongest when the source material is already in Docs or Drive | Most others are wrappers on these three for writing |
| Coding | Fast debugging, scaffolds, tests, mature tool and function calling | Careful review, refactor reasoning, architecture tradeoffs, terminal agent | Solid, and strong when the repo or spec is very large | GitHub Copilot for in-editor completion; DeepSeek for cheap bulk work |
| Research with sources | Good synthesis, verify the citations yourself | Careful with uncertainty, conservative about claims | Good, and ties into Google Search results | Perplexity is the category specialist: answers arrive with links attached |
| Long documents and PDFs | Handles most business documents comfortably | Strong on dense text and careful extraction | Usual first pick at the extreme end, million-token class context | Most smaller tools truncate silently, so test with your actual file |
| Images | GPT Image generation and editing inside chat | No image generation, reads images only | Native image generation and editing in the Gemini apps | Flux and Stable Diffusion lead for production image work |
| Price (standard paid tier) | Plus $20/mo, Go $8/mo | Pro $20/mo, or $17/mo billed annually | AI Pro $19.99/mo, AI Plus $4.99/mo | Perplexity Pro $20/mo, SuperGrok $30/mo, DeepSeek near free |

One row of that table matters more than the rest for most readers: the price row. The paid tiers have converged on roughly $20 a month each, which means the decision is rarely "which single model is best" and usually "how many $20 charges am I prepared to carry." Three of them is close to $60 a month. That arithmetic, rather than any benchmark, is what actually decides most stacks.

If you have never used any of these, start at [the best AI apps for beginners](https://whizi.io/resources/best-ai-apps-for-beginners), which covers signup, free tiers, and what to install first. If you want the decision framework rather than the verdicts, [how to choose an AI model](https://whizi.io/resources/how-to-choose-ai-model) has a ten-minute test protocol and a scorecard.

#### Best AI for writing and editing

Verdict: Claude for the draft that has to sound like a person wrote it, ChatGPT for volume and range. This is the most consistent split in the field and the easiest one to verify yourself in an afternoon.

Claude tends to win on editorial restraint. Give it a rough memo and it usually returns something with a defensible structure, fewer filler transitions, and a voice closer to the one you started with. It is the common first choice for executive summaries, sensitive customer email, fundraising updates, and long-form pieces where flattening the author's voice would be the failure mode.

ChatGPT wins when you need range rather than polish: twenty headline directions, a landing page structure, three email versions, a comparison matrix, and an outline, all in one sitting. It is also the better partner for pressure-testing an argument, because asking it for counterarguments reliably produces specific ones rather than a hedge.

Gemini is a real option for writing and is underrated when the source material already lives in Google Docs or Drive, because the integration removes the copy-paste step that quietly eats the time savings. If your writing starts from documents you did not write, that matters more than a small quality difference.

The test that settles it takes ten minutes. Take one real piece of writing from last week, paste the identical prompt and identical source material into two of them, and score the outputs on one axis: how many minutes of editing stand between this and something you would send. That number is the whole comparison. [Claude vs ChatGPT](https://whizi.io/resources/claude-vs-chatgpt) walks through the same test with prompts you can copy, and [ChatGPT alternatives for writing](https://whizi.io/resources/chatgpt-alternatives-for-writing) scores the wider field on the same axis.

#### Best AI for coding

Verdict: it depends on where the code is. In an editor, GitHub Copilot still owns inline completion because it sees the file you are in, and the Microsoft assistant sold inside Word and Outlook is a separate product with a separate argument, weighed in [Copilot vs ChatGPT](https://whizi.io/resources/copilot-vs-chatgpt). In a chat window, Claude and ChatGPT split the work along a fairly clean line.

ChatGPT is the faster first stop for a concrete, bounded problem: a failing test, an unfamiliar API, a small implementation, a set of unit tests, a regex you do not want to write. Its function calling and tool patterns are the most mature, which matters if the model is going to sit inside an application rather than a chat box.

Claude is the better second opinion and the better reviewer. Ask it what could break, where the hidden coupling is, which tests would actually prove the fix works, and whether an abstraction earns its keep, and the answers tend to be specific rather than generic. It is at its strongest when the code sits inside a longer architecture discussion instead of a snippet.

Gemini earns its place when the input is genuinely enormous: a whole repository, a long specification plus the code that implements it, or a migration where the relevant context is spread across dozens of files. Cost-sensitive teams should also price DeepSeek, which is dramatically cheaper per token and good enough for bulk mechanical work like test generation or docstring passes, though most people should verify its output more carefully.

The pattern that outperforms picking one: draft with one model, review with another. Ask the first for the smallest safe change and the tests, then paste the proposed patch into the second and ask for regression risk. [AI for coding](https://whizi.io/resources/ai-for-coding) covers that loop in detail, and it is the single highest-return habit in this whole article.

#### Best AI for research and checking facts

Verdict: Perplexity if you need the sources attached, one of the big three if you need the synthesis. These are different jobs and conflating them is how people end up citing a hallucinated study.

Perplexity is built around search-then-answer, so a response arrives with links you can open. That does not make it more accurate in some abstract sense, and it can still summarize a source badly, but it makes the answer checkable in seconds instead of minutes. For anything you intend to publish, quote, or act on, checkability is worth more than eloquence.

ChatGPT, Claude, and Gemini all have web search built in now and all of them synthesize better than a search-first tool does, which is the real reason to use them for research: they are better at holding six sources in mind and telling you where the sources disagree. The trade is that a citation from a general chat model is a claim about a source, not the source, and it needs opening.

Claude is noticeably more willing to say it does not know, which is an asset in research and an irritation elsewhere. Gemini has the advantage of sitting next to Google Search results. Grok pulls current posts from X, which is genuinely useful for breaking events and genuinely unreliable for anything requiring an authoritative source.

Whatever you use, the discipline is the same and it is not optional: ask for claims and sources separately, ask what would falsify the conclusion, and open at least the citations that carry weight. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) explains why confident wrong answers are a structural feature rather than a bug that will be patched out.

#### Best AI for long documents, PDFs, and files

Verdict: Gemini at the extreme end, Claude for dense text, and test with your real file rather than trusting a stated limit.

Google documents context windows of one million tokens and above on several Gemini models, and the long-context guidance is aimed squarely at the case where you hand over a large body of material up front: many documents at once, long transcripts, extended reference sets, whole codebases. When the core problem is keeping a lot of source available simultaneously, that is the first thing to try.

Claude is the stronger pick for dense text where careful extraction matters more than raw volume: contracts, filings, research packets, interview transcripts you need quoted accurately. ChatGPT handles the ordinary business document range comfortably and is usually the fastest route from a document to a written recommendation, which is often the actual deliverable.

The trap in this category is that a large stated context window is not a promise of attention across all of it. Models can and do lose detail in the middle of very long inputs. The defense is a workflow rather than a model choice: ask for a document map first, then ask for extraction with section references, then ask a second model to check the extraction for anything missing or overstated, and verify every number and quote by hand before it leaves your desk.

[What is a context window](https://whizi.io/resources/what-is-a-context-window) explains the underlying limit in plain terms, and [summarize a PDF with AI](https://whizi.io/resources/summarize-pdf-with-ai) has the step-by-step version of the workflow above.

#### Best AI for images, audio, and video

Verdict: the chat apps are fine for a quick image and wrong for production work. This is the category where "best AI" genuinely means a different set of products.

For a quick illustration inside a conversation, the image tools built into ChatGPT and Gemini are convenient and good enough. Claude does not generate images at all, though it reads them well. Convenience is the entire argument here: the image is already in the thread where you were working.

For work that ships, the specialist models lead. Flux from Black Forest Labs and Stable Diffusion from Stability AI are where control lives: prompt adherence, editing that preserves the parts you liked, consistency across a set, and the ability to iterate without rebuilding an image from scratch. The things that break in production, hands, typography, exact aspect ratios, keeping a product shape stable across ten variants, are exactly the things a chat-window image tool handles least well.

The full comparison, including a scorecard and the licensing question, is in [best AI image generators](https://whizi.io/resources/best-ai-image-generators), and [how to create AI images](https://whizi.io/resources/image-generation) covers the six things a description has to name before any model can give you what you pictured. One point worth repeating here: check the provider terms for your intended use before publishing anything, because the rules differ by tool and there is no universal commercial-use rule to fall back on.

#### What the best AI costs, and how to pick yours this week

The prices below are United States list prices as of August 2026 and change often, so treat them as the shape of the market rather than a quote. Several providers charge different amounts in other countries.

| Plan | Provider | Price per month | What it is for |
| --- | --- | --- | --- |
| ChatGPT Free | OpenAI | $0 | Generous enough that many people never need to pay |
| ChatGPT Go | OpenAI | $8 | Higher limits without the full Plus feature set |
| ChatGPT Plus | OpenAI | $20 | The default paid AI subscription for most people |
| Claude Pro | Anthropic | $20, or $17 billed annually | Writing, long documents, careful review |
| Google AI Plus | Google | $4.99 | The cheapest paid tier from a major provider |
| Google AI Pro | Google | $19.99 | Long context, Google Workspace integration |
| Perplexity Pro | Perplexity | $20 | Research with sources attached |
| SuperGrok | xAI | $30 | Grok with higher limits |
| DeepSeek | DeepSeek | Near free | Cost-constrained bulk work |

Three observations that matter more than any single price. First, the free tiers are better than most people assume, and a large share of readers genuinely do not need to pay at all. Second, the paid tiers have converged, so switching between them saves you nothing and the only real saving is buying fewer of them. Third, the stack is what gets expensive: ChatGPT Plus for general work, Claude Pro for writing, Google AI Pro for documents and Perplexity Pro for research is roughly $80 a month for four tools that overlap heavily.

Here is the decision, compressed. If you will only ever open one AI, buy ChatGPT Plus and stop reading comparison articles. If your work is mostly writing, buy Claude Pro instead. If you live inside Google Workspace, buy Google AI Pro. If your work genuinely spans several of these, the choice is between paying each provider separately and using a workspace that puts several models behind one bill, which is a real category with real tradeoffs: you get model switching and one history, you give up each provider's deepest integrations and first access to new features. [One subscription for all AI models](https://whizi.io/resources/one-subscription-all-ai-models) covers that trade honestly, including who should skip it. If the two you actually want are GPT and Claude, [the cheapest way to use ChatGPT and Claude together](https://whizi.io/resources/cheapest-way-to-use-claude-and-gpt) prices four routes for a light, a medium, and a heavy user.

Then run the test, because a verdict from an article is worth less than fifteen minutes with your own work. Take one real task from last week, send the identical prompt to two or three models, and score the outputs on accuracy, format compliance, and minutes of cleanup. Write down which one won and for what kind of task. That note, repeated three or four times, is a routing rule, and a routing rule beats a favorite model every time.

Key points:

- Every serious ranking collapses once you name the task: the model that writes the best memo is not the one that reads the best 400-page PDF
- If you only ever open one AI, ChatGPT is the safest single pick in 2026, and that is a statement about breadth rather than about winning any one category
- The paid tiers cluster at roughly $20 a month each, so the real question for most people is how many of them you are willing to buy

Checklist:

- Name the task before naming a model: writing, coding, research, documents, or images
- Start with ChatGPT if you are only going to use one AI, and Claude if the work is mostly writing
- Use Perplexity or a search-enabled model when the answer needs sources you can open
- Test long-document work with your actual file, not with a stated context limit
- Use a specialist image model for anything that ships, and check the terms for your intended use
- Count the paid subscriptions you are carrying before adding another at $20 a month
- Run the same prompt through two models on a task that matters, and save the winner as a routing rule

FAQ:

Q: What is the best AI to use in 2026?
A: There is no single best AI, and the leaders have separated by task rather than by overall rank. ChatGPT is the strongest all-round default and the safest pick if you will only use one. Claude leads on polished writing and careful review, Gemini on very large documents and Google integration, and Perplexity on research with sources attached.

Q: Which AI is best for writing?
A: Claude is the usual first pick for polished drafts, editing, and tone-sensitive work, because it tends to preserve the author's voice and needs less cleanup. ChatGPT is better when you need range rather than polish: outlines, many variants, and turning one idea into several formats quickly. Judge them by minutes of editing left, not by which reads better at first glance.

Q: Which AI is best for coding?
A: GitHub Copilot leads for inline completion inside an editor. In a chat window, ChatGPT is faster for bounded problems like a failing test or an unfamiliar API, and Claude is stronger as a reviewer for architecture tradeoffs, refactor reasoning, and regression risk. The pattern that beats picking one is drafting with one model and reviewing the patch with the other.

Q: Is the best AI free, or do I have to pay?
A: The free tiers from OpenAI, Anthropic, and Google are good enough that many people never need a paid plan. Paying buys higher limits, longer documents, faster models, and access to the newest releases. If you hit limits mid-task more than once a week, that is the signal to upgrade, and the standard paid tiers cluster around $20 a month.

Q: Should I pay for more than one AI subscription?
A: Only if you have evidence that you use more than one regularly. Four separate plans at roughly $20 each is about $80 a month for tools that overlap on most tasks. Spend a normal week noting which model you actually opened for which job, then either cancel what you do not use or look at a single multi-model plan that covers the ones you do.

Q: Which AI is best for images?
A: For a quick image inside a conversation, the tools built into ChatGPT and Gemini are convenient, and Claude does not generate images at all. For work that ships, specialist models like Flux and Stable Diffusion lead on prompt adherence, editing, and consistency across a set. Check the provider terms for your intended use before publishing.

### The best all-in-one AI platforms in 2026

URL: https://whizi.io/resources/best-all-in-one-ai-platforms/
Updated: August 2026

Quick answer: The best all-in-one AI platform depends on whether you want to manage API keys. Hosted subscriptions like Whizi and Poe charge one fee and need no setup. Bring-your-own-key clients such as TypingMind, ChatHub, and big-AGI are cheaper at spiky volume. Self-hosted LibreChat or Open WebUI is the only option that keeps data on your own hardware.

Every serious way to get GPT, Claude and Gemini in one place, grouped by the three architectures that actually exist, with an honest account of who each one suits.

#### The short answer

Yes, you can get GPT, Claude, Gemini and a few hundred open models through a single interface, and there are three fundamentally different ways to do it. Which one is right depends almost entirely on whether you want to manage API keys.

| Architecture | Examples | What you pay | Who it suits |
| --- | --- | --- | --- |
| Hosted subscription | Whizi, Poe | One monthly fee, metered in credits or points | People who want it to work with nothing to configure |
| Bring your own key | TypingMind, ChatHub, big-AGI | The interface, then each provider directly at cost | Technical users with moderate, spiky usage |
| Self-hosted open source | LibreChat, Open WebUI | Nothing for the software, plus running the service yourself | Anyone whose data must not leave their hardware |
| Not really in the category | OpenRouter, Perplexity | Per token, or a separate subscription | Developers, and people who want sourced answers |

**Hosted subscriptions** charge one monthly fee and handle everything: Whizi, Poe. Best for people who want it to work and do not want a second bill to reason about.

**Bring-your-own-key clients** give you an interface and you supply provider API keys: TypingMind, ChatHub, big-AGI. You pay providers directly at cost, which is cheaper at low volume and requires you to hold accounts with each provider.

**Self-hosted open source** runs on your own machine or server: LibreChat, Open WebUI. Free software, real setup, and the only option that can also run local models with no data leaving your hardware.

A fourth thing often listed alongside these is not really in the category: OpenRouter is a developer API gateway rather than a consumer workspace, and answer engines like Perplexity are a different product shape entirely.

#### Hosted subscriptions

One price, no keys, no setup. The provider absorbs the complexity and meters your usage in some internal unit.

**Whizi.** Over 280 models across three plans. Starter is $15.99 per month or $10.99 billed yearly and includes three fast models plus file upload. Pro is $29.99 or $19.99 yearly and adds Claude Sonnet 5, GPT-5.6 Terra, Grok 4.6, Kimi K3 and roughly 37 more, with 2,000 credits per month. Powerhouse is $49.99 or $34.99 yearly, opens the full catalogue including Claude Opus 5 and GPT-5.6 Sol, and adds side-by-side comparison, image, video, music and 500 minutes of voice. Web, iOS and Android on one login. Usage is metered in credits, from 1 per message on roughly half the catalogue up to 20 for the frontier tier.

**Poe.** Quora product, large model catalogue, a points system rather than credits, and a strong bot and persona ecosystem that no one else in this list matches. It is the most established option in the category and the one most people have heard of.

The honest trade for both: you are buying a metered allowance across many models rather than unlimited use of one. If your actual pattern is thirty long conversations a day with a single frontier model and nothing else, a direct subscription to that provider is a better deal and you should take it.

#### Bring-your-own-key clients

You pay for the interface once or on a small subscription, then supply API keys and pay each provider directly for tokens.

**TypingMind.** The most polished of these. One-time or subscription licence, your keys, your data stored locally or in your own cloud. Strong plugin and prompt library support.

**ChatHub.** Browser extension shape rather than a web app, with side-by-side chat across several models as the headline feature. Convenient if you live in the browser and want models next to whatever you are reading.

**big-AGI.** Open source, feature-dense, and unusually fast to adopt new model capabilities. More of an enthusiast tool than a product for a team.

Who this suits: people with moderate, spiky usage who are comfortable holding accounts with OpenAI, Anthropic and Google, and who would rather pay $6 in tokens one month and $60 the next than a flat fee. Who it does not suit: anyone who does not want to think about API keys, rate limits, per-provider billing, or what happens when a key leaks.

The real cost here is not money, it is administration. Three provider accounts is three sets of credentials, three billing relationships and three places to check when something stops working.

#### Self-hosted open source

**LibreChat** and **Open WebUI** are the two serious options. Both are free, actively developed, and run on your own infrastructure. Both connect to hosted provider APIs with your keys, and both can also run local open-weight models with nothing leaving your machine.

That last capability is the reason to choose this category, and it is a genuinely important one. If your constraint is that certain material cannot go to a third party at all, no hosted product solves that and a local model on your own hardware does.

The cost is real and recurring: you are running a service. Updates, storage, authentication, backups and the occasional broken deployment are yours. Teams that already run infrastructure absorb this easily. Individuals usually underestimate it, enjoy the first month, and quietly stop updating by the third.

#### What is not actually in this category

**OpenRouter** appears on most lists like this one and does not belong. It is an API gateway for developers: one endpoint and one bill in front of hundreds of models, which is genuinely excellent, but it is infrastructure rather than a place to work. If you are building software, use it. If you want somewhere to chat with Claude about a PDF, it is the wrong shape.

**Perplexity** is an answer engine. It is very good at sourced answers to questions about the world and it is not trying to be a multi-model workspace, so comparing them on model count misses what either is for.

**Microsoft Copilot and Google Gemini** are single-vendor products with deep integration into their own ecosystems. If you live in Microsoft 365 or Google Workspace, that integration may be worth more than model choice. That is a real argument and it is the opposite of the one this category makes.

#### The question that settles it

Ask yourself: do I want to manage API keys?

If no, you want a hosted subscription. Compare Whizi and Poe on which models are on which tier and how the metering works, and take the trial.

If yes and you are technical and your usage is spiky, a bring-your-own-key client is cheaper and gives you more control. TypingMind if you want polish, ChatHub if you want a browser extension, big-AGI if you want everything immediately.

If yes and your real constraint is that data must not leave your hardware, self-host Open WebUI or LibreChat with local models. Nothing else solves that problem.

The second question, only if the first left you undecided: how many models do you actually use in a week? Most people who count honestly find the answer is two or three. If that is you, the size of a catalogue is close to irrelevant and what matters is whether the two you use are on the tier you can afford.

#### How to evaluate any of them in a week

1. **Run it in parallel with what you have.** Do not cancel anything yet. Send the same real tasks to both for a week.
2. **Use real work, not test questions.** Testing an AI on questions you can already answer measures the wrong thing and is why so many trials end in a shrug.
3. **Count your actual model usage.** Note which model answered each task that mattered. Most people discover they use two.
4. **Check the metering against that.** Whatever the unit is, work out what your real week costs in it. This is where a plan either fits or does not, and it is invisible from the pricing page.
5. **Cancel the loser.** One of them is now clearly better for your work. This is the step people skip, and it is why so many end up paying for three things.

Key points:

- The three architectures in this category, and why the choice between them matters more than the choice of product
- Where each option genuinely wins, including where Whizi loses
- The one question that settles it for most people in about a minute

Checklist:

- Decide first whether you want to manage API keys, which settles the architecture
- Hosted subscriptions suit people who want one bill and no setup
- Bring-your-own-key is cheaper at spiky volume and costs administration
- Self-hosting is the only option where data need not leave your hardware
- OpenRouter is developer infrastructure, not a workspace
- Count how many models you actually use in a week before comparing catalogues
- Run a new tool in parallel for a week before cancelling anything

FAQ:

Q: What is the best all-in-one AI platform?
A: There is no single answer, because the category contains three different architectures. For most people who want GPT, Claude and Gemini in one place without managing API keys, a hosted subscription like Whizi or Poe is the right shape. For technical users with spiky usage, a bring-your-own-key client like TypingMind or ChatHub is cheaper. For anyone whose data cannot leave their own hardware, self-hosted Open WebUI or LibreChat is the only option that solves the problem.

Q: Is there one app that gives me access to all the major AI models?
A: Yes, several. Whizi includes over 280 models across OpenAI, Anthropic, Google, xAI, DeepSeek, Meta, Mistral, Moonshot and Qwen on one subscription starting at $10.99 per month billed yearly, with Claude and GPT both available from the Pro tier up. Poe offers a comparable catalogue on a points system. Both work on web and mobile with no API keys required.

Q: Is an all-in-one AI subscription cheaper than paying for ChatGPT Plus and Claude Pro?
A: Usually yes on price and sometimes no on fit. ChatGPT Plus and Claude Pro are $20 each, so $40 per month for two model families, while Whizi Pro at $19.99 billed yearly covers both plus a few dozen more. The trade is metering: you get a monthly allowance across many models rather than unlimited use of one, so a very heavy single-model user can still be better off paying that provider directly.

Q: What is the difference between an all-in-one platform and OpenRouter?
A: OpenRouter is an API gateway aimed at developers: one endpoint, one bill, hundreds of models, and no consumer interface to speak of. All-in-one platforms are workspaces aimed at people doing work, with file upload, conversation history, sharing and mobile apps. They are complementary rather than competing, and comparing them on model count obscures that they solve different problems.

Q: Can I run all these models locally instead?
A: You can run open-weight models locally through Open WebUI or LibreChat with nothing leaving your hardware, which is the strongest privacy position available. You cannot run GPT, Claude or Gemini locally, because those weights are not published, so a local setup means accepting open models for the private work and using an API for the rest.

### What is ChatGPT and How to Use It (For Beginners)

URL: https://whizi.io/resources/chatgpt/
Updated: August 2026

Quick answer: ChatGPT is a program you type to in ordinary English, and it types back. It is not looking anything up. It learned patterns from an enormous amount of text and predicts what should come next, which is why it can write a good cover letter and state a wrong date in the same confident tone.

Learn what ChatGPT is, how it works, and practical ways beginners can use it every day to write emails, solve problems, and save time.

#### What ChatGPT actually is

ChatGPT is a program you type to, in ordinary English, and it types back. No commands to learn, no setup. If you can send a text message you can use it.

The one thing worth understanding, because nearly every beginner mistake comes from not knowing it: **ChatGPT is not looking things up.** It is not searching the internet unless you specifically use a version that does. It learned patterns from an enormous amount of text and predicts what should come next.

That explains both the magic and the problem. It can write a decent cover letter for a job it has never seen, because it has seen thousands of cover letters. It can also state a wrong date, a made-up statistic, or a legal detail that does not apply where you live, in exactly the same confident tone as everything else.

It is not a search engine and it is not a database. It is closer to a very well read assistant who never says "I am not sure". [AI vs Google search](https://whizi.io/resources/ai-vs-google-search) turns that difference into a rule for which box to type a given question into.

#### The formula for getting good answers

Most disappointing results are prompt problems. Vague in, generic out. Four parts fix nearly all of it.

| Part | What to add | Why |
| --- | --- | --- |
| The task | What you want, specifically | "Help with my email" is not a task |
| Who it is for | The reader and what they care about | Changes the output more than any style instruction |
| The facts | Your actual details, pasted in | Otherwise it invents plausible ones |
| The limits | Length, format, what to avoid | Stops it padding and rambling |

Weak: `write an email asking for time off`.

Strong: `Write a short email to my manager asking for Friday 14th off. I have no deadlines that week and my colleague has agreed to cover the client call. Tone: polite, not apologetic. Under 80 words. No subject line needed.`

The second one takes twenty seconds longer to type and saves you three rounds of "make it shorter" and "less formal".

#### Ten prompts to copy

- **The awkward email.** `Write a polite message to [person] about [situation]. Facts: [paste]. Tone: friendly but clear. Under 100 words. End with one specific next step.`
- **Fix my draft.** `Rewrite this so it is clearer and shorter. Keep every fact and name exactly as written. Do not make it more formal. Draft: [paste].`
- **Explain it to me.** `Explain [topic] to someone who knows nothing about it. Use one everyday example. Then ask me one question to check whether I understood.`
- **Plan something.** `Make a [duration] plan for [goal]. I have [time available] and [constraints]. Give me a table with the day, the task, and how long it takes. No motivational filler.`
- **Use what I have.** `Here is what is in my fridge: [list]. Give me three dinners using mostly these, with anything extra I would need to buy listed separately.`
- **Prepare me.** `I have [meeting or interview] about [subject]. Give me the eight questions I am most likely to be asked, hardest first, and no answers.`
- **Compare options.** `I am choosing between [A] and [B] for [purpose]. Give me a table of the real trade-offs, then say which you would pick and what would change your mind.`
- **Sort out my notes.** `Here are messy notes. Turn them into a clear summary with headings and a list of anything that needs a decision. Do not add anything that is not in the notes. Notes: [paste].`
- **Argue against me.** `I am planning to [decision]. Make the strongest case that this is a mistake. Do not conclude that both sides have merit.`
- **Second opinion.** `Here is an answer I got. What is wrong with it, or what did it miss? Quote the specific parts.`

That last one is worth trying in a different AI than the one that produced the answer. Any model reviewing its own work mostly tells you it is fine.

#### When to believe it

A simple rule: the more specific and checkable a claim is, the more you should verify it.

**Usually fine.** Explaining a concept, rephrasing your own text, structuring your ideas, brainstorming, general how-things-work knowledge.

**Check it.** Any number, date, name, price, quote, statistic, or citation. Anything about a specific person or company. Anything current.

**Do not rely on it alone.** Medical, legal, tax, and financial specifics. It does not know your jurisdiction, your situation, or what changed last year. It is fine for understanding the vocabulary before you talk to someone qualified.

A useful test: `How confident are you about this, and what would you want me to check?` It is imperfect at self-assessment, but the things it flags are usually worth flagging.

#### Five habits that separate good users from frustrated ones

**Give it your material.** It edits far better than it invents. Rough notes plus "make this good" beats "write me something" every time.

**Keep going in the same conversation.** It remembers the current chat, so refining beats restarting. "Shorter, and lose the last paragraph" works.

**Say what you do not want.** "No bullet points", "do not start with In today's fast-paced world", "no exclamation marks".

**Ask what is weak before asking for better.** Otherwise you get a smoother version of the same thing rather than a different one.

**Do not trust it just because it pushed back confidently, or fold just because it agreed.** Both happen. If you disagree, ask it to make the case rather than restating.

#### What it costs, and what else is out there

There is a free version and it is genuinely useful. The paid plan is around $20 a month, mainly for higher limits and access to the more capable models.

Worth knowing before you subscribe: ChatGPT is not the only one, and it is not the best at everything. Claude writes noticeably better, which matters if your main use is emails and documents. Gemini is better for anything recent and for very long documents. Those are not marginal differences and you will notice them within a week.

That is the reason a combined workspace exists. Whizi puts ChatGPT, Claude, and Gemini in one window for less than any one of them costs alone, so you can send each task to whichever is better and get a second opinion when something matters. See [ChatGPT vs Claude](https://whizi.io/compare/chatgpt-vs-claude) and [ChatGPT vs Gemini](https://whizi.io/compare/chatgpt-vs-gemini) if you want the detail first.

Key points:

- What it actually is, and the one misunderstanding that causes most beginner mistakes
- Ten prompts you can copy today, with the four-part formula behind them
- When to believe it and when not to, explained plainly

Checklist:

- Include the task, the reader, your real facts, and the limits in every prompt
- Give it your rough draft instead of asking it to start from nothing
- Keep refining in the same conversation rather than starting over
- Say what you do not want, not just what you do
- Verify every number, date, name, and citation
- Ask what is weak before asking for a rewrite
- Try the same prompt in a different AI once, to see the difference

FAQ:

Q: Do I have to pay to use ChatGPT?
A: No. The free version is genuinely useful and covers a lot of everyday work, with limits on the more capable models that tend to appear when you are busiest. Paying makes sense at the point where being interrupted mid-task costs more than the subscription. If you find yourself considering a second AI subscription as well, a combined workspace usually costs less than either plan alone.

Q: Can ChatGPT write my resume?
A: It is very good at structuring and polishing one, and poor at inventing one. Give it your actual job history and the job description, and ask it to match your experience to what the posting asks for. Then check every line, because it will occasionally smooth a detail into something slightly untrue, and that is your name on the document.

Q: Does ChatGPT know about recent events?
A: Only if you are using a version connected to the web, and even then you should open the sources it cites. Its underlying knowledge stops at a training date, and it will answer a question about last month from general knowledge in exactly the same confident tone as everything else. This is the single most common cause of a beginner being confidently misled.

Q: Is it safe to paste personal or work information?
A: Use judgement. Never paste passwords, card numbers, or government identifiers into any AI tool. Be careful with other people’s personal information and with anything confidential at work, where your employer may have a policy. For most tasks you can remove names and identifying details without affecting the answer at all, which is the simplest safe habit.

Q: Is ChatGPT the best AI?
A: It is the best known and a very good all-rounder, particularly for structure, plans, and formatted output. It is not the best at everything: Claude writes better for anything a person reads end to end, and Gemini is stronger on recent information and very long documents. Trying the same prompt in two of them once is the fastest way to see this for yourself.

### How to choose a ChatGPT alternative: an evaluation guide

URL: https://whizi.io/resources/chatgpt-alternatives/
Updated: August 2026

Quick answer: Choose a ChatGPT alternative by running one real task from your week through each contender and scoring the results, rather than by reputation. Judge task fit, output quality, reasoning visibility, context handling, tool support, verification path, and cost. Claude suits writing, Gemini suits multimodal and document work, and coding assistants suit development.

A practical rubric for evaluating ChatGPT alternatives for writing, coding, research, and documents, so you choose the right AI setup without stacking subscriptions.

#### What people mean by "ChatGPT alternative"

When people search for ChatGPT alternatives, they are usually not asking for a clone. They are asking a more practical question: "Which AI assistant should I use for the work I actually do?" That might mean a different writing style, stronger long-document handling, better coding help, better image understanding, clearer citations, or simply a lower monthly cost.

The best answer is rarely "replace ChatGPT everywhere." ChatGPT remains one of the top AI chatbots, and OpenAI documents a broad model lineup for text generation, reasoning, images, vision, structured outputs, tool use, and more. But Claude, Gemini, research tools, [open-source models such as Llama](https://whizi.io/resources/llama-3), coding assistants, and unified workspaces can all be better fits depending on the task.

A useful ChatGPT competitor should be judged by workflow fit, not brand loyalty. If you write long client deliverables, you may care about tone and context. If you code, you may care about debugging and tests. If you research, you need traceability. If you use AI every day, you may care most about comparing outputs without paying for three or four separate subscriptions. If you just want a ranked list of tools with pricing, start with our [ChatGPT alternatives](https://whizi.io/alternatives/chatgpt) page. This guide is the companion piece: it shows you how to run that evaluation against your own work.

#### How we evaluate tools

The fastest way to choose an AI chatbot alternative for work is to run the same task through each contender and score the result. Do not ask vague demo prompts like "write a blog post about productivity." Use a real task from your week, with real constraints, and compare the outputs side by side.

Use this workflow-based rubric:

1. **Task fit:** Does the model understand the job: writing, coding, research, data extraction, PDF analysis, brainstorming, or image review?
2. **Output quality:** Is the answer specific, structured, and usable without heavy rewriting?
3. **Reasoning visibility:** Does the response explain assumptions, tradeoffs, and uncertainty clearly enough for you to review it?
4. **Context handling:** Can it follow long briefs, documents, code snippets, or multi-step instructions without drifting?
5. **Tool support:** Does it work with files, images, web research, structured outputs, or automations when your workflow requires them?
6. **Verification path:** Can you check sources, reproduce code suggestions, or inspect how conclusions were reached?
7. **Cost and access:** Are you paying for one assistant, several subscriptions, or a consolidated workspace that gives you multiple models in one place?

The key is to score the workflow, not the first impression. A model that sounds polished may still invent details. A model that is excellent at code may be less pleasant for marketing copy. The best ChatGPT alternative is the one that repeatedly produces the cleanest result for your actual work.

#### Criteria

Here is the checklist we use when comparing the best ChatGPT alternatives in 2026:

- **Writing quality:** voice control, structure, clarity, editing ability, and ability to avoid generic phrasing.
- **Coding usefulness:** debugging, code explanation, refactoring, test writing, and awareness of edge cases.
- **Research quality:** source handling, citation discipline, synthesis, and ability to separate evidence from opinion.
- **Document and context handling:** PDFs, long briefs, meeting notes, product specs, contracts, and research packets.
- **Multimodal work:** ability to understand or generate across text, images, files, and structured data.
- **Workflow speed:** how quickly you can move from prompt to usable output.
- **Subscription efficiency:** whether the tool reduces or increases the number of paid AI accounts you manage.

Also run a "failure test." Ask each tool to identify missing context, list assumptions, and name what it cannot verify. The best assistants are not just confident; they are reviewable.

#### Best alternatives at a glance

The table below is a practical starting point. Specific model names, limits, and pricing can change, so treat this as a use-case map rather than a permanent ranking.

| Tool or model family | Best fit | Strengths | Watchouts | Best next step |
| --- | --- | --- | --- | --- |
| **Claude** | Writing, editing, long-form work, careful synthesis | Strong prose, thoughtful restructuring, useful for long briefs and nuanced documents | May not be the default choice for every tool-heavy workflow | Test it against ChatGPT on one real writing brief |
| **Gemini** | Multimodal work, Google-style productivity, long-context document tasks | Strong fit for text + image + document workflows; official Gemini docs emphasize multiple model options for different tasks | Output quality still depends heavily on prompt detail and verification | Try it on a PDF, screenshot, or research packet |
| **Perplexity-style answer engines** | Fast research discovery and web-backed answers | Useful for source discovery and quick landscape scans | Summaries still need source checking and synthesis | Use it to collect sources, then synthesize elsewhere |
| **Open-source models** | Privacy-sensitive experiments, local workflows, customization | More control, flexible deployment, active ecosystem | Setup, hardware, and quality vary widely | Use for specialized or private workflows, not as a universal answer |
| **Coding assistants** | IDE help, code review, refactors, tests | Integrated with developer workflows and repositories | May be less useful for non-code work | Pair with a general model for planning and explanation |
| **Whizi** | Comparing multiple AI chatbots in one workspace | Lets you test prompts across models and reduce subscription sprawl | You still need to choose the right model for each task | Start with a side-by-side prompt test in Whizi |

The move is not memorizing a winner. It is building a repeatable comparison loop: same prompt, same inputs, same scoring criteria, then choose the output you would actually ship.

#### Best for writing/coding/research

Different tools feel "best" depending on the task. The sections below give a more useful decision path than a single universal winner.

### Best for writing

For writing, the best ChatGPT alternative is the one that can preserve your intent while improving structure, tone, and specificity. Claude is often a strong contender for long-form editing, voice-sensitive rewrites, emails, reports, and strategy documents. ChatGPT can still be excellent for ideation, outlining, and turning rough notes into useful drafts. Gemini can be useful when writing depends on documents, images, or other context-heavy inputs.

Use this test prompt:

> Rewrite this draft for a busy executive audience. Keep the meaning, cut repetition, preserve any concrete numbers, and make the recommendation impossible to miss. After the rewrite, list the three biggest edits you made and why.

Score the result on clarity, voice, structure, and how much editing remains.

### Best for coding

For coding, do not judge only by whether the first answer looks plausible. Judge by whether the assistant helps you reduce risk. A good AI coding workflow includes a reproduction, likely root causes, the smallest safe fix, tests, and a review of side effects.

Use this test prompt:

> Here is the bug, the expected behavior, the current behavior, and the relevant code. First identify the most likely root causes. Then propose the smallest fix. Then write tests that would fail before the fix and pass after it. Do not rewrite unrelated code.

The best AI for coding is the one that gives you a path you can verify, not the one that writes the most code.

### Best for research

For research, prioritize traceability. The best ChatGPT alternative for research is not just the assistant that writes a smooth summary; it is the one that helps you collect sources, extract claims, compare evidence, and mark uncertainty.

Use this test prompt:

> Build a research brief from these sources. Separate direct evidence, interpretation, and open questions. Flag any claim that is not supported by the provided material. End with a decision-ready summary and a list of facts I should verify manually.

That last line matters. AI research should make verification easier, not optional.

#### Free vs paid vs consolidated subscription

Free AI chatbots are useful for experimentation, but free usually comes with tradeoffs: lower limits, slower access, fewer advanced models, weaker file handling, or less predictable availability. Paid plans can unlock better models and workflows, but the cost problem starts when you subscribe to every promising tool separately.

A common AI stack can quietly become expensive: one subscription for ChatGPT, one for Claude, one for Gemini or another model, and another for research or image work. Even when each tool is valuable, the combined spend can be harder to justify.

That is where consolidation becomes a real decision factor. Instead of asking "Which single AI chatbot should I marry forever?", ask "Where can I compare the best model for each task without turning my budget into a pile of recurring charges?"

If you already pay for multiple tools, run your current stack through the [AI subscription savings calculator](https://whizi.io/tools/calculator). Then compare that number with the cost of a unified workspace on [Whizi pricing](https://whizi.io/pricing). The goal is not to use fewer models; the goal is to stop managing more subscriptions than your workflow actually needs. If consolidation is the direction you are leaning, [the best all-in-one AI platforms](https://whizi.io/resources/best-all-in-one-ai-platforms) groups the category by how each option is actually built.

#### Try multiple models in one workspace

The most practical way to choose among ChatGPT alternatives in 2026 is to stop treating model choice as a one-time decision. Your writing model, coding model, research model, and image/document model may not be the same. That is normal.

In Whizi, the workflow is simple: paste the same brief, run it across multiple AI models, compare the outputs, and keep the result that best matches the job. For a marketing brief, compare structure and voice. For code, compare fix strategy and tests. For research, compare how clearly each model separates evidence from assumptions.

If you are still building your AI workflow, use the broader [Whizi resources](https://whizi.io/resources) library as a starting point, then come back to this page when you are ready to compare tools more deliberately.

**Mid-article action:** [Try Whizi for $0.99](https://whizi.io/register) with one real task from your week. Then [run the savings calculator](https://whizi.io/tools/calculator) to see whether consolidating your AI stack makes financial sense.

A good first test takes less than ten minutes:

1. Pick one real task you already need to finish.
2. Write one detailed prompt with context, constraints, and desired format.
3. Run it through two or three model options.
4. Score each answer on usefulness, accuracy, clarity, and time saved.
5. Save the best prompt as a reusable workflow.

#### Common questions

**What is the best ChatGPT alternative overall?**

There is no single best alternative for every user. Claude is often strong for writing and editing, Gemini is worth testing for multimodal and document-heavy work, coding assistants can be best inside development environments, and research-focused tools can help with source discovery. The best choice depends on your workflow.

**Are free ChatGPT alternatives good enough?**

Sometimes. Free tools are good for learning, casual drafting, and light brainstorming. For daily work, paid or consolidated access is often more reliable because you get better limits, stronger models, and smoother file or workflow support.

**Which ChatGPT alternative is best for writing?**

Test Claude, ChatGPT, and Gemini on the same real writing brief. Look for specificity, structure, voice control, and how much editing remains. Do not pick based on one polished sample; pick based on repeatable results.

**Which ChatGPT alternative is best for coding?**

Use the model that helps you reproduce the issue, identify likely causes, make the smallest safe change, and write tests. For serious coding work, always review and run the code yourself.

**Should I subscribe to several AI chatbots?**

Only if each subscription clearly earns its place. Many users are better served by a workspace where they can compare multiple models, then upgrade only when the workflow and savings are obvious.

Key points:

- Compare leading AI chatbots by writing, coding, research, document, and multimodal workflows
- Use a practical evaluation rubric instead of picking a tool because it is trending
- See when a consolidated AI subscription is cleaner than paying for several separate apps

Checklist:

- Define the job before choosing a model: writing, coding, research, documents, images, or daily productivity
- Run the same prompt across at least two AI chatbots before deciding
- Score outputs on usefulness, accuracy, clarity, verification, and time saved
- Calculate your current AI subscription stack before adding another paid plan
- Use Whizi when you want multiple model options in one workspace instead of scattered subscriptions

FAQ:

Q: Do I have to stop using ChatGPT to try an alternative?
A: No. The best answer is rarely to replace ChatGPT everywhere. ChatGPT remains one of the top AI chatbots, and it is still strong for ideation, outlining, and turning rough notes into usable drafts. Treat model choice as a per task decision instead: your writing model, coding model, and research model may not be the same tool.

Q: What should I score when I compare AI chatbots?
A: Score seven things: task fit, output quality, reasoning visibility, context handling, tool support, verification path, and cost and access. Run one real task from your week through each contender with the same inputs, then compare the outputs side by side. Score the workflow, not the first impression, because a polished answer can still invent details.

Q: What makes a good test prompt for comparing AI tools?
A: Use a real task from your week with real constraints, not a vague demo prompt like write a blog post about productivity. Give context, constraints, and the format you want. Strong test prompts also ask for review material: the three biggest edits made, likely root causes plus tests, or evidence separated from open questions.

Q: What is a failure test for an AI chatbot?
A: Ask each tool to identify missing context, list its assumptions, and name what it cannot verify. It shows whether an assistant is reviewable rather than just confident. Pair it with a verification path check: can you inspect sources, reproduce code suggestions, or see how the conclusion was reached? The best assistants make review easier.

Q: How do I know if I am paying for too many AI subscriptions?
A: Add up what you already pay. A stack can quietly grow into one subscription for ChatGPT, one for Claude, one for Gemini, and another for research or image work. Run your current stack through the AI subscription savings calculator, then compare that number against one unified workspace. The goal is fewer subscriptions, not fewer models.

### ChatGPT alternatives for coding: best picks and workflows

URL: https://whizi.io/resources/chatgpt-alternatives-for-coding/
Updated: August 2026

Quick answer: The strongest ChatGPT alternatives for coding are Claude for code review, refactor planning, and tradeoff analysis, and Gemini for long files, screenshots, logs, and other mixed input. ChatGPT stays a solid default for implementation plans and step-by-step debugging. Choose by task rather than brand, and test each one on your own bug, diff, and tests.

Compare ChatGPT alternatives for coding by debugging, refactoring, code review, tests, and workflow fit so you can choose the right AI model for each job.

#### Evaluation criteria

A good ChatGPT alternative for programming is not the model that writes the longest patch. It is the model that helps you ship a smaller, safer change with less confusion. Coding work has a different quality bar than ordinary writing: the answer must fit the existing codebase, preserve behavior, avoid hidden security problems, and include a way to prove the change works.

Start by evaluating AI code assistant alternatives across five criteria: context handling, debugging discipline, implementation restraint, test quality, and review usefulness. The model should use the files, stack, logs, and constraints you provide without inventing missing details. It should ask for a reproduction, propose the smallest useful change, name tests that prove the change, and spot regression risk.

| Criterion | What good looks like | Red flag |
|---|---|---|
| Reproduction | Restates the failing path, expected behavior, and observed behavior | Starts coding from a vague symptom |
| Scope control | Changes the smallest area that explains the bug | Rewrites modules that were not involved |
| Codebase fit | Follows local patterns, naming, framework conventions, and test style | Introduces a new abstraction without a reason |
| Testing | Suggests unit, integration, or regression tests tied to the failure | Says "add tests" without naming cases |
| Review | Calls out tradeoffs, edge cases, and rollback risk | Presents the patch as guaranteed correct |

Official model docs from OpenAI, Anthropic, and Google show that models differ by context windows, tool use, multimodal input, and API behavior. Those capabilities matter, but they do not replace a real coding test. Use your own stack: one bug, one refactor, one review, and one test-writing task.

#### Best picks by scenario

There is no single best AI model for coding in every situation. A model that explains a stack trace well may be weaker at reviewing a large diff. Treat the choice as routing: pick the first model based on the job, then use a second model as the reviewer when the risk is high.

| Scenario | What to optimize for | Model-selection rule |
|---|---|---|
| Debugging a failing test | Root-cause reasoning, logs, minimal fix | Use the model that asks for missing context and ties the patch to the reproduction |
| Refactoring legacy code | Behavior preservation, dependency awareness, staged migration | Use the model that creates a plan before code and names tests for each stage |
| Code review | Regression risk, security, maintainability, edge cases | Use the model that gives specific line-level concerns and avoids style-only noise |
| Writing unit tests | Boundary cases, fixtures, mocks, deterministic assertions | Use the model that maps each test to a behavior claim |
| Explaining unfamiliar code | Plain-language summary, call flow, data ownership | Use the model that separates facts from guesses and points to exact code paths |
| API integration | Docs awareness, input/output contracts, error handling | Use the model that asks for version, endpoint, auth, and failure modes |

ChatGPT remains a strong default for many coding workflows because it is broad, fast, and good at turning a problem into structured steps. Claude is worth testing for code review, refactor planning, long-context reasoning, and tradeoff analysis. Gemini is worth testing when your task includes long files, screenshots, logs, documentation, or multimodal context.

A practical team workflow is to keep three saved prompts: one for debugging, one for refactors, and one for review. When the work is risky, run the prompt in two models inside Whizi and compare which answer makes the fewest assumptions and gives you the most testable path.

#### Workflow: repro -> fix -> tests

The most reliable AI for debugging code workflow is simple: reproduction first, fix second, tests third. Most bad AI coding sessions skip the first step. A better workflow forces the model to reason from evidence.

Step 1: capture the reproduction. Include the failing command, failing test name, exact error, expected behavior, observed behavior, environment details, and the smallest code excerpt that explains the path. For UI bugs, include the route, user action, console error, and network response. For API bugs, include the request, response, status code, and logs.

Step 2: ask for causes before code. A good model should list likely root causes, rank them, and say what evidence supports each one. This slows the session down just enough to prevent a fantasy patch. If the model cannot explain why a cause is likely, it should ask for more context.

Step 3: request the smallest fix. Tell the model not to rewrite unrelated code, change public behavior, introduce new dependencies, or rename things unless necessary. Ask for files touched, functions changed, and why each change is needed.

Step 4: require tests. Ask for a failing test that captures the bug, a passing test after the fix, and at least one edge case. For risky code, ask a second model to review the proposed tests.

Use this debugging checklist before you paste anything into an AI assistant:
- I can name the exact failing behavior.
- I know the command or action that reproduces it.
- I have the relevant logs, stack trace, request, or test output.
- I know what behavior must not change.
- I can identify the files most likely involved.
- I have a test or verification step for the fix.
- I will ask the model for assumptions before accepting code.

This workflow also works for an AI refactoring assistant. Replace "failing behavior" with "behavior to preserve." Ask for a staged plan, public interfaces, invariants, and tests before moving code.

#### Prompt templates

Use these templates as starting points. The bracketed fields matter more than the model name. Strong context produces stronger answers across ChatGPT, Claude, Gemini, and other coding assistants.

Debugging prompt:
`You are a senior engineer helping debug a production-quality codebase. Do not write code yet. First restate the reproduction, expected behavior, observed behavior, and the three most likely root causes. Rank the causes by evidence. Then ask for any missing context. Bug: [describe bug]. Command or user action: [paste]. Error/logs: [paste]. Relevant code: [paste]. Constraints: [stack, style, files not to touch].`

Smallest-fix prompt:
`Based on the reproduction and code below, propose the smallest safe fix. Return: 1) root cause, 2) files/functions to change, 3) patch outline, 4) behavior that must not change, 5) tests that prove the fix. Do not introduce new dependencies or refactor unrelated code. Context: [paste].`

Code review prompt:
`Review this diff like a careful maintainer. Focus on correctness, regression risk, security, edge cases, and missing tests. Ignore minor style unless it affects maintainability. Return a table with issue, risk, evidence, suggested fix, and test needed. Diff: [paste]. Product behavior: [paste].`

Refactor planning prompt:
`Create a staged refactor plan for this code. Goal: [goal]. Constraints: preserve public behavior, minimize churn, follow existing patterns, and keep each stage testable. Return: dependency map, invariants, stages, files touched, tests per stage, rollback risk, and a final review checklist. Code: [paste].`

Unit-test prompt:
`Write test cases for this behavior before changing implementation. Return test names, setup, input, expected output, and why each test matters. Include happy path, boundary case, error case, and regression case. Use the existing test style shown here: [paste example test]. Code under test: [paste].`

Model-comparison prompt for Whizi:
`I am comparing models for a coding workflow. Solve the task using only the context provided. Do not assume missing files. Return root cause, smallest safe fix, tests, risks, and questions. After the answer, grade your confidence from 1-5 and list what would change your recommendation. Task: [paste]. Context: [paste].`

Run the last prompt across models. Compare which answer gives you the cleanest path to a patch, the most relevant tests, and the clearest assumptions. If one model writes the best patch and another gives the best review, use both roles deliberately.

#### CTA blocks

When you are evaluating ChatGPT alternatives for coding, do not rely on benchmark headlines or one-off opinions. Use your own code. Pick a real bug, a real refactor, and a real review. Run the same prompt in multiple models and compare output quality against your engineering checklist.

Whizi is built for that comparison habit. You can keep the prompt fixed, compare model outputs inside one workspace, and decide which answer is safest. That is useful when the choice is not obvious: ChatGPT for a quick implementation plan, Claude for review depth, Gemini for long-context or mixed-input tasks, or another model for a specialized workflow. Those roles map onto the three models most teams already have access to.

| Model | Where it stands out | Reach for it when |
| --- | --- | --- |
| ChatGPT | Broad and fast, and good at turning a problem into structured steps | You want an implementation plan or a way into unfamiliar code |
| Claude | Code review, refactor planning, long-context reasoning, and tradeoff analysis | The change is risky and you want objections before you merge |
| Gemini | Long files, screenshots, logs, documentation, and other multimodal context | The task carries far more context than a pasted snippet |
| A second model as reviewer | A wider review surface on the same fixed prompt | One model wrote the patch and you want the risk checked independently |

If your team is already paying for several AI coding tools, compare the workflow cost too. Start with the broader [ChatGPT vs Claude vs Gemini guide](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini), check the main [ChatGPT alternatives](https://whizi.io/resources/chatgpt-alternatives) guide or the [ChatGPT alternative with Claude and Gemini built in](https://whizi.io/alternatives/chatgpt), then compare plans on [Whizi pricing](https://whizi.io/pricing). When you are ready, [create your Whizi account](https://whizi.io/register) and run the same coding prompt across models.

Key points:

- The best AI for coding depends on the task: debugging, refactoring, reviewing, tests, or architecture.
- Use a repro -> fix -> tests loop so the model reasons from evidence instead of guessing.
- Compare model outputs in Whizi before you standardize on one coding assistant workflow.

Checklist:

- Use a real bug, refactor, review, and test-writing task to evaluate coding models.
- Require the model to restate the reproduction before proposing a fix.
- Ask for root-cause options and evidence before accepting code.
- Prefer the smallest safe patch over broad rewrites.
- Require tests that would fail before the fix and pass after.
- Use a second model to review risky patches, refactors, and missing edge cases.
- Compare model outputs in Whizi before paying for another standalone AI coding subscription.

FAQ:

Q: What is the best ChatGPT alternative for coding?
A: The best ChatGPT alternative for coding depends on the task. Claude is often worth testing for code review and refactor reasoning, while Gemini is worth testing for long-context, document-heavy, or multimodal workflows. The safest approach is to compare models on your own bug reports, diffs, and tests.

Q: Can AI write unit tests for code?
A: Yes, AI can help draft unit tests, but you should require specific behavior coverage. Ask for happy path, boundary, error, and regression cases, then review whether each test would actually fail before the fix and pass after it.

Q: How should I use AI for debugging code?
A: Use a repro-first workflow. Provide the failing command, logs, expected behavior, observed behavior, and relevant code. Ask the model to identify likely causes before writing code, then request the smallest fix and tests.

Q: Should developers use more than one AI coding model?
A: Often, yes. One model may be stronger at drafting a fix while another is better at reviewing risk. For important work, run the same prompt across models and use the output that is easiest to verify.

### ChatGPT alternatives for research: best tools for cited work

URL: https://whizi.io/resources/chatgpt-alternatives-for-research/
Updated: August 2026

Quick answer: For research, judge ChatGPT alternatives on traceability rather than fluency: whether the tool cites the specific source behind each claim, summarizes long documents without flattening nuance, and admits when a source does not support a conclusion. Work in stages, with extraction before synthesis, and run the same source pack across models to see which output is easiest to verify.

Compare ChatGPT alternatives for research by source handling, traceability, synthesis quality, and workflow fit. Includes prompts and verification templates.

#### What research requires: sources, traceability, and restraint

The best ChatGPT alternatives for research are not simply the chatbots that sound the most confident. Research work has a different quality bar than brainstorming or everyday productivity. A useful AI research assistant needs to show where claims came from, separate source evidence from interpretation, preserve uncertainty, and make the final answer easy to verify.

Most research mistakes look polished: one unsupported statistic, one outdated product claim, or one missing caveat. In a literature review, that can distort the argument. In market research, it can send a founder toward the wrong segment. In competitive analysis, it can turn an old pricing page into a false positioning insight.

So when you compare ai research assistant alternatives, evaluate the workflow. Does the tool help you collect sources? Can it summarize long documents without flattening nuance? Does it cite the specific source behind each claim? Does it admit when a source does not support the conclusion?

A strong research workflow has four layers. First, define the question tightly. Second, gather source material and label it clearly. Third, extract evidence before asking for synthesis. Fourth, verify the answer against the original sources. OpenAI, Anthropic, and Google all publish model and long-context guidance, but the practical lesson is consistent: structure the input, make the model work from source material, and review the output before you trust it.

Whizi is useful here because research is rarely a one-model job. One model may create a cleaner plan, another may extract better from long documents, and another may write the clearest synthesis. Run the same source pack across models and keep the answer that is most traceable.

#### Best picks by research scenario

There is no universal best ai for research. The right choice depends on the job: web scan, literature review, market map, PDF summary, interview analysis, or decision memo. Use the table below as a practical starting point, then test your own prompts inside Whizi.

| Research scenario | What matters most | Best-fit workflow | What to verify |
|---|---|---|---|
| Market research | Competitor claims, positioning, pricing signals, customer language | Collect source pages, extract claims into a table, synthesize patterns | Current pricing, customer segment, date of source, whether claims are from the company or customers |
| Literature review | Accurate source summaries, terminology, methods, limitations | Summarize each paper separately, extract key findings, group by theme | Citations, methodology, sample size, whether the AI overstates a finding |
| Deep research brief | Multi-source synthesis and uncertainty tracking | Start with a research question, build a source pack, ask for claims with citations, then synthesize | Unsupported claims, missing counterevidence, stale sources |
| PDF or document summary | Long-context handling and structured extraction | Ask for outline, entities, claims, evidence, and open questions before summary | Whether each important point appears in the original document |
| Customer or interview analysis | Exact language and theme grouping | Extract verbatim phrases, tag pains and desired outcomes, then synthesize | Invented quotes, overgrouped themes, missing outliers |
| Competitive positioning | Differences that buyers can understand | Compare messaging, feature emphasis, proof points, and pricing page language | Whether the comparison uses the same evidence type across competitors |

For cited answers, avoid asking "What is the answer?" too early. Ask for a source table first: title, type, date, key claims, relevant passage, confidence, and caveats. Only after that should you ask for synthesis.

For literature reviews, use two passes. First summarize each source independently. Then compare sources by theme. This reduces the chance that the model blends findings together or attributes one paper's conclusion to another paper.

For market research, start with the [Founder Research Stack](https://whizi.io/templates/founder-research-stack). It gives you a reusable structure for competitor scans, pricing analysis, messaging extraction, and synthesis. If you are researching a category, a customer segment, or a new product direction, the template is usually more valuable than a blank chat box.

If your next project is competitor, pricing, or positioning research, pair this article with the guide to [AI for market research](https://whizi.io/resources/ai-for-market-research). If it is academic work instead, [the researcher workspace](https://whizi.io/use-cases/researchers) covers long-context reading and cited drafting. Then bring the source pack into Whizi and compare outputs.

#### Workflow: question to sources to synthesis

The fastest way to improve AI research quality is to stop treating research as one prompt. Use this workflow instead: question, source map, extraction, claim table, synthesis, verification.

| Stage | What you produce | What it prevents |
| --- | --- | --- |
| Question | A testable question with audience, decision, scope, and output named | A broad request that returns an equally broad answer |
| Source map | The list of source types you need, written before you collect them | Research shaped by whatever was easiest to find |
| Extraction | Facts, claims, quotes, caveats, and contradictions in a table | Conclusions written before the evidence is on the page |
| Claim table | Every claim tied to a source, evidence snippet, confidence, and verification note | Inference presented as sourced fact |
| Synthesis | What the sources show, what they suggest, and what stays unknown | Certainty the sources do not support |
| Verification | A manual read of the passages behind the highest-impact claims | Stale dates, wrong pricing, and invented quotes |

Step 1: Define the research question. A weak question is broad: "Research the market." A stronger question is testable: "What pricing and positioning patterns appear across AI meeting note tools for small teams in 2026?" Add audience, decision, scope, and output.

Step 2: Build a source map. List the source types you need before collecting them. For market research, that might include homepages, pricing pages, docs, reviews, customer interviews, and comparison pages. For a literature review, it might include papers, abstracts, methods sections, datasets, and review articles.

Step 3: Extract before synthesis. Ask the model to extract facts, claims, quotes, caveats, and contradictions into a table. Do not ask for conclusions yet. Extraction keeps the model close to the source material and makes verification easier.

Step 4: Build a claim table. Every major claim should have a source, evidence snippet, confidence level, and verification note. If the model cannot point back to a source, mark the claim as inference or remove it.

Step 5: Synthesize with constraints. Require the synthesis to distinguish what the sources show, what they suggest, and what remains unknown. Ask for implications, caveats, and next research steps.

Step 6: Verify manually. Read the passages behind the most important claims. Check dates, pricing, product names, definitions, sample sizes, and any claim that would affect a decision.

Mid-workflow CTA: use the [Founder Research Stack](https://whizi.io/templates/founder-research-stack) to turn this process into a reusable workspace. Then [create your Whizi account](https://whizi.io/register) to run the same prompt stack across models.

#### Template: research prompt stack

Use this prompt stack when you need cited, traceable research. It works for market research, literature review notes, competitive analysis, customer language research, and long-document synthesis. Replace the bracketed fields and run the same prompts across models in Whizi.

Prompt 1: Research plan. "Act as a careful research analyst. I am researching [topic] to decide [decision]. Audience: [audience]. Scope: [scope]. Create a plan with key questions, source types, exclusion criteria, risks, and final deliverable format. Do not answer yet."

Prompt 2: Source intake. "Using only the sources below, create a source inventory table. Columns: source ID, title, source type, date, author or company, useful evidence, reliability concerns, and questions this source can answer. Sources: [paste material]."

Prompt 3: Source extraction. "Extract evidence from the source pack. Columns: source ID, exact claim, supporting passage, topic tag, confidence, caveat, and evidence type. Do not synthesize yet. Do not invent missing details."

Prompt 4: Claim check. "Review the extraction table. Flag claims that are unsupported, stale, vague, duplicated, contradicted by another source, or too broad. Suggest what source would be needed to verify each weak claim."

Prompt 5: Synthesis. "Now synthesize the research for [audience]. Use only the extracted evidence. Structure the answer as: executive summary, strongest findings, evidence table, counterevidence or caveats, implications, and recommended next research. Mark every important claim with its source ID. Clearly label inference versus directly supported evidence."

Prompt 6: Verification pass. "Act as a skeptical reviewer. Audit this research synthesis for unsupported claims, citation mismatch, missing caveats, outdated evidence, overgeneralization, and decision risk. Return: issues to fix, claims to verify manually, and a revised version that is more careful."

This stack is slower than a single prompt. That is the point. Good research separates collection, extraction, synthesis, and verification so the final answer is easier to inspect.

#### Source extraction, synthesis, and verification template

Use this table format whenever the research needs to survive scrutiny. It keeps the model from hiding uncertainty inside polished prose.

| Field | What to capture | Why it matters |
|---|---|---|
| Source ID | A short label like S1, S2, S3 | Lets you cite and audit claims quickly |
| Source type | Paper, pricing page, interview, docs, review, report | Helps distinguish evidence quality |
| Date | Published or updated date when available | Prevents stale claims from driving decisions |
| Claim | One specific statement from the source | Avoids vague summaries |
| Evidence | Quote, passage, table, metric, or observation | Keeps synthesis tied to source material |
| Confidence | High, medium, low | Forces uncertainty into the open |
| Caveat | Limitation, missing context, possible bias | Stops overclaiming |
| Use in synthesis | Include, exclude, verify, or background only | Makes the final answer cleaner |

Run this verification pass before you ship anything:

- Every major claim points to a source, and every source ID actually supports the sentence it is attached to.
- Dates are checked, and any pricing or product claim is current rather than remembered.
- No quote is invented. Check them against the original, not against the summary.
- Limitations are visible instead of smoothed over.
- Counterevidence is included, because a brief with no disagreement in it is usually incomplete.
- Recommendations carry a confidence level, so the reader knows what is solid and what is a guess.

This is where comparing models helps. Run the same extraction table through two models and ask each to audit the other for missed caveats and weak evidence.

#### Use Whizi for repeatable research

The practical reason to use Whizi for research is simple: stop guessing which model is "best" and test which model is best for this source pack, question, and deliverable.

Start with the [Founder Research Stack](https://whizi.io/templates/founder-research-stack), paste in your research question and source material, then run the extraction prompt across models. Compare which output is easiest to verify. Then run the synthesis prompt and keep the answer that is clearest, most cautious, and most useful.

When the workflow is working, save it. Your goal is a repeatable system for cited answers, market scans, literature notes, customer language analysis, and decision memos. Create an account at [Whizi](https://whizi.io/register), compare plans at [pricing](https://whizi.io/pricing), or read the [ChatGPT alternative overview](https://whizi.io/alternatives/chatgpt) to see what one multi-model plan replaces.

Key points:

- Choose research tools by traceability, source handling, and synthesis quality instead of brand preference.
- Use a question-to-sources-to-synthesis workflow for market research, literature reviews, competitive analysis, and source summaries.
- Copy a research prompt stack that separates extraction, claims, citations, synthesis, and verification.

Checklist:

- Define the research question before asking for an answer.
- Create a source map that lists the evidence you need and the source types you trust.
- Ask for source extraction before asking for synthesis.
- Require every important claim to include a source ID, confidence level, and caveat.
- Label direct evidence separately from inference.
- Verify dates, pricing, quotes, product claims, and statistics manually.
- Use the Founder Research Stack for repeatable market and competitor research.
- Run the same prompt stack across models in Whizi and keep the most traceable output.

FAQ:

Q: What is the best ChatGPT alternative for research?
A: The best option depends on the workflow. Prioritize source handling, long-context performance, citation discipline, and verification support. In Whizi, compare the same source pack across models.

Q: Can AI cite sources accurately?
A: AI can help organize citations and source notes, but verify important citations yourself. A cited sentence is only trustworthy if the source supports the claim.

Q: How should I use AI for literature review work?
A: Summarize each source separately, extract methods and findings into a table, compare sources by theme, then write the synthesis. Avoid asking for a broad literature review before the source-level extraction is complete.

Q: How do I use AI for market research?
A: Start with a specific question, collect competitor and customer sources, extract pricing and messaging claims, synthesize patterns, and verify high-impact claims manually.

### ChatGPT alternatives for writing: best picks in 2026

URL: https://whizi.io/resources/chatgpt-alternatives-for-writing/
Updated: August 2026

Quick answer: Claude is the strongest ChatGPT alternative for polished prose and careful editing, Gemini for document-heavy drafting, and ChatGPT itself for fast variants. The choice depends on writing type, so run one identical brief across models and score each draft on clarity, voice, factual discipline, and how much editing remains.

Compare ChatGPT alternatives for writing emails, marketing copy, docs, and long-form work with prompts, scoring criteria, and an editing checklist.

#### How we evaluate writing output

The best ChatGPT alternative for writing is not the model that sounds most impressive in a demo. It is the one that turns your actual brief into the cleanest usable draft with the least human cleanup. Writing quality is contextual: an email needs judgment and brevity, marketing copy needs audience tension and specificity, documentation needs accuracy and structure, and long-form content needs coherence over many sections.

Use this scoring rubric before you pay for another AI writing assistant. Score every output from 1 to 5 on six criteria: clarity, audience fit, voice control, factual discipline, structure, and edit time. A polished paragraph that invents claims should lose. A plain draft that follows the brief, keeps facts intact, and gives you a strong editing base should win.

Here is the practical test: write one detailed brief, run the same brief across models, and compare outputs side by side. Include the audience, goal, source notes, constraints, tone, length, and desired format. Then ask: Which version would I actually send, publish, or hand to a teammate after one editing pass?

| Evaluation factor | What good looks like | Red flag |
| --- | --- | --- |
| Clarity | The main point is obvious on the first read | Pretty sentences hide the recommendation |
| Voice | It sounds like your brand or person, not generic AI | It uses vague phrases like "in today's fast-paced world" |
| Structure | The piece has a useful order and scannable flow | The answer is a wall of plausible paragraphs |
| Specificity | It uses your examples, constraints, and audience language | It replaces detail with buzzwords |
| Factual discipline | It keeps claims grounded in your notes or flags uncertainty | It adds unsupported stats, features, or promises |
| Edit time | You can improve it quickly | You have to rewrite the whole thing |

Official model docs from OpenAI, Anthropic, and Google all point to a practical reality: model lineups change, capabilities differ, and good prompting matters. For writing, your workflow matters as much as your tool choice.

#### Best picks by writing type

There is no single best AI writing tool for every job. The right choice depends on what you are writing, how much source material you have, how strict the tone needs to be, and whether you need many variations or one polished draft.

### Email

For email writing, choose the model that understands social context. A good email draft is not just grammatical; it gets the ask right, respects the relationship, and removes friction. Test each writing assistant with the same follow-up, client update, delicate disagreement, or short sales reply.

Claude is often worth testing when the email needs nuance, diplomacy, or a polished human tone. ChatGPT is often useful when you want several versions quickly. Gemini can help when the email depends on document context or material you need summarized first. The winner is the draft that preserves facts, lowers tension, and makes the next step unmistakable.

### Marketing

For marketing copy, the common failure is generic confidence. Many AI tools can produce a headline, landing page section, or ad concept that sounds like marketing but says little. A strong chatgpt alternative for marketing copy should ask for the audience, pain, desired outcome, proof, offer, objection, and channel. It should create options that differ by angle, not just by adjective.

Use AI to generate positioning routes, headlines, email sequences, product descriptions, and ad variants, but judge the output by specificity. Does the copy name a real customer problem? Does it include proof? Does it avoid claims your product cannot support? If not, the model is giving you polish, not strategy.

### Docs

For documentation, internal guides, SOPs, support articles, and writing reports, accuracy beats style. The best ai for writing reports should organize evidence, preserve source details, and make missing information visible. Long-context support can matter here because docs often rely on meeting notes, specs, tickets, transcripts, or research packets.

A strong docs workflow has three steps. First, ask the model to extract facts from the source material before drafting. Second, ask for an outline with assumptions and gaps. Third, ask for the final doc in the required format. For reports, add a final pass that separates findings, evidence, interpretation, and recommendations.

### Long-form

Long-form writing is where many AI writing assistant alternatives look good for two paragraphs and then drift. The test is not whether the introduction sounds fluent. The test is whether the article, guide, memo, or thought leadership piece keeps a consistent argument, avoids repetition, and uses examples that make the piece worth reading.

For ai for long form writing, use a staged workflow instead of asking for a full article in one prompt. Start with the thesis and reader problem. Build an outline. Add source notes. Draft section by section. Then run a cohesion pass. Give the model a map before asking it to write the journey.

#### Prompt pack

Use this reusable prompt pack to compare ChatGPT alternatives for writing without changing the test between models. Paste the same prompt into each model, then score the output with the rubric above.

**1. Same-brief model comparison**

`You are a senior editor. I am comparing AI writing models. Use the brief below to create the requested deliverable. Do not add facts that are not in the brief. If important information is missing, mark it as [needs input] instead of inventing. Return: 1) the draft, 2) three editing notes, 3) claims that need verification. Brief: [paste brief]. Audience: [audience]. Goal: [goal]. Tone: [tone]. Format: [format]. Length: [length].`

**2. Email rewrite**

`Rewrite this email so it is clearer, shorter, and more useful to the recipient. Keep the relationship warm but do not over-apologize. Preserve all facts, dates, names, and commitments. Return a subject line, the revised email, and a one-sentence reason for the tone. Email: [paste email].`

**3. Marketing copy angles**

`Create five distinct marketing angles for this product. For each angle, include the audience pain, promise, proof needed, headline, subhead, and CTA. Avoid generic claims. If proof is missing, say what evidence we need before publishing. Product notes: [paste notes]. Audience: [audience]. Channel: [landing page/ad/email].`

**4. Long-form outline**

`Turn this topic into a useful long-form outline. Start with the reader's search intent. Then produce H2 sections, the job of each section, examples to include, internal links, and evidence needed. Topic: [topic]. Primary keyword: [keyword]. Source notes: [paste notes].`

**5. Report draft from notes**

`Create a report from these notes. Separate findings, evidence, interpretation, risks, and recommendations. Use concise headings. Do not invent numbers or sources. End with a list of missing inputs that would improve the report. Notes: [paste notes].`

**6. Voice match**

`Rewrite the draft to match the example voice. Preserve meaning and factual claims. Match sentence length, directness, level of detail, and vocabulary. Do not copy unique phrases unless they are product terms. Example voice: [paste example]. Draft: [paste draft].`

**7. Human editing pass**

`Edit this draft like a strict but practical editor. Return: the biggest structural issue, the three most generic lines to replace, unsupported claims, a tighter draft, and a final checklist before publishing. Draft: [paste draft].`

The important move is consistency. If you give one model a better prompt than another, you are testing prompt quality instead of model fit. In Whizi, run the same brief across models, compare drafts, keep the best parts, and save the prompt that performed well.

#### Editing checklist

AI can get you to a draft faster, but publishing still requires human judgment. Use this checklist for emails, docs, reports, marketing copy, and long-form content before anything leaves your workspace.

- **Purpose:** Is the reader supposed to decide, understand, reply, click, approve, or act?
- **Audience:** Does the piece use the reader's language and knowledge level?
- **Specificity:** Replace generic claims with concrete examples, constraints, numbers, or proof.
- **Factual review:** Check every product claim, date, price, feature, statistic, quote, and citation.
- **Voice:** Remove phrases your team would never say. Add examples of your real style when needed.
- **Structure:** Move the main point higher. Cut sections that repeat the same idea.
- **Risk:** Flag legal, medical, financial, compliance, privacy, or customer-sensitive claims for expert review.
- **Conversion:** Make the next step clear, whether it is reply, book, sign up, read more, or compare plans.
- **Final pass:** Read it out loud once. If it sounds like a brochure for "innovative solutions," cut harder.

For commercial writing, also check search intent. Readers do not only want a list of brands. They want to know which tool helps with their email, campaign, report, or long-form draft. Compare outputs by writing type instead of treating every AI assistant as interchangeable.

When you are ready to test this for real, pick one writing task you already need to finish. Run the same prompt across models in Whizi. Keep the output that best matches your audience, then use the checklist above to make it publishable.

#### Run the same brief across models

A side-by-side writing test is the cleanest way to choose among AI writing assistant alternatives. Here is the full workflow.

First, create one brief with the reader, job to be done, source notes, tone, length, must-include points, must-avoid points, and final format. Second, run that exact brief across at least two models. Third, score each draft using the writing rubric. Fourth, ask the strongest draft for a revision using the strongest idea from another draft. Fifth, save the prompt and model pairing.

This works because different models often win different jobs. One may be better at warm emails, another at structured reports, and another with long source material. You do not need one permanent choice. You need a repeatable way to get the best draft for the task in front of you.

Whizi is built for that kind of writing workflow: compare outputs, reduce tab switching, and avoid paying for every separate AI writing subscription before you know what actually improves your work. Start with one prompt from this article, test it on a real draft, and let the result decide. If the test convinces you that you want Claude and GPT together rather than either alone, the [ChatGPT alternative with both built in](https://whizi.io/alternatives/chatgpt) covers that in one plan.

Key points:

- A practical writing-score rubric for judging AI output by clarity, voice, structure, and edit time
- Best-fit guidance for email, marketing copy, docs, reports, and long-form content workflows
- A reusable prompt pack and side-by-side model test you can run inside Whizi

Checklist:

- Score AI writing output by clarity, voice, structure, factual discipline, specificity, and edit time
- Use one identical brief when comparing ChatGPT, Claude, Gemini, or other writing tools
- Choose models by writing type: email, marketing, docs, reports, or long-form content
- Ask for claims that need verification before publishing AI-assisted writing
- Save the prompts that work so your writing workflow becomes repeatable

FAQ:

Q: What is the best ChatGPT alternative for writing?
A: The best choice depends on the writing task. Claude is often worth testing for polished prose and careful editing, ChatGPT for fast ideation and variants, and Gemini for document-heavy or multimodal workflows. Run the same brief across models before choosing.

Q: How do I make AI writing sound less generic?
A: Give the model a real audience, examples of your voice, source notes, constraints, and a clear format. Then edit for specificity by replacing vague claims with concrete proof, examples, and sharper reader language.

Q: Should I use different AI models for different writing tasks?
A: Yes. Email, marketing copy, documentation, reports, and long-form content reward different strengths. A side-by-side model test helps you find the best model for each repeatable writing workflow.

### ChatGPT says I've reached my limit: what to do next

URL: https://whizi.io/resources/chatgpt-limit-reached/
Updated: August 2026

Quick answer: Read the exact wording first, because three different messages look alike. A cap on one model means switching to the fallback it offers or to another provider. Maximum conversation length is a context limit: summarize the thread into a new chat and continue immediately. A feature cap leaves text chat working.

You hit a wall mid task. Here is which limit you actually hit, when it resets, what still works while you wait, and how to stop it happening again.

#### What to do in the next sixty seconds

Read the exact words on screen before you do anything else. ChatGPT shows at least three different messages that all feel like "you are out of ChatGPT", and they have nothing to do with each other. Getting the right one takes five seconds and saves you from waiting for a reset that was never going to fix your problem. OpenAI revises the exact wording, so match the sense of what you are seeing rather than the letter of the table.

| What the message says | What actually happened | The fix that works |
| --- | --- | --- |
| You have hit your limit on this model, or it offers you a smaller model instead | A rate cap on that specific model in a rolling time window | Keep working on the fallback model it offers, or switch to another provider now |
| You have reached the maximum length for this conversation | The chat is too long for the model to hold in one go, which is a context limit and not a usage cap | Ask for a summary of the thread, paste it into a new chat, and carry on immediately |
| A limit tied to one feature: image generation, file uploads, voice, deep research | A separate cap on that feature alone | Text chat still works. Do the writing part now and the image or file part after the reset |

If it is the middle row you are not blocked at all, and you can be working again in about thirty seconds. Send "Summarize everything we have decided in this chat, including the constraints and the current draft, so I can paste it into a fresh conversation." Open a new chat, paste, keep going. That row is the most commonly misdiagnosed of the three, because the wording sounds punitive when it is really an out of space notice.

If it is the first or third row, you are waiting on a clock. The next sections explain what that clock is doing, because it does not work the way most people assume, and the assumption is what makes the wait feel longer than it is.

Two nearby problems look almost identical on screen and are not this one. If nothing mentions your plan or your limit and every conversation errors out including brand new ones, that is an outage, and [what to use when ChatGPT is down](https://whizi.io/resources/what-to-use-when-chatgpt-is-down) is the right page. If nothing is blocked at all but the answers suddenly got shorter, flatter and more forgetful, you were probably moved onto a smaller fallback model without being told, which [why ChatGPT switched you to a weaker model](https://whizi.io/resources/why-chatgpt-switched-to-a-weaker-model) explains. A cap names you, an outage does not know who you are, and a downgrade says nothing at all.

#### Free, paid, and what changed in August 2026

Something shifted very recently, and much of the advice you will find on this query is now out of date. On 6 August 2026 OpenAI announced that Free and Go accounts get unlimited text chats, rolling out the following week. TechCrunch reported it the same day, and the change explicitly does not remove the other caps: separate limits remain for files, images, voice, and image generation. Engadget's rundown adds deep research, Codex, and Work to that list.

**If you are on a free account in late 2026 and you hit a limit, it is probably not a chat cap.** It is far more likely a feature cap, or the conversation length limit, or an abuse guardrail firing on unusual traffic. People who spend an afternoon waiting for a message quota to refill are often waiting for something that no longer applies to them.

Paid tiers are a different story, but not on chat. OpenAI's pricing page lists everyday text chats as unlimited on Free, Go, Plus and Pro alike, so upgrading does not buy you more messages. What still has ceilings, and tightens as you go down the tiers, is file uploads, image generation, voice, deep research, memory, and Codex access. That is why this article does not print a number for those limits: any figure published today has a real chance of being wrong by the time you read it, and a wrong number sends you off to wait for a reset that is not coming. Your own account is the closest thing to a current source. Engadget reported in August 2026 that you can see what is left of your limit under Settings, then Usage. OpenAI does not document that screen in its help pages, so treat it as the first place to look rather than a guarantee, and note that some accounts see nothing there until they actually hit a wall.

For orientation only, the individual plan ladder in August 2026 runs Free, Go at about 8 dollars a month, Plus at about 20, and Pro above that, with the actual price varying by country. If text chat is the whole of what you do, the August change means the free tier may now be enough, and upgrading buys you nothing you were missing. The upgrade case is about the capped features and about which models you can reach, not about running out of messages. [Is ChatGPT Plus worth it](https://whizi.io/resources/is-chatgpt-plus-worth-it) works through the break even math instead of assuming it.

#### What resets, and when

The reset mechanism is not fixed, and it changed recently. For a long time OpenAI metered usage on a rolling window of a few hours, where each message aged out and freed its slot, so waiting twenty minutes often bought you another message or two. Engadget's August 2026 rundown reports that OpenAI appears to have dropped that structure around July 2026, and that the remaining feature caps now run on a single weekly allowance instead.

That difference matters for what you do next. Under a rolling window, a short wait is worth trying. Under a weekly allowance, it is not: capacity comes back when the week turns over, and refreshing every ten minutes buys you nothing. So the first move is not to guess which one you are on, it is to look. If Settings, then Usage, shows you a remaining allowance and a reset, that figure beats anything printed here.

Two things stay true whichever structure is in force:
- **Nothing resets at midnight.** Neither a rolling window nor a weekly allowance is tied to your local calendar day, so "try again tomorrow" is a guess rather than a plan.
- **A burst costs the same as a steady pace, but hurts more.** Spending an allowance in one sitting leaves you with nothing for the rest of the period. The same volume spread out is far less likely to strand you mid task.

Anthropic frames Claude's limits differently, which is a useful corrective to thinking in message counts at all. Its help center says usage depends on the length and complexity of your conversations, the features you use, which model you are on, and the effort level you picked, and that usage across claude.ai, Claude Desktop and Claude Code all draws on the same budget. What you spend is a measure of work done, not a tally of messages, so a long chat with three attached files costs far more than a one line question.

Neither OpenAI nor Anthropic publishes exact allowance numbers in its help pages, and OpenAI appears to have changed the reset structure in July 2026 without announcing it. Any article quoting you a precise message count is quoting something it cannot know.

#### Feature limits are separate, which is good news

Image generation, file uploads, voice, and deep research each carry their own budget, and running out of one does not touch the others. Use up your image quota and the text side of the product is completely untouched. So when you hit a wall, ask what you were actually spending on, and split the task:
- **Blocked on image generation?** Write and lock the copy, the brief, or the prompt text now. Generate the visual after the reset, or generate it with a different image model somewhere else.
- **Blocked on file uploads?** Paste the relevant section as plain text instead. Most of the time you needed four paragraphs of a document, not the whole PDF, and pasting has no upload cost at all.
- **Blocked on deep research?** These carry the tightest budgets of anything on the list. Do the reading yourself and use the model to structure and challenge it.
- **Blocked on voice?** Type. It is slower and it always works.

One thing not to do: a second account to double your allowance breaks the terms of service of every major provider, and the usual outcome is losing both accounts and their chat history at a bad moment. A shared or resold login is worse, because whoever controls it can read your conversations. The routes that hold up are a higher tier, spreading work across the day, or a second provider under your own account.

#### How to stop landing here again

A few habits cut the frequency sharply, and none of them require paying anyone more:

1. **Start a new chat per task.** Long threads use up context, cost more per message, and are what triggers the maximum length message. A fresh chat with a two line brief usually beats a fifty turn thread anyway.
2. **Front load the context.** One well specified message beats six clarifying rounds. Say who it is for, how long you want it, and what you will do with it.
3. **Do not attach what you do not need.** Uploading a 200 page PDF to ask about one clause spends your file budget and buries the clause. Paste the clause.
4. **Reserve the expensive features.** Deep research and image generation have the tightest budgets. Do not spend a research run on something a normal answer covers.
5. **Keep an account with a second provider.** There is no shared counter between providers, so a free account elsewhere costs nothing and turns a hard stop into a five second detour. This is not a duplicate account with the same provider, which is the thing that gets you banned. Check the provider's own supported countries page first, because Claude and Gemini do not operate everywhere, and a refused sign up in your country is a location limit rather than a payment problem. [ChatGPT alternatives](https://whizi.io/resources/chatgpt-alternatives) compares the options by use case.
6. **Watch when you get throttled.** If it is always the same point in your week, you have found the burst that causes it, and spreading that work out is easier than paying to absorb it.

One last thing. When a provider offers a smaller fallback model instead of cutting you off, that is not a punishment and it is often perfectly adequate: fallback models handle rewriting, summarizing, and ordinary questions fine. Save the flagship for the reasoning you actually needed it for and you will hit the ceiling far less often.

Key points:

- Read the exact wording of the message first: chat caps, feature caps, and conversation length limits look similar and have completely different fixes
- How caps reset is not stable: OpenAI ran a rolling multi hour window until mid 2026, and August 2026 reporting describes a single weekly allowance instead, so your own account is the only reliable source
- Nothing stops you using a different model in the meantime, because caps are set per provider and never move together

Checklist:

- Read the exact message: a rate cap, a conversation length limit, and a feature cap are three different problems.
- If it says maximum conversation length, summarize the thread and paste it into a new chat right now.
- Check Settings, then Usage, for a stated reset instead of assuming a short wait will do it.
- Check whether text chat still works before assuming the whole account is blocked.
- Paste text instead of uploading files when the file budget is the thing that ran out.
- Spread heavy work across the period instead of spending the whole allowance in one sitting.
- Never make a duplicate account or buy a shared login to get more capacity.
- Keep a working account with a second provider, where one is available in your country, so a cap costs you seconds rather than the rest of the day.

FAQ:

Q: How long until my ChatGPT limit resets?
A: Look it up in your own account rather than in an article, because the structure itself changes. OpenAI metered on a rolling window of a few hours for years, and Engadget reported in August 2026 that the remaining feature caps had moved to a single weekly allowance around July 2026. Nothing resets at your local midnight under either scheme. Engadget also reported that Settings, then Usage, shows what is left of your limit, so check there first, though OpenAI does not document that screen and not every account displays a reset time.

Q: Does the free plan still limit how many messages I can send?
A: OpenAI announced on 6 August 2026 that Free and Go accounts get unlimited text chats, rolling out the following week, with abuse safeguards still in place. Separate limits remain for file uploads, image generation, voice, and deep research. So a limit message on a free account in late 2026 is more likely a feature cap or a conversation length limit than a chat quota.

Q: What does "maximum length for this conversation" mean?
A: That is a context limit, not a usage cap, and paying more rarely moves it much, since the ceiling belongs to the model rather than to your subscription. The thread has grown longer than the model can hold at once, counting your messages, its replies, and any attached files. Nothing resets and waiting does not help. Ask for a summary of the decisions and the current draft, paste it into a new chat, and continue.

Q: Will paying for a higher plan remove the limits?
A: Not on everyday text chat, which OpenAI's pricing page lists as unlimited on Free, Go, Plus and Pro alike. What paying raises is the ceiling on features: file uploads, image generation, voice, deep research, memory, and Codex access, plus which models you can reach. Before upgrading, check whether you were hitting a feature cap or the conversation length limit, since both are usually cheaper to solve than a subscription.

Q: Do limits on ChatGPT, Claude, and Gemini reset together?
A: No. Each provider meters its own service independently, with its own windows and its own rules, and there is no shared counter between them. Exhausting one leaves the others completely untouched.

### ChatGPT is not available in your country: what that message actually means

URL: https://whizi.io/resources/chatgpt-not-available-in-your-country/
Updated: August 2026

Quick answer: That message means OpenAI does not offer ChatGPT where it thinks you are, and nothing is broken on your device. The check can fire on your IP address, on the country code of your phone number, or at payment. Claude and Gemini publish separate country lists, so check those next.

Why ChatGPT blocks signups by country, how the IP check and the phone number check are separate gates, why unsupported and sanctioned are different problems, and what is legitimately available instead.

#### The short answer

That message means OpenAI has decided it does not offer ChatGPT where it thinks you are. Nothing is broken on your device and clearing your cache will not help. OpenAI publishes a list of the countries and territories where it supports access, and if your location is not on it, signup and login are refused. The page carrying that list is blunt about going around it: "Accessing or offering access to our services outside of the countries and territories listed below may result in your account being blocked or suspended."

Two things decide what you can do next. First, availability is checked in more than one place, so a block can come from your network or from your phone number independently. Second, there is a real difference between a country a company has not launched in yet and a country under a comprehensive United States sanctions program. The first can change next quarter. The second is law that no consumer app can waive for you.

If ChatGPT specifically is unavailable to you, that does not mean AI is. The country lists for Claude and Gemini are written by different companies under different legal advice, and they do not match.

#### Three separate gates: your network, your phone number, your card

People describe this error as one wall, but it is usually three checks in sequence, and knowing which one you hit explains why the symptoms look inconsistent.

- **The network check.** The service reads the IP address your connection presents and maps it to a country. If that country is not on the supported list, the page refuses before you have typed anything. This is the error on the very first screen.
- **The account check.** Even when the page loads, creating an account can require verifying a phone number, and the country code of that number is its own signal. A number issued in an unsupported country can fail this step regardless of where you are sitting.
- **The payment check.** On a paid plan, the card country and billing address are checked again by the payment processor, which follows its own compliance rules and may refuse where the app did not. A decline here is often a bank setting rather than a country block, and [why your card gets declined for AI subscriptions](https://whizi.io/resources/paying-for-ai-when-your-card-is-declined) separates the two.

This is why forum reports contradict each other. They are describing different gates. It also explains the reverse case, where a country is on the supported list and access still fails: that is usually a network level block imposed inside the country rather than by OpenAI, and OpenAI cannot lift it either. The list feels arbitrary because it is a business and legal decision rather than a technical one.

#### Not on the list is not the same as sanctioned

These two situations produce the same error message and have completely different outlooks.

A country can be missing for ordinary commercial reasons: no local payment support, no legal entity, an unresolved regulatory question, or simply a launch queue. Lists like this grow. Companies add countries quietly, and the page that refused you in January may accept you in September. Waiting is a real strategy here, not a polite way of dismissing the question.

Sanctions are a different category. The United States maintains a small number of comprehensive country or region programs that broadly prohibit exporting services, so a US company refusing an account there is following the law, not making a customer service decision. As of August 2026 those programs cover Cuba, Iran and North Korea, along with the Crimea, Donetsk and Luhansk regions of Ukraine, which is why OpenAI lists Ukraine "with certain exceptions". Some providers apply the same treatment to the Kherson and Zaporizhzhia regions as their own conservative reading rather than because Treasury has named them, so read the provider's wording alongside the Treasury sanctions programs page linked at the foot of this article. If you are in one of those places, no support ticket changes that, and anyone selling you a way in is selling you a problem.

The two categories also move independently, and Syria is the clean example. The comprehensive Syria program was revoked by Executive Order 14312, signed on 30 June 2025, and the Syrian Sanctions Regulations were removed from the Code of Federal Regulations in August 2025. More than a year later, Syria still does not appear on OpenAI's published list. Sanctions relief does not automatically restore a commercial launch, because a company still needs banking, payments and its own legal comfort first. The lesson runs both ways: your country may be absent for reasons that have nothing to do with sanctions.

#### Why the results above this one are a bad trade

Search this phrase and most of the first page is VPN affiliates and sellers of foreign phone numbers, because the payout per signup is high. We are not going to give you those instructions, and the reasons are practical rather than preachy.

- **You can lose the account, including the paid one.** The supported countries page states outright that access from outside the listed territories may result in suspension. Accounts built this way tend to fail at a payment or a re-verification, and appeals go nowhere because the terms were clear.
- **Rented phone numbers are not yours.** A number bought for a one time code can be reissued to someone else, and whoever holds it next can trigger a password reset on your account.
- **Payment gets ugly fast.** Card services promising no identity checks are a well documented fraud surface. Putting real money through one to reach a subscription that runs about 20 dollars a month as of August 2026 is a poor risk, and chargebacks on a mismatched billing country are their own headache.
- **In a sanctioned country the exposure is legal, not commercial.** Circumventing a comprehensive sanctions program is not a terms of service issue, and it is not a risk worth taking for a chatbot.

There is also a plain quality argument. An account that only works while a particular tunnel is up, on a number you rent monthly, is a fragile foundation for work you depend on.

#### The country lists genuinely differ between providers

These lists are not shared. Each company publishes its own, they are revised without announcement, and the differences are real. Checking the other two takes about a minute and is the step that most reliably helps.

| Provider | What the provider publishes | What we saw in August 2026 |
| --- | --- | --- |
| OpenAI (ChatGPT) | One published list, 186 entries when we counted it | China, Hong Kong, Macau, Russia, Belarus, Iran, Cuba, North Korea, Syria and Venezuela absent. Ukraine listed "with certain exceptions" |
| Anthropic (Claude) | Two separate lists, one for Claude.ai and one for the commercial API | Broad coverage across Africa, the Middle East and Latin America. China and Russia absent from both |
| Google (Gemini) | Over 70 languages and over 230 countries and territories | The widest published footprint of the three. Mainland China noted as Workspace accounts only |

Do not treat that table as current. Treat it as proof that the lists disagree, then read the live pages yourself, all three linked at the foot of this article. Three notes on reading them. The counts are not comparable, because Google enumerates small territories that other lists fold into a parent country. The OpenAI page we counted is the one on its developer documentation site, and it refers to "our services" without naming a product, so treat it as the list for the whole account rather than a ChatGPT specific document. And Anthropic publishes the chat product and the API as two separate lists. They held the same countries when we checked, but they are separate documents and can drift apart, so read the one that matches what you want to do.

For who else is in this market, [ChatGPT alternatives](https://whizi.io/resources/chatgpt-alternatives) covers the field by use case, and [free ChatGPT alternatives](https://whizi.io/resources/free-chatgpt-alternatives) is the version for when budget rather than geography is the constraint. If a different assistant turns out to be the one available to you, [the beginner guide to AI apps](https://whizi.io/resources/best-ai-apps-for-beginners) will get you productive on it without assuming you started on ChatGPT.

#### What to actually do next

Work through this in order. It takes a few minutes and it ends with a real answer rather than a workaround.

1. **Confirm which gate you hit.** Before the email field means the network check. At the code step means the phone number check. At checkout means the payment check.
2. **Read OpenAI's current list.** Countries get added. If yours is on it and access still fails, the block is very likely inside your country, and the answer is a different provider rather than a different setting.
3. **Check Anthropic and Google separately.** This is the step that most often resolves the problem, because the lists are different documents written by different companies.
4. **Establish whether sanctions are in play.** If you are in a comprehensively sanctioned country or region, most large US AI services will remain unavailable to you, and the honest move is to look at what is offered where you are.
5. **Look at what is available locally.** Some capable assistants are built outside the United States and publish their own country rules: the French company Mistral runs the assistant it renamed from Le Chat to Vibe in May 2026, covered in [the Le Chat alternative guide](https://whizi.io/alternatives/le-chat), and Alibaba's Qwen Chat and DeepSeek are both Chinese. Each has its own restrictions, so check the one you want rather than assuming a non US company will take you. Open models you run on your own hardware have no country check at all and cost nothing per message, but they need a machine with a capable GPU and a willingness to set it up.

One lesson applies regardless of geography. Building your week around a single provider is fragile, and a country block is only the most visible form of that. Outages, sudden free tier limits, a retired model, a price change: all of them land the same way if you have exactly one account. That is worth planning for whether or not your country is on anyone's list. And it is worth saying plainly about our own product, since you will see a signup button on this page: Whizi bills through its own payment processor with its own country coverage. It is not a route into a service that is legally unavailable where you are, and its checkout can decline a card issued in a country it does not serve. If that happens to you it is a location limit, not a problem with your card.

Key points:

- The message is a country check, not a bug on your side, and it can fire on your IP address, on the country of your phone number, or on both
- A country that is simply not on a provider list is a different situation from a country under a comprehensive sanctions program, and only the first one tends to change
- The country lists differ between OpenAI, Anthropic and Google, so a blocked signup on one service says almost nothing about the others

Checklist:

- Identify whether the block came from your network, your phone number, or your payment method.
- Read the provider's current supported countries page rather than a blog summary of it.
- Check Anthropic and Google separately, because their country lists are different documents.
- Find out whether your country is under a comprehensive sanctions program before spending time on this.
- Skip any service selling foreign phone numbers, resold accounts, or cards with no identity checks.
- Recheck the lists every few months, since countries are added without announcement.

FAQ:

Q: Why does ChatGPT say it is not available in my country?
A: OpenAI restricts access to a published list of countries and territories and refuses signup and login from anywhere else. The check runs on the country your IP address maps to, and separately on the country of the phone number you verify with. It is a policy decision on their side, not a fault on your device.

Q: Is my country unsupported or is it sanctioned? How do I tell?
A: Compare two published documents: the provider's own supported countries page, and the US Treasury sanctions programs page, both linked at the foot of this article. As of August 2026 the comprehensive United States programs cover Cuba, Iran, North Korea, and the Crimea, Donetsk and Luhansk regions of Ukraine. If your country is missing from a provider list but is not under a comprehensive program, you are most likely waiting on a commercial launch, which does happen.

Q: If ChatGPT is blocked, will Claude and Gemini be blocked too?
A: Not necessarily, and this is the step most people skip. The three companies publish separate lists that do not match. As of August 2026 Google states Gemini is available in over 230 countries and territories, which is a wider published footprint than the 186 entries on OpenAI's list. Checking all three takes a minute and often ends the problem.

Q: Should I use a VPN to sign up for ChatGPT?
A: We do not recommend it and we will not walk you through it. OpenAI's own supported countries page says access from outside the listed territories may result in your account being blocked or suspended, which puts anything you build on it at risk. Where a comprehensive sanctions program applies, the issue is legal rather than contractual and no workaround is appropriate.

Q: My country is on the supported list but ChatGPT still will not load. What now?
A: That usually points to a block applied inside your country by a network or a regulator rather than by the provider, and OpenAI cannot lift it from their end. One message to support is still worth sending in case the issue is with your account. Otherwise the practical answer is a service that is reachable where you are.

### ChatGPT Plus vs Claude Pro vs Gemini Advanced: which $20 plan

URL: https://whizi.io/resources/chatgpt-plus-vs-claude-pro-vs-gemini-advanced/
Updated: August 2026

Quick answer: Claude Pro is the pick for writing and code review, Google AI Pro for long documents and anyone working in Gmail and Docs, and ChatGPT Plus for breadth and integrations. All three sit around $20 a month and all three are good, so your ecosystem and writing preferences decide it rather than model quality.

Three AI subscriptions, all around $20 a month. A practical comparison of what each one actually includes, who each is built for, and how to pick just one.

#### The honest starting point

These three plans all sit at roughly $20 a month, all give you a frontier model, and all are good enough that you would get useful work done with any of them. Anyone who tells you one is objectively best across the board is either selling something or has only used one of them properly.

A note on names before we start. Gemini Advanced was folded into **Google AI Pro**, so if you are searching for Gemini Advanced you are looking for that plan. Claude also sells much heavier **Max** tiers, and ChatGPT has a **Pro** tier around ten times the price. This comparison is about the $20 rung, since that is where almost every real decision happens.

The tiebreakers that actually decide it are not benchmark scores. They are, in rough order of how much they matter day to day:

1. **Which ecosystem you already live in.** Gmail and Docs all day, or Microsoft, or neither.
2. **What is bundled.** One of these three includes 2TB of cloud storage, which is not a small detail.
3. **Whose writing voice needs less editing.** Completely subjective and completely decisive if you write for a living.
4. **Whether you code, and where.** In a terminal, in an editor, or in a chat window.

Verify current pricing on each vendor's site before you buy. Plan names and inclusions in this category change several times a year, and the bundled extras change more often than the prices do.

#### What each plan actually includes

Feature lists blur together, so this table focuses on what is different rather than what all three share. Every one of them gives you a frontier model, file uploads, image understanding, web access, and a mobile app.

| | ChatGPT Plus | Claude Pro | Google AI Pro |
| --- | --- | --- | --- |
| Around | $20 / month | $20 / month, cheaper annually | $20 / month |
| Best known for | Breadth and ecosystem | Writing quality and code | Long context and Google apps |
| Signature feature | Custom GPTs, code execution, connectors | Artifacts, Projects, Claude Code | Gemini inside Gmail, Docs, Drive, Sheets |
| Context window | Large | Large, with much larger tiers available | The largest of the three by a wide margin |
| Bundled extras | Image and video generation allowances | Coding agent access with limits | 2TB cloud storage, video generation, NotebookLM limits |
| Weakest at | Sourced research with citations | Ecosystem breadth and integrations | Consistency of tone in long writing |
| Free tier | Good | Good | Good |

Look at the bundled extras row again. If you already pay Google for 2TB of storage, the AI plan effectively costs you the difference rather than the full price. That single fact settles the decision for a lot of people before any model comparison starts, and almost no comparison article mentions it.

#### Where each one genuinely wins

### ChatGPT Plus wins on breadth

It is the tool that does the most different things acceptably well. Code execution on your uploaded data, image generation, voice, connectors into workplace tools, custom GPTs for repeatable jobs, and by far the largest body of community knowledge when something goes wrong. If you want one assistant that will handle whatever lands on your desk, this is the safe pick, and "safe" is a real virtue when the alternative is being stuck.

It is also the best pick for anyone building workflows or automations, because the API and the tooling around it are the most mature and the most documented.

### Claude Pro wins on writing and reading

Two things stand out consistently. First, the default prose needs less editing. It holds a register, it hedges less mechanically, and it produces first drafts that read like a competent human wrote them rather than like a template got filled in. If you write for other people to read, test this before you commit to anything else.

Second, it is very good at reading. Long documents, dense contracts, code review, sprawling research notes: Claude handles synthesis carefully and is more willing to say what a document does not say. For developers, Claude Code brings that same quality into the terminal, and it is included with the plan at usage limits that suit a solo developer.

The catch is breadth. Fewer integrations, a smaller ecosystem, and tighter usage limits than people expect from a $20 plan if you push it hard.

### Google AI Pro wins on scale and on Google

The context window is the headline. Gemini can take in far more material at once than the other two, which changes what is possible: a whole codebase, a year of meeting notes, a 900-page manual. Long-context work is where it feels like a different class of tool rather than a slightly different one. Our [context window explainer](https://whizi.io/resources/what-is-a-context-window) covers why bigger is not automatically better, but for the "read all of this and find the pattern" job, it is.

Then there is the integration. Gemini in the side panel of Gmail, Docs, Sheets, and Drive means the assistant is where the work already is, with access to your files without uploading anything. Add 2TB of storage, video generation, and higher NotebookLM limits, and the value per dollar is the strongest of the three for anyone in the Google ecosystem.

The weakness is consistency. Output quality varies more between tasks than the other two, and long-form writing needs more editing.

#### Where each one genuinely loses

This is the part that matters more than the strengths, because you will feel the weakness of whichever plan you pick every single week.

| Plan | The thing that will annoy you |
| --- | --- |
| ChatGPT Plus | Writing that needs a pass to remove its default rhythm, and research answers that summarise rather than cite |
| Claude Pro | Hitting usage limits sooner than you expect on heavy days, and a thinner ecosystem when you want an integration |
| Google AI Pro | Uneven output between tasks, and a product surface that changes names and layouts often |

None of these are dealbreakers. They are the friction you are choosing. It is easier to accept friction you picked deliberately than friction you discover in month three.

#### Pick one in sixty seconds

Answer the first question that applies to you and stop reading.

- **Do you live in Gmail and Google Docs all day?** Google AI Pro. The integration and the storage bundle make it the best value, and you will use it more because it is already there.

- **Do you write things other people read, professionally?** Claude Pro. The editing time you save is worth more than any feature list difference.

- **Do you build software, mostly in a terminal or reviewing code?** Claude Pro, for Claude Code and code review quality. If you work primarily inside an IDE, note that a dedicated in-editor assistant may serve you better than any of these three, and read our [AI for coding guide](https://whizi.io/resources/ai-for-coding).

- **Do you need one tool that can attempt anything, including data analysis, images, and automations?** ChatGPT Plus.

- **Do you work with very long documents, video, or huge codebases?** Google AI Pro, for the context window.

- **Is most of your work research where every claim needs a source?** None of these three are ideal. Look at a citation-first tool and read [Perplexity vs ChatGPT](https://whizi.io/resources/perplexity-vs-chatgpt) first.

- **Do you genuinely need two of them?** Read the next section before you pay for two.

#### The two-subscription problem

Here is what actually happens to people who take this decision seriously. They pick one, use it for a few months, and hit its weakness. Then they add a second plan to cover the gap. Now they are at $40 a month, with their history split across two products, pasting prompts between tabs to compare answers, and never quite sure which tool to open first.

The instinct is right. Different models really are better at different jobs, and the ranking changes with every release, so committing to one vendor means accepting whatever they ship next. The instinct is also expensive, and it gets worse each time a new model makes headlines.

The alternative worth knowing about: a multi-model workspace gives you these models behind one subscription. [Whizi](https://whizi.io/pricing) includes GPT, Claude, Gemini, Grok, and DeepSeek in one place, so you can start a draft in Claude, ask GPT to poke holes in it, and hand the 300-page appendix to Gemini, in one conversation and on one bill. [How to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers the workflow, and [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus), [Whizi vs Claude Pro](https://whizi.io/vs/whizi-vs-claude-pro), and [Whizi vs Gemini Advanced](https://whizi.io/vs/whizi-vs-gemini-advanced) give the direct comparisons.

The honest counterpoint: a single vendor plan gives you that vendor's deepest product features, and some of them are excellent and not replicable elsewhere. Claude Code, Gemini inside Google Docs, and ChatGPT's connectors are product features, not just model access. If one of those is central to your day, buy that plan and do not overthink it.

The decision rule is simple. **If one product's features are the reason you want it, buy that product. If model access is the reason, buy model access.** Most people who own two subscriptions are in the second group and bought the first thing.

#### Test all three before you commit

Every one of these has a free tier that is good enough to judge the model, even if the limits are tighter. Spend one afternoon rather than three months of subscription fees.

Take four prompts from real work: a piece of writing you will actually send, a document you need summarised accurately, a problem where you already know the correct answer, and whatever your most repeated task is. Run all four through all three free tiers, on the same day, with identical prompts.

Score only three things per answer: how many minutes to a usable result, how many edits it needed, and whether you would have caught a mistake if you had not been checking. That last one is the most revealing and the one people never measure.

Then total your existing AI spend before adding anything with the [savings calculator](https://whizi.io/tools/calculator), and pick a plan from the [pricing page](https://whizi.io/pricing) once the numbers and the test agree with each other.

Key points:

- All three cost about the same and all three are good, so the tiebreaker is rarely raw model quality
- Google's plan bundles 2TB of storage and Workspace integration, which changes the real price
- Claude wins on writing and code review, Gemini on long documents and Google apps, ChatGPT on breadth

Checklist:

- Check whether you already pay for Google storage, which changes the real price of one plan
- Decide whether you are buying product features or model access, since that settles most of it
- Run the same four real prompts through all three free tiers on the same day
- Score answers on editing time, not on which one sounded most impressive
- Note which weakness you can live with, since you will meet it every week
- Total your current AI subscriptions before adding a second or third plan

FAQ:

Q: Which is best: ChatGPT Plus, Claude Pro, or Gemini Advanced?
A: Claude Pro for writing and code review, Google AI Pro for long documents and anyone working in Gmail and Docs, ChatGPT Plus for breadth and integrations. All three are around $20 a month and all three are good, so the deciding factors are usually your ecosystem and your writing preferences rather than model quality.

Q: Is Gemini Advanced still called that?
A: It was folded into Google AI Pro, which includes Gemini's paid features plus 2TB of storage, Workspace integration, video generation, and higher NotebookLM limits. If you are searching for Gemini Advanced, that is the plan you want.

Q: Which one has the best free tier?
A: All three free tiers are usable for casual work, with daily caps on the flagship models. Google's is often the most generous for long-context tasks and ChatGPT's has the widest feature surface. Test all three before paying, because a free tier may be enough for you.

Q: Should I subscribe to more than one?
A: Only if you need a specific product feature from each, such as Claude Code and Gemini in Google Docs. If what you actually want is access to different models for different tasks, a multi-model workspace costs less than two subscriptions and keeps your history in one place.

Q: Which is best for coding?
A: Claude Pro is the strongest of the three for code review, refactoring, and terminal work through Claude Code. Google AI Pro has an edge when you need a model to read an entire large codebase at once. For work inside an IDE all day, a dedicated in-editor assistant often beats all three.

### ChatGPT Projects alternative: the same workflow, every model, free tier included

URL: https://whizi.io/resources/chatgpt-projects-alternative/
Updated: August 2026

Quick answer: A ChatGPT Projects alternative needs three things: folders that group related chats, files the AI can see in every conversation, and instructions that apply to the whole folder. Whizi projects do all three on every plan including free, with one difference that matters: the chats inside a project can use any major model, not just one vendor.

Looking for a ChatGPT Projects alternative? What Projects actually does, where its limits bite, and how Whizi projects give you folders, pinned files, and instructions with every major model.

#### What people actually want from Projects

Strip away the branding and the Projects workflow is three features working together. A **folder** that groups the conversations about one piece of work. **Files** that the AI can see in every conversation in that folder, so you stop re-uploading the same brief. And **instructions** that apply to the whole folder, so every chat starts already knowing the client, the tone, or the format you need.

Once you have worked that way, going back to loose chats feels broken. The same document pasted into the fifth conversation, the same three paragraphs of context retyped, the same instructions repeated. That is the workflow people are trying to keep when they search for an alternative.

The reasons people look elsewhere are usually one of three: the feature is gated or limited on their plan, they want the workflow with a different model than the one vendor offers, or they are consolidating tools and the standalone subscription is the thing being cancelled.

#### Where ChatGPT Projects is strong, and where it stops

Credit where due: ChatGPT Projects is a well-built version of this workflow, with project files, project-level instructions, and project-scoped memory, inside the most polished mainstream AI app. If you live entirely inside ChatGPT and pay for it, Projects is a good reason to stay.

The limits worth knowing before you commit your working structure to it. Projects began as a paid-tier feature, and free access arrived later with tighter caps that have shifted over time, so check what your plan includes today. File and project counts scale with the tier you pay for.

The structural limit does not change with any tier: **everything inside a ChatGPT project is answered by OpenAI models**. If the writing project would be better served by Claude, or the research folder wants a different model for verification, the folder cannot follow you. The same is true in reverse for Claude Projects, which is a paid-tier feature on Anthropic's side. Your organized workspace is locked to one vendor's strengths and one vendor's limits.

#### Whizi projects: the same three features, every model

Whizi projects implement the same workflow: a project is a folder of conversations that also carries pinned files and its own instructions, and on every message in the project the instructions and the pinned file text are folded into the model prompt.

The caps, published in the [product guide](https://whizi.io/docs/use-projects):

| Limit | Value |
| --- | --- |
| Projects per account | 50 |
| Project instructions | 32,000 characters |
| Pinned files per project | 10 |
| Size of one pinned file | 10 MB |
| Accepted file types | PDF, text, Markdown, CSV, JSON, .docx, .xlsx |
| Available on | Every plan, including free |

The difference that motivates this page: a conversation inside a Whizi project can run on any model in the picker. The client folder can draft in Claude, fact-check in GPT, and summarize a spreadsheet in Gemini, and every one of those chats sees the same pinned brief and follows the same project instructions. You organize once and the structure serves whichever model is best for the task in front of you.

Two honest caveats so this stays a fair comparison. Pinned files are text-first: images cannot be pinned, because pinned content rides along as extracted text on every turn. And project context is rebuilt on every message, which is what makes it reliable and also what makes very large pinned files expensive in context. The [full guide](https://whizi.io/docs/use-projects) explains both behaviors.

#### Moving a working setup across

Migrating the Projects workflow is lighter than it sounds, because the valuable part is not the chats, it is the structure: which files matter, and what the standing instructions say.

1. Create the project in Whizi and paste the same instructions you maintained in ChatGPT. Up to 32,000 characters is more room than most standing briefs need.
2. Pin the working documents: the brief, the style guide, the reference data. Ten files of 10 MB each covers a serious working folder.
3. Start the next conversation in the project rather than importing the old ones, and keep the old thread open in a tab for the first week as reference.
4. If you do want the history, [Whizi imports ChatGPT and Claude exports](https://whizi.io/docs/import-chats-from-chatgpt), and imported chats can be moved into projects like any other chat.

From there the setup habits are the same ones that made Projects useful in the first place: one project per client or deliverable, instructions that describe the reader rather than the task, and pinned files kept current instead of accumulating stale drafts.

#### Which alternative fits which situation

The honest matrix, since not everyone should switch:

- **You pay for ChatGPT, use only OpenAI models, and Projects works for you.** Stay. The feature is good and switching costs a week of habit change.
- **You want the Projects workflow without paying for a single-vendor plan.** Whizi projects are on the free tier, so the workflow itself costs nothing to try.
- **You want the folder, files, and instructions to work across models.** This is the case Whizi is built for, and no single-vendor Projects feature can follow you there. See the [ChatGPT alternative overview](https://whizi.io/alternatives/chatgpt) for the wider comparison.
- **You are consolidating two AI subscriptions into one.** Run the numbers in the [savings calculator](https://whizi.io/tools/calculator), then read [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) for the direct head-to-head, including where ChatGPT still wins.

Whichever way you go, the three-feature test is the useful filter: folders, files the AI sees everywhere, instructions that stick. Any tool that does all three preserves the workflow. Anything less is a chat app with a sidebar.

Key points:

- The Projects workflow is three features in one: chat folders, pinned files, and per-project instructions
- The real limitation of ChatGPT Projects is not the caps, it is that everything inside is answered by one vendor
- Whizi projects are on every plan including free, with up to 10 pinned files and 32,000 characters of instructions per project

Checklist:

- List the projects you actually maintain and what their standing instructions say
- Check what your current plan gates: file counts, project counts, and instruction length
- Test the same project structure on a second tool before committing to a switch
- Keep pinned files current; a stale brief applied to every chat is worse than none
- Judge the alternative by the three-feature test: folders, ever-present files, sticky instructions

FAQ:

Q: Is there a free alternative to ChatGPT Projects?
A: Yes. Whizi projects are available on every plan including the free tier: folders of chats with up to 10 pinned files and 32,000 characters of per-project instructions. The caps are abuse guards rather than a paid gate.

Q: Can project chats use different AI models?
A: In single-vendor tools, no: a ChatGPT project answers with OpenAI models and a Claude project with Anthropic models. In Whizi, each conversation inside a project can run on any model in the picker while seeing the same pinned files and instructions.

Q: Do pinned files work like chat attachments?
A: No, and the difference is the point. A chat attachment is seen once in that conversation. A pinned file rides along with every message in the project as extracted text, so every chat in the folder can rely on it without re-uploading.

Q: Can I move my ChatGPT project chats to an alternative?
A: The structure moves in minutes: recreate the instructions and re-pin the files. For the conversations themselves, Whizi imports ChatGPT and Claude export files, and imported chats can be moved into projects afterwards.

Q: Why can I not pin images to a Whizi project?
A: Pinned files contribute extracted text to every turn in the project, and an image has no text to extract. Images still work as normal attachments in a single conversation with a vision-capable model.

### ChatGPT vs Claude vs Gemini: which AI model is best in 2026?

URL: https://whizi.io/resources/chatgpt-vs-claude-vs-gemini/
Updated: August 2026

Quick answer: No single model wins outright. ChatGPT is often the strongest default for structured drafts, implementation plans, and step-by-step debugging. Claude is often the better fit for nuanced writing, code review, and careful synthesis. Gemini is worth testing for long-context document work and multimodal input. Route by task, then test all three on one real job.

Compare ChatGPT, Claude, and Gemini by task fit, model capabilities, workflow needs, and cost so you can choose the right AI model for real work.

#### Models vs apps vs APIs

The first trap in the ChatGPT vs Claude vs Gemini debate is comparing the wrong layer. ChatGPT, Claude, and Gemini are often used as shorthand for three different things: the underlying model family, the consumer app, and the developer API. Those layers overlap, but they are not the same decision.

A model is the reasoning and generation engine. An app is the interface around that model, including chat history, file upload, voice, connectors, memory, projects, and collaboration features. An API is the programmable version a developer can connect to a product or internal workflow. When someone asks "which AI model should I use?", the practical answer depends on whether they are drafting a sales email, reviewing a pull request, analyzing a 60-page PDF, or building an automated workflow.

That is why this guide frames ChatGPT vs Claude vs Gemini as a task-fit comparison. OpenAI, Anthropic, and Google all publish model documentation that changes over time, and each vendor offers multiple models optimized for different tradeoffs. The better question is not "which brand wins?" It is "which model gives me the best answer for this job, under my constraints, at a cost I can justify?"

If you are still building your shortlist, start with the broader guide to [ChatGPT alternatives](https://whizi.io/resources/chatgpt-alternatives). If you already know you want to compare the big three, keep reading and use the test protocol below before you standardize your workflow.

#### Capabilities comparison: vision, tools, long context

Capability claims age quickly, so treat this section as a practical checklist rather than a permanent scoreboard. Official OpenAI, Anthropic, and Gemini docs describe current model families, context windows, tool use patterns, and multimodal capabilities, but the right choice still comes down to your input, output, and review process.

| Capability | ChatGPT / OpenAI models | Claude / Anthropic models | Gemini / Google models | What to test |
|---|---|---|---|---|
| Text generation | Strong for structured drafts, plans, explanations, and general-purpose work | Strong for nuanced writing, editing, long-form synthesis, and careful tone | Strong for broad productivity, document work, and Google ecosystem-adjacent workflows | Ask each model to rewrite the same messy draft with the same audience and constraints |
| Coding help | Useful for explanations, implementation plans, debugging, tests, and code generation | Useful for code review, refactors, reasoning through large diffs, and careful explanations | Useful for code tasks, especially when paired with long context and structured prompts | Give each model the same bug report, failing test, and file excerpt |
| Vision and multimodal input | Official model docs include models with vision and multimodal capabilities | Claude supports multimodal and tool-oriented workflows depending on model and product surface | Gemini documentation emphasizes multimodal models and long-context workflows | Test one image or PDF extraction task with a required output schema |
| Tool use | OpenAI models can be used in app and API workflows with tool integrations depending on surface | Anthropic documents tool use patterns for connecting Claude to external functions and tools | Gemini API supports tool and structured workflow patterns depending on model | Ask for a tool plan first, then compare whether the model calls for the right data |
| Long context | Available in supported models, with limits varying by model | Available in supported Claude models, with model-specific context limits | Gemini documentation specifically highlights long-context use cases and guidance | Upload or paste a long source and ask for cited extraction, not a vague summary |

The main lesson: do not choose a model from a feature checklist alone. A model can support a capability and still be the wrong fit for your workflow if it produces answers that are too verbose, too brittle, too slow, too expensive, or too hard for your team to verify.

#### Best by use case

Most teams eventually learn that the best AI model 2026 decision is not one permanent decision. It is a routing system. Use a default model for everyday work, then switch when the task has a special constraint: sensitive tone, large documents, code risk, visual input, research traceability, or structured extraction. The routing most teams settle on looks like this.

| Model | Where it usually wins | Where to test a challenger |
| --- | --- | --- |
| ChatGPT | Outlines, frameworks, implementation plans, and step-by-step debugging from a clear reproduction | Long documents and voice-sensitive final prose |
| Claude | Long-form editing, tone-sensitive rewriting, code review, refactor plans, and research synthesis | Multimodal input and very large source packs |
| Gemini | Long-context document work, large source material, and mixed image, log, and text input | Short structured drafts where fast iteration matters more |

#### Writing

For writing, compare models on audience fit, specificity, editability, and how well they preserve facts you provide. Claude is often a strong candidate for long-form editing, voice-sensitive rewriting, and polished narrative structure. ChatGPT is often strong for outlines, frameworks, campaign ideas, and turning messy notes into organized drafts. Gemini can be useful when the writing task is connected to large source material, document review, or multimodal context. For a two-model breakdown of the Google and OpenAI options, see [Gemini vs ChatGPT](https://whizi.io/resources/gemini-vs-chatgpt).

Use this writing prompt to compare outputs: "Rewrite the draft below for a skeptical operations leader. Keep the claims factual, remove generic AI phrasing, preserve the concrete examples, and return: 1) final draft, 2) three edits you made, 3) two claims I should verify before publishing."

The winner is not the prettiest paragraph. The winner is the output you can publish with the least cleanup while still trusting the factual spine.

#### Coding

For coding, the best model is the one that helps you reduce risk, not just the one that writes the most code. Compare ChatGPT, Claude, and Gemini on how they reason from a reproduction, whether they ask for missing context, whether they propose small changes, and whether they include tests.

Use this coding prompt: "You are reviewing a bug fix. First restate the likely root cause from the reproduction. Then propose the smallest safe change. Then list tests that would fail before the fix and pass after. Do not rewrite unrelated code. Here is the bug report, relevant code, and test output."

ChatGPT can be excellent for implementation planning and explaining unfamiliar code. Claude can be especially useful when you want a careful review of tradeoffs or a large-context refactor plan. Gemini is worth testing when the code task includes long files, screenshots, logs, or broader document context. In every case, require tests and human review.

#### Research

For research, do not reward confident prose. Reward traceability. The best AI model for research is the one that separates source collection, extraction, synthesis, and uncertainty. If the model cannot show what came from the source versus what it inferred, the output is not ready for decision-making.

Use this research prompt: "Answer the question using only the sources I provide. Create a table with claim, source, confidence, and notes. Then write a synthesis in 250 words. End with open questions and what evidence would change the conclusion."

ChatGPT, Claude, and Gemini can all support research workflows when used carefully, but they should be evaluated on citation discipline, quote handling, and whether they flag missing evidence. For web-connected or document-heavy research, test the same source pack across models instead of assuming one brand is always better.

#### Docs

Document work is where long context, file handling, and structured extraction matter. Gemini is a serious contender when long-context document workflows are central to the job, especially because Google publishes specific guidance around long context in the Gemini API. Claude can be strong for preserving nuance across long documents and turning dense material into readable deliverables. ChatGPT can be strong for building structured summaries, action plans, and reusable templates from docs.

A good document prompt should not ask for "a summary" and stop. Ask for an outline, key entities, decisions, risks, contradictions, and page-specific items if your workflow supports page references. Then ask the model to mark anything it is unsure about.

#### Decision table

Use this table as a starting point, then run the A/B/C protocol below. It is intentionally framed as task fit, not a universal ranking.

| If your task is... | Start with... | Also test... | Why |
|---|---|---|---|
| First draft of a plan, outline, or structured answer | ChatGPT | Claude | Strong general-purpose structure and fast iteration |
| Polishing a long article, memo, or client deliverable | Claude | ChatGPT | Often a strong fit for tone, nuance, and editing workflows |
| Large document analysis or long-context extraction | Gemini | Claude | Gemini long-context guidance makes it worth testing for big inputs |
| Code review or safer refactor planning | Claude | ChatGPT | Careful reasoning and review-style outputs can help reduce risk |
| Debugging with logs and tests | ChatGPT | Claude | Strong step-by-step reasoning when the reproduction is clear |
| Multimodal task with images, docs, and text | Gemini | ChatGPT | Worth testing when inputs span formats |
| Research synthesis from a provided source pack | Claude | Gemini | Evaluate source discipline, uncertainty, and synthesis quality |
| Everyday mixed work | Whizi model comparison | All three | The fastest way to learn your own routing rules |

The table should not replace judgment. It should help you decide which model gets the first attempt and which model gets the challenger slot.

#### Reusable A/B/C prompt test protocol

The cleanest way to answer "ChatGPT vs Claude vs Gemini" is to stop debating and run the same prompt across all three. Use this protocol for any important workflow before you choose a default model.

1. Pick one real task. Do not use a toy prompt. Choose a task you actually need to complete this week: a sales email, a code review, a market research synthesis, a PDF extraction, or a support response.

2. Create a fixed input pack. Include the same source text, constraints, audience, desired format, and quality bar for every model. If one model gets more context than the others, the test is not fair.

3. Use a scoring rubric before you read the answers. Score each output from 1 to 5 on accuracy, usefulness, format compliance, edit time, risk, and confidence calibration.

4. Run the same prompt in each model without changing wording. If the first prompt is flawed, revise it once and rerun it everywhere.

5. Do a second-round challenge. Ask each model: "What could be wrong with this answer? What assumptions did you make? What should I verify?"

6. Choose a workflow rule. For example: "Use Claude for final prose, ChatGPT for implementation plans, Gemini for long docs, and Whizi when the task matters enough to compare."

Copy-paste test prompt: "I am comparing AI models for this workflow. Complete the task below using only the context provided. Follow the output format exactly. After the answer, include: assumptions, risks, verification checklist, and one suggestion to improve the prompt. Task: [paste task]. Context: [paste context]. Output format: [paste format]."

#### Cost: one subscription vs multiple

The model decision is also a subscription decision. Many people start with one paid AI subscription, then add another for writing, another for research, another for image or document work, and another for team experiments. The monthly cost becomes hard to justify because each tool is only clearly best for part of the workflow.

That is the consolidation argument for Whizi: you do not need to pretend there is one permanent winner. You can compare model outputs in one workspace, keep the model that fits the job, and avoid bouncing between separate subscriptions when you only need access for specific tasks.

If you are already paying for more than one AI tool, run the [AI subscription savings calculator](https://whizi.io/tools/calculator). Then compare the result against [Whizi pricing](https://whizi.io/pricing). If the two you keep opening are GPT and Claude, [the cheapest way to use ChatGPT and Claude together](https://whizi.io/resources/cheapest-way-to-use-claude-and-gpt) prices four routes side by side. The goal is not just cheaper access. The goal is a cleaner workflow: fewer tabs, fewer logins, and faster decisions about which model should handle which task.

#### Try the same prompt across models

The most useful conclusion is simple: ChatGPT vs Claude vs Gemini is not a one-model cage match. ChatGPT may be your fastest structured thinker for one task. Claude may be your best editor for another. Gemini may be the better test for long-context or multimodal document work. The smart move is to build a routing habit instead of a brand habit.

In Whizi, you can run the same prompt across models, compare outputs side by side, and turn the winning pattern into a reusable workflow. Start with one important task, use the A/B/C test above, and save the model choice that actually performs best for your work.

When you are ready, [try comparisons in Whizi](https://whizi.io/register) and use the results to choose your own best AI model for writing, coding, research, docs, and everyday work.

Key points:

- Use ChatGPT, Claude, and Gemini as task-fit options instead of treating one model as universally best.
- Compare vision, tools, long context, writing, coding, research, document work, and subscription cost.
- Run a reusable A/B/C prompt test in Whizi before you commit to a model or workflow.

Checklist:

- Decide whether you are comparing models, apps, or APIs before choosing a tool.
- Use the same task, context, and output format when testing ChatGPT, Claude, and Gemini.
- Score outputs on accuracy, usefulness, format compliance, edit time, risk, and verification quality.
- Use ChatGPT, Claude, and Gemini as a routing system instead of forcing one universal winner.
- Run the savings calculator if you are paying for multiple AI subscriptions.

FAQ:

Q: Which is better: ChatGPT, Claude, or Gemini?
A: There is no universal winner. ChatGPT is often a strong default for structured general work, Claude is often a strong fit for nuanced writing and careful review, and Gemini is worth testing for long-context and multimodal workflows. The best choice depends on the task and your review process.

Q: What is the best AI model for writing?
A: For writing, test models on voice, factual preservation, structure, and edit time. Claude is often a strong candidate for polished prose, ChatGPT is often strong for outlines and structured drafts, and Gemini can be useful when writing from large source material.

Q: What is the best AI model for coding?
A: For coding, choose the model that best works from a reproduction, explains risk, proposes small changes, and suggests tests. ChatGPT and Claude are both worth testing for debugging, refactors, reviews, and unit test planning.

Q: Should I pay for multiple AI subscriptions?
A: Only if the extra cost clearly improves your workflow. Many users are better served by comparing models in one workspace, then using a consolidated plan instead of stacking separate subscriptions.

### The cheapest way to use ChatGPT and Claude together

URL: https://whizi.io/resources/cheapest-way-to-use-claude-and-gpt/
Updated: August 2026

Quick answer: The cheapest route depends on volume. For occasional use, the free tiers of both cost nothing. For light to medium daily use, pay-as-you-go API access through a chat client often runs a few dollars a month. For heavy steady use, a flat plan wins: ChatGPT Go plus Claude Pro at about $28, or one multi-model subscription.

Four real routes to running Claude and GPT side by side, priced for a light, medium, and heavy user, with the account sharing risk named plainly.

#### The short answer

There are four real ways to use Claude and GPT together, and the cheapest one depends on how much you actually use them, not on which route sounds cheapest in the abstract.

- **Two direct subscriptions**: about $28 a month for ChatGPT Go plus Claude Pro, or about $40 for ChatGPT Plus plus Claude Pro, flat, regardless of usage.
- **Free tiers of both, used side by side**: $0 a month, with daily or weekly caps that reset and models that step down during busy periods.
- **Pay-as-you-go API access through a chat client**: no monthly fee, billed per token, so the cost is pennies for light use and can run past $150 a month for heavy use.
- **A multi-model subscription** that includes both in one plan: one monthly fee, $15.99 on Whizi's entry plan ($10.99 a month if billed annually), against $20 for either provider's own standard plan.

A fifth thing shows up on page one of the search results for this exact question, and it is not a real option: paying a stranger for a shared or resold ChatGPT Plus or Claude Pro login. That breaks both providers' terms of service, so it is covered plainly further down rather than counted as a fifth real route.

#### Route 1: two direct subscriptions

The obvious route is also the easiest to price, and most articles on this question price it wrong by assuming ChatGPT Plus. There are two paid ChatGPT tiers. As of August 2026, ChatGPT Go is $8 a month in the US and ChatGPT Plus is $20 a month, while Claude Pro is $20 a month, or $17 a month if you pay annually (Anthropic bills that as $200 up front). So the cheapest pairing of two direct subscriptions is Go plus Claude Pro at about $28 a month, or about $25 if you pay Claude annually. ChatGPT Plus with Claude Pro is $40 a month, or $37 with Claude annual, since OpenAI does not currently list a discounted annual rate for Plus.

Go is a capped tier rather than a smaller Plus: more uploads and images than the free plan, and everyday text chat that OpenAI's pricing page lists as unlimited on both, but not the deep research and agent style work Plus is bought for, and it can carry ads in a growing list of markets. Price the $28 pairing first and move up to $40 only if you can name the Plus feature you need. Two more caveats on those figures. OpenAI took Go worldwide in January 2026 and prices it locally outside the US, and Claude subscriptions are localized in many markets too, so the number at checkout may not be the dollar figure here. And both services are country gated: [ChatGPT supported countries](https://help.openai.com/en/articles/7947663-chatgpt-supported-countries) and [Claude supported countries](https://www.anthropic.com/supported-countries) are the lists that decide whether you can subscribe at all. If your country is not on them, that is a location limit rather than a payment problem, and no card fixes it.

What the money buys you is not usage. It buys you access, at whatever volume the provider considers reasonable for one person, inside two completely separate products. You get two logins, two histories, two apps, and no way to ask GPT to critique something Claude just wrote without copying the text yourself. If you already juggle more tools than these two, [ai subscription costs](https://whizi.io/resources/ai-subscription-costs) has a worksheet for finding out what your whole stack actually adds up to.

This route is the right one when your usage is heavy and steady, because the price does not move whether you send fifty messages a month or five thousand. It is the wrong one if your actual usage is light, because you are paying a flat fee for capacity you are not using, and it stops being obviously right the moment a cheaper metered option can cover the same volume for less. The next two sections cover exactly that case.

#### Route 2: free tiers of both, used together

Both companies give away a genuinely usable free tier. ChatGPT's free plan and Claude's free plan both let you send messages to a capable model at no cost, and running them side by side costs nothing.

The catch is the same on both sides, and it is a mechanism rather than a fixed number: free tiers cap how much you can send in a given window, typically resetting every few hours or every day, and during busy periods some free plans quietly route you to a smaller, faster model instead of the flagship one. Those caps and reset windows move often enough that a specific figure printed here would be wrong within a few months, so check the current limits on each provider's own help page before you plan around them: [OpenAI Help Center](https://help.openai.com/) and [Anthropic Help Center](https://support.anthropic.com/) both keep this current.

This route is genuinely the cheapest one that exists, and it is the right answer for a real slice of readers: anyone whose actual usage is a handful of questions a week, comparing an occasional answer between the two, well under whatever the current caps are. It stops being the right answer the moment you hit the reset wall mid-task, because a free tier gives you no way to buy more room for that one busy week without upgrading the whole month.

#### Route 3: pay-as-you-go API access

The route almost nobody tries, because it sounds technical and mostly is not. Both OpenAI and Anthropic sell direct API access billed per token, with no monthly subscription fee. You need a client to talk to it, since the raw API has no chat window of its own: OpenAI's own developer platform and Anthropic's own developer console both work as a starting point, and several chat interfaces let you paste in your own API key and talk to both providers from one screen without a subscription to either. The usual names are TypingMind, a paid web app you buy a licence for, LibreChat, a free open source app you host yourself, and Chatbox, a desktop and mobile client with a free tier. All three accept an OpenAI key and an Anthropic key side by side, so the choice is mostly about whether you would rather pay once, self host, or install an app.

Anthropic's published API pricing as of August 2026: the small, fast Claude model runs $1 per million input tokens and $5 per million output tokens; the mid-tier Claude model most people reach for runs $2 in and $10 out per million tokens; the most expensive Claude model most people would consider runs $5 in and $25 out, and Anthropic lists specialist models above that. OpenAI's published API pricing over the same period: its smallest model runs $0.05 in and $0.40 out per million tokens; a smaller general-purpose model runs $0.25 in and $2 out; its flagship model runs $5 in and $30 out. A million tokens is roughly 750,000 words, so most single conversations use a tiny fraction of one.

Three numbers, worked through for a light, medium, and heavy user, using those published rates and the mid-tier model on each side unless noted. Treat these as an illustration of the mechanism rather than a promise about your own bill, since the real number depends on which model you pick and how long your conversations run.

| User | Assumption | GPT side | Claude side | Total |
| --- | --- | --- | --- | --- |
| Light | About 50,000 tokens a month total, a few short exchanges a week | Under $0.03 | Under $0.20 | Under $0.25 a month |
| Medium | About 2,000,000 tokens a month in total, daily use with some pasted context | About $1.10 | About $6.00 | Roughly $7 a month |
| Heavy | About 20,000,000 tokens a month in total on the flagship model, split evenly between the two, with long documents and extended sessions | About $175 | About $150 | Roughly $325 a month |

One mechanism matters more than any single number here: a back-and-forth conversation resends the whole thread on every turn unless your client caches it, so a long conversation burns through tokens far faster than the same total words spread across short, independent questions. Two things drive the heavy row above, then, and it is worth keeping them apart: the token count itself, and the flagship rate, which is several times the mid-tier one. That is why anyone doing sustained daily work should watch the running total rather than trust a monthly estimate made in advance. Most API clients show you the running cost per conversation, so check it against the number here instead of taking it on faith.

This route wins clearly for light and medium use, where a $7 month beats even the $28 Go pairing and, for the lightest users, comes close to the free tier while removing the reset-window ceiling. It stops winning once usage gets heavy enough that a flat subscription is cheaper than the metered total. On these rates that crossover sits at roughly 8,000,000 mid-tier tokens a month against the $28 pairing, and roughly 11,000,000 against the $40 one, both well inside the gap between the medium and heavy rows above.

#### Route 4: a multi-model subscription

The fourth route is a single subscription that includes both GPT and Claude models in one plan, which is what [Whizi](https://whizi.io/register) is. It exists for the same reason [one subscription for all AI models](https://whizi.io/resources/one-subscription-all-ai-models) explains at more length: providers price by seat, not by task, and nobody bundles a competitor's model into their own app, so a third product has to do the bundling.

Whizi's Starter plan is $15.99 a month, or $10.99 a month billed annually, and includes GPT alongside a Claude model in the same workspace. That is above the $8 of ChatGPT Go on its own and below the $20 of either ChatGPT Plus or Claude Pro, against $28 to $40 for running two direct subscriptions. Higher Whizi plans add Claude's larger models and more monthly messages, and they range from below the $28 pairing to above it, so price the tier you actually need rather than assuming the whole ladder undercuts two direct subscriptions. Prices and message limits move, so check [pricing](https://whizi.io/pricing) before you commit either way, and run your own numbers through the [savings calculator](https://whizi.io/tools/calculator) if you are also paying for other AI tools.

This route is the right answer for the medium user who wants the convenience of a normal chat app, a shared history between models, and the ability to switch models mid-conversation, none of which the API route gives you without building it yourself. [How to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers what that switching is actually good for beyond convenience: drafting in one model and asking a different one to find the flaw is a workflow the two separate direct subscriptions cannot do at any price, because nothing connects them.

#### The route to skip: account sharing

Search results for this exact question are full of sellers offering a shared or resold ChatGPT Plus or Claude Pro login, sometimes bundled with a foreign phone number or a prepaid card, at a price below what either provider charges directly. It is worth naming plainly why this is not a real option rather than a cheap one.

Sharing a paid account violates both OpenAI's and Anthropic's terms of service, and both sets of terms reserve the right to suspend or terminate an account for that breach, at the provider's discretion and without warning. Neither company publishes how it enforces this, so treat suspension as a risk you are carrying rather than a certainty either way. If it happens you lose access with no notice and no refund, and you were never the accountholder, so you have no standing to dispute it. Paying a third party for login credentials also means handing a stranger a live session tied to your payment method or a card they control, which is a real exposure with no upside if the arrangement collapses, which these arrangements routinely do.

None of the four routes above require breaking either provider's terms, and all four are genuinely available to anyone with a card or, for the free tiers, nothing at all. If a service you are looking at cannot say plainly that you get your own account, it is not a cheaper version of one of these four routes. It is a different, riskier thing wearing the same price tag.

#### The one-week test that picks your route

The one-week test settles this faster than more reading will. Track how many real exchanges you send to each model over seven normal days, note roughly how long the conversations run, and multiply by four for a rough month. Under a few hundred short exchanges a month, free tiers or the API route are genuinely cheapest. Daily use with some longer sessions is where a multi-model subscription usually beats both a two-subscription stack and an API bill that has started climbing. Sustained heavy use, especially long documents and extended sessions every day, is where a flat plan, whether that is two direct subscriptions or a higher multi-model tier, stops looking expensive and starts looking like the only option that does not require watching a running total.

Key points:

- Two direct subscriptions run about $28 a month if you pair ChatGPT Go with Claude Pro, or about $40 with ChatGPT Plus, and neither price moves whether you send five messages or five thousand
- Pay-as-you-go API access is close to free for light use and can pass $150 a month for heavy use, because the cost scales with every token you send and receive
- The right answer depends on volume: light users do best on free tiers, heavy users do best on a flat plan, and medium users are the ones worth doing the math for

Checklist:

- Track real exchanges with each model for one normal week before picking a route
- Check both providers' supported countries lists first, since a blocked signup is a location limit rather than a payment problem
- Verify current prices on each provider's own pricing page, since many markets are priced locally rather than in dollars
- Price ChatGPT Go before assuming Plus, and only pay the Plus difference for a feature you can name
- If trying the API route, watch the running cost per conversation, not just the token rate
- Check whether your usage fits inside free tier reset windows before assuming it is enough
- Treat any shared or resold login as a terms of service violation, not a cheap subscription
- Compare the flat monthly total of two direct subscriptions against a multi-model plan before renewing either
- Re-run the comparison if your usage shifts from occasional to daily, since the cheapest route changes with volume

FAQ:

Q: What is the actual cheapest way to use ChatGPT and Claude together?
A: It depends on volume. Occasional use: the free tiers of both, at $0. Light to medium daily use: pay-as-you-go API access through a client, often a few dollars a month. Heavy, steady daily use: a flat plan, either two direct subscriptions at about $28 a month for ChatGPT Go plus Claude Pro, or a multi-model subscription that includes both from $15.99 a month, or $10.99 a month billed annually.

Q: Can I use ChatGPT and Claude for free at the same time?
A: Yes. Both providers offer a free tier, and using them side by side costs nothing. The tradeoff is a message cap that resets on a schedule, and some free plans step down to a smaller model during busy periods, so check each provider's current limits before relying on it for daily work.

Q: Is paying for ChatGPT Plus and Claude Pro together worth it?
A: It is worth it if your usage is heavy and steady enough that the flat $40 a month combined price beats what metered API access would cost at the same volume. Before paying it, check whether ChatGPT Go at $8 covers your ChatGPT side, which brings the pair to about $28. For lighter or occasional use, a free tier or pay-as-you-go API access almost always costs less for the same work.

Q: Is it safe to buy a shared ChatGPT or Claude account to save money?
A: No. Sharing or reselling a paid account violates both OpenAI's and Anthropic's terms of service, and both sets of terms let the provider suspend the account for it without warning. You have no standing to dispute a suspension since you are not the accountholder, and you are handing a stranger access tied to someone's payment method with no recourse if it goes wrong.

Q: What is a cheaper alternative to paying for ChatGPT Plus and Claude Pro separately?
A: A multi-model subscription that includes both GPT and Claude in one plan, such as Whizi at $15.99 a month for Starter ($10.99 a month billed annually), against $20 for either provider's own standard plan and $28 to $40 for both. Pay-as-you-go API access through a client is cheaper still for light to medium use, though it takes more setup and has no flat monthly cost cap.

### What is Claude AI: The Best AI for Writing and Beginners

URL: https://whizi.io/resources/claude/
Updated: August 2026

Quick answer: Claude is an AI assistant made by Anthropic. You type, it answers, the same as ChatGPT or Gemini. What sets it apart is narrow and specific: it writes better, so a draft usually needs the least fixing before you would send it, and it reads long documents without flattening the caveats.

Discover what Claude AI is, how it differs from ChatGPT, and why it’s considered the best AI for writing, summarizing, and natural conversation.

#### What Claude is

Claude is an AI assistant made by Anthropic. You type, it answers, the same as ChatGPT or Gemini. If you have used any of them, you already know how to use this one.

The reason people single Claude out is narrower and more specific than most descriptions suggest: it writes better. Ask three assistants to draft the same email and Claude's version usually needs the least fixing before you would send it. It uses fewer filler phrases, does not reach for words like "revolutionary" and "seamless", and handles awkward situations, like declining something or delivering bad news, with more tact.

Its second strength is reading. You can give Claude a long document and ask questions about it, and it is unusually good at noticing qualifications and caveats rather than flattening everything into a summary. It is also good at telling you what a document does not say, which is often what you actually wanted to know.

#### What it is not the best at

Worth knowing up front so you are not disappointed by the wrong thing.

**Recent events.** Claude's knowledge stops at its training date. Unless you are using a version connected to the web, it will answer a question about last month from general knowledge, and it will sound confident doing it.

**Very long documents.** It handles long documents well, but Gemini handles much longer ones. If you are working with a 300 page report or twenty files at once, that is a Gemini job.

**Rigid formatting at scale.** If you need two hundred rows extracted into exactly the same shape, ChatGPT is more reliable at holding a strict format without drifting.

None of that makes Claude worse. It makes it a specialist, and knowing what it specialises in is the whole point of this page.

| What you are doing | Best choice | Why |
| --- | --- | --- |
| An email or message that matters | Claude | The draft needs the least fixing before you would send it |
| Bad news, declining something, honest criticism | Claude | Handles awkward situations with more tact |
| Reading a document where the caveats matter | Claude | Keeps the qualifications instead of flattening them |
| Anything from the last few weeks | Another model | Claude answers from its training date unless it is connected to the web |
| A 300 page report or twenty files at once | Gemini | Handles much longer documents |
| Two hundred rows extracted into exactly the same shape | ChatGPT | More reliable at holding a strict format without drifting |

#### The four things beginners actually use it for

**1. Making rough thoughts into decent writing.** This is the most common and the most useful. Type your messy version and hand it over.

`Here is a rough version of what I want to say. Rewrite it so it sounds professional but still like a person. Keep every fact exactly as I have it. Do not add flattery. Under 150 words. Draft: [paste].`

**2. The difficult message.** Complaints, declines, apologies, chasing someone who owes you money. This is where Claude noticeably beats the alternatives.

`Write a message telling [person] that [difficult thing]. Be direct but not cold. Do not over-apologise. Do not add hedging that invites an argument about the fact itself. End with one clear next step and a date.`

**3. Understanding a long document.** Upload it rather than pasting a summary of it.

`Read the attached document and tell me: the five things that matter most, anything that would cost me money or create an obligation, and anything important that is missing or left vague. Quote the exact text for each point so I can check it.`

**4. Honest feedback.** Claude gives real criticism, where most assistants give encouragement.

`Tell me the three weakest things about this. Quote each one. Do not rewrite it and do not tell me what is good about it. Draft: [paste].`

#### The habit that makes it sound like you

Everything above works better with one addition, and it is the single most valuable trick on this page.

Paste two or three paragraphs of your own actual writing into the prompt and say `match this voice`. Anything works: an old email you were happy with, something from your website, a message you sent a colleague last week.

This does more than any amount of describing your tone. "Write in a friendly, professional voice" gives the model the average of everything friendly and professional ever written. Two paragraphs of your real writing gives it you.

Keep those paragraphs somewhere you can copy them from, and use them every time. It takes five seconds and it is the difference between output you rewrite and output you edit.

#### Getting the most out of it

- **Say who is reading it.** "For my manager, who is busy and dislikes long emails" changes the output more than any instruction about style.
- **Say what not to do.** Claude follows negative instructions well. `No bullet points`, `do not use the word solution`, `do not start with In today's`.
- **Give it the facts rather than asking it to know them.** It writes well from material you provide and invents plausibly when you leave a gap.
- **Ask for criticism before a rewrite.** If you ask it to improve something, it will polish forever. If you ask what is weak first, you get a different draft rather than a shinier one.
- **Push back if it caves too easily.** Like all assistants, it will sometimes abandon a correct answer when you disagree. Ask `is that actually right, or are you agreeing with me?`

#### Getting access to it

Anthropic offers a free tier that is genuinely usable, with limits on how much you can use the more capable models. For occasional writing help it may be all you need, and the honest catch is that limits arrive when you are busiest.

The paid plan is around $20 a month for Claude alone. That is the calculation worth pausing on, because most people who want Claude also want something else: a model that knows what happened this week, or one that reads a 300 page document. Paying $20 for Claude and another $20 for one of those is a lot for two chat windows.

A workspace like Whizi includes the Claude models alongside ChatGPT and Gemini for less than Claude alone costs, which also lets you draft in Claude and get a second opinion from a different model without paying twice. See [Whizi vs Claude Pro](https://whizi.io/vs/whizi-vs-claude-pro) if you already subscribe.

#### Things to be careful about

**It invents things.** Less than some, but it happens, and with no change in tone. Check any specific fact, date, number, or citation before you rely on it.

**It does not know what happened recently** unless connected to the web.

**It agrees too readily.** Ask what it thinks before telling it what you think.

**Be careful what you paste.** Not the place for passwords, financial details, or other people's personal information. Removing names costs nothing in answer quality.

**Editing beats generating.** The best results come from giving it your material and asking it to improve the writing, not from asking it to produce something out of nothing. That is true of every assistant and most true of this one.

Key points:

- What Claude is genuinely better at, and what it is not, without the marketing
- Copy-ready prompts for the four things beginners use it for most
- The one habit that makes Claude write like you rather than like a chatbot

Checklist:

- Keep two paragraphs of your own writing handy and paste them into every writing prompt
- Give Claude your rough draft rather than asking it to start from nothing
- Say who will read it, and what not to do
- Use it specifically for difficult messages, where it is clearly better
- Upload long documents and ask for exact quotes you can check
- Ask what is weak before asking for a rewrite
- Use a different model for anything recent or very long

FAQ:

Q: Is Claude better than ChatGPT?
A: For writing that a person reads end to end, yes, and it is the most consistently reported difference between them. ChatGPT is better at structure, formatting, and anything with a fixed shape such as plans and tables. Neither knows about recent events without web access, and for very long documents Gemini beats both. The realistic answer is that they are specialists rather than competitors.

Q: Do I need a separate subscription for Claude?
A: Going directly through Anthropic, yes, at around $20 a month, and there is a free tier that is genuinely usable for occasional work. The reason people look for alternatives is that most who want Claude also want something else, such as live web access or much longer document handling, and paying separately for each adds up quickly. A combined workspace covers the Claude models alongside the others for less than Claude alone costs.

Q: What is Claude best at, specifically?
A: Three things. Writing that sounds like a person rather than a template, which is where it clearly leads. Difficult tone, meaning bad news, declining requests, and critical feedback delivered without being cold. And careful reading of documents where meaning depends on qualification, including telling you what a document does not say, which is frequently the actual question.

Q: Can Claude read PDFs and long documents?
A: Yes, and it handles nuance in them better than most, keeping the caveats rather than flattening them into a summary. For very long documents, meaning hundreds of pages or many files at once, Gemini can hold considerably more at a time and is the better choice. Whichever you use, ask for the exact quote supporting an answer so you can verify it came from your document.

Q: How do I stop it sounding like AI?
A: Paste two or three paragraphs of your own real writing and ask it to match that voice. This works far better than describing your tone in adjectives, because a description gives the model an average and an example gives it you. Then give it your actual facts and ask what is weak before requesting a rewrite, which produces a different draft rather than a smoother one.

### Claude vs ChatGPT: which should you use?

URL: https://whizi.io/resources/claude-vs-chatgpt/
Updated: August 2026

Quick answer: Claude is usually the stronger choice for polished writing, long-context review, and careful synthesis, while ChatGPT is usually stronger for fast iteration, broad daily tasks, coding scaffolds, and tool-heavy workflows. Neither wins outright. Run the same real prompt in both, score accuracy and cleanup time, then set your default per task.

A practical Claude vs ChatGPT comparison for writing, coding, research, and daily workflows, with a decision checklist you can use before paying for another AI plan.

#### At-a-glance differences

Short answer: use Claude when the job depends on nuance, tone, long context, careful synthesis, or a polished first draft. Use ChatGPT when the job depends on broad utility, fast back-and-forth iteration, coding help, structured tool use, or a flexible everyday assistant. That is not a permanent rule. It is a starting point for deciding which model gets the first attempt.

The honest Claude vs ChatGPT answer is not "one is better." The better question is: which model gives you the best output for this specific task, with the least cleanup, at a cost you can justify? A product manager writing a strategy memo may prefer Claude for the first draft and ChatGPT for counterarguments. A developer may use ChatGPT to sketch a fix, Claude to review the reasoning, then ChatGPT again to turn it into test cases. A founder researching pricing pages may run both and trust the model that extracts clearer evidence.

| Decision factor | Claude is usually a strong fit when... | ChatGPT is usually a strong fit when... |
| --- | --- | --- |
| Writing | You need a natural draft, executive memo, rewrite, or careful tone pass | You need ideation, outlines, variants, or fast copy experiments |
| Coding | You want a patient review of architecture, risk, or a larger code snippet | You want quick debugging, implementation help, tool-assisted iteration, or tests |
| Research | You need synthesis across long notes or documents with careful caveats | You need broad exploration, query brainstorming, or structured extraction |
| Tool use | You want an agentic workflow that follows explicit tool instructions | You want mature function calling patterns and app-style workflows |
| Buying decision | You want one model that feels especially strong for text-heavy work | You want a broad assistant that can handle many different daily jobs |

A useful rule: do not choose from brand reputation alone. Pick three prompts from your real work, run them in both models, score the outputs, then decide. If you are comparing paid subscriptions, use the [AI subscription savings calculator](https://whizi.io/tools/calculator) before stacking more plans. If you want the broader comparison hub, read [ChatGPT vs Claude vs Gemini](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini).

#### Writing

For many teams, the strongest argument for Claude is writing quality. Claude often performs well when the task requires a human-feeling voice, long-form organization, editorial restraint, or sensitivity to context. It can be especially useful for rewriting a rough memo, polishing customer-facing explanations, condensing interview notes, or turning messy bullet points into a coherent narrative.

That does not make ChatGPT weak for writing. ChatGPT is excellent for brainstorming angles, creating options, pressure-testing positioning, drafting social variants, building outlines, and converting one idea into many formats. If you want 20 headline directions, a landing page structure, a comparison matrix, and three email versions in one sitting, ChatGPT can be a very efficient writing partner.

Try this same prompt in both models: `You are editing a founder memo. Keep the argument direct, remove vague claims, preserve the founder voice, and return: 1) a sharper draft, 2) five edits you made, and 3) three claims that need evidence.` Claude may give you a smoother narrative pass. ChatGPT may give you sharper options and more structured alternatives. The winner is the one that needs less human editing for your audience.

Use Claude first for executive summaries, sensitive customer emails, fundraising updates, policy explanations, and long-form thought leadership. Use ChatGPT first for campaign ideas, outlines, short-form variants, structured briefs, and rapid copy exploration. For high-stakes content, run both: ask Claude for the polished draft, ask ChatGPT to critique gaps, then ask the stronger model to produce the final version. If neither voice fits the way you write, [ChatGPT alternatives for writing](https://whizi.io/resources/chatgpt-alternatives-for-writing) compares the rest on tone and on how much editing they leave behind.

#### Coding

For coding, the ChatGPT vs Claude decision depends on the workflow. ChatGPT is often a strong first stop for quick debugging, code generation, tests, API usage, and step-by-step implementation help. OpenAI also documents model and function-calling patterns that make ChatGPT a natural fit for structured app workflows where a model needs to call tools, return arguments, or fit into a developer system.

Claude can be excellent for code review, refactoring plans, larger-context reasoning, and explaining tradeoffs. Many developers reach for it when they want a careful second opinion. What could break if I change this? Where is the hidden coupling? What tests would actually prove this works? Is this abstraction worth it? Those are the questions it answers well, and it is at its most useful when the code sits inside a longer product or architecture discussion rather than a tiny snippet.

Use this reproducible coding workflow instead of asking either model to "fix my code." First, provide the failing behavior, relevant files, expected behavior, logs, and constraints. Second, ask for likely causes before code changes. Third, ask for the smallest fix. Fourth, ask for tests. Fifth, ask the other model to review the proposed patch for regression risk.

Copy-paste prompt: `You are reviewing a bug fix. Before writing code, list the top three failure modes. Then propose the smallest safe change, the tests needed, and any behavior that should not change. If information is missing, ask for it instead of guessing.` Run it in both Claude and ChatGPT. If one model jumps too quickly to code without naming assumptions, use the other as the reviewer. The best AI for coding is often a two-model loop: one drafts, one challenges.

#### Tool use and workflows

Tool use is where the comparison becomes less about personality and more about systems. ChatGPT and OpenAI models are widely used in structured workflows where a model needs to call functions, produce tool arguments, use files, or participate in app logic. Anthropic also documents tool use for Claude, including patterns for giving the model external capabilities through well-defined tools. In practice, both ecosystems support serious workflows, but the right choice depends on the product environment and the task.

For individual users, the workflow question is simpler: where do you want to spend your attention? If you open separate subscriptions, separate chats, and separate histories, you may end up comparing models in your head instead of comparing outputs. That is expensive in both money and focus. If you run the same prompt in both models inside one workspace, you can judge the actual result without turning your day into model management.

A practical Claude vs ChatGPT workflow looks like this: define the task, write one strong prompt, run it in both, score each output on usefulness, accuracy, completeness, and cleanup time, then save the winning prompt as a reusable workflow. For example, a content lead might compare both models on a product launch email. A developer might compare both on a failing test. A founder might compare both on pricing page synthesis.

This is also where pricing becomes real. Claude vs ChatGPT pricing is not just the sticker price of one plan. The real cost is the stack: one subscription for writing, another for coding, another for image work, another for research, plus the switching cost of managing them. Before you buy another standalone plan, see [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) for the direct head-to-head, or run your stack through the [savings calculator](https://whizi.io/tools/calculator). If the conclusion is that you want Claude and GPT together, that is exactly what a [ChatGPT alternative with both built in](https://whizi.io/alternatives/chatgpt) is for, and [the cheapest way to use ChatGPT and Claude together](https://whizi.io/resources/cheapest-way-to-use-claude-and-gpt) prices every route to it, including the two-subscription one.

#### Decision checklist

Use this checklist when someone asks, "Is Claude better than ChatGPT?" The better answer is: better for which job, under which constraints, and with what evidence from your own prompts?

- Choose Claude first when the output must sound polished, careful, diplomatic, or deeply edited.
- Choose Claude first when the prompt includes long notes, transcript material, research excerpts, or a dense document.
- Choose ChatGPT first when you need many ideas quickly, structured options, or fast iteration.
- Choose ChatGPT first when the task is connected to tool calls, app workflows, coding scaffolds, or rapid implementation.
- Run both when the task affects revenue, customers, code quality, legal review, or strategy.
- Do not pay for both separately until you know your usage pattern and total AI subscription cost.

The simplest A/B test takes 10 minutes. Pick one writing task, one coding or analysis task, and one research task. Use the same prompt in both models. Score each response from 1 to 5 on accuracy, completeness, clarity, and edit time. The model with the higher score is your default for that workflow. The model with the lower score may still be useful as a critic, reviewer, or alternative angle generator.

For most teams, the durable answer is not choosing Claude forever or ChatGPT forever. It is building a workflow where the right model handles the right job. Whizi is designed around that reality: compare outputs, keep your prompts reusable, avoid subscription sprawl, and move from decision to execution without opening five different AI tabs. Start with the model comparison, then move to the plan that fits your actual usage at [Whizi pricing](https://whizi.io/pricing) or create an account at [register](https://whizi.io/register).

Key points:

- Claude is often strongest for careful drafting, editing, and long-document synthesis
- ChatGPT is often strongest for fast iteration, broad task coverage, and tool-heavy workflows
- The best answer is usually task-based: run the same prompt in both before you commit

Checklist:

- Run the same real prompt in Claude and ChatGPT before choosing a default
- Use Claude first for polished writing, long context, and careful review
- Use ChatGPT first for fast iteration, coding scaffolds, and tool-heavy workflows
- Score outputs by accuracy, completeness, clarity, and cleanup time
- Check total subscription cost before paying for multiple standalone plans

FAQ:

Q: Is Claude better than ChatGPT?
A: Claude is often better for polished writing, long-context review, and careful synthesis. ChatGPT is often better for broad everyday tasks, fast iteration, coding help, and structured workflows. The best choice depends on the task.

Q: Which is better for writing, Claude or ChatGPT?
A: Claude is a strong first choice for memos, rewrites, long-form editing, and sensitive tone. ChatGPT is a strong first choice for brainstorming, outlines, copy variants, and turning one idea into many formats.

Q: Which is better for coding, Claude or ChatGPT?
A: ChatGPT is often useful for quick debugging, implementation, tests, and tool-assisted development. Claude can be excellent for code review, refactor reasoning, architecture tradeoffs, and larger-context analysis.

Q: Should I pay for both Claude and ChatGPT?
A: Only if your usage justifies it. Many users are better served by comparing models in one workspace and checking total AI subscription cost before adding another separate plan.

### Claude vs DeepSeek: what the price gap does and does not buy

URL: https://whizi.io/resources/claude-vs-deepseek/
Updated: August 2026

Quick answer: Use Claude when a person will read the output and judge you by it: client documents, careful editing, and code changes that need scope discipline. Use DeepSeek for volume work where good enough clears the bar, such as summaries, extraction, and first-pass code, since it costs roughly a tenth as much per answer.

A practical Claude vs DeepSeek comparison: writing quality, coding, reasoning value, hosting and privacy, and when the order-of-magnitude price gap actually matters.

#### The short answer

Use **Claude** when a human being will read the output and judge you by it. Client documents, long-form writing, careful editing, and code changes where the discipline to not rewrite unrelated things matters as much as the fix. That judgment layer is what the price buys.

Use **DeepSeek** for the enormous middle of AI work where good-enough clears the bar and volume matters: summaries, extraction, translation drafts, classification, first-pass code, internal notes. The cost table on this page shows the gap; per answer it is roughly an order of magnitude, and at volume that is not a rounding error.

The mistake in most Claude vs DeepSeek arguments is treating it as one decision. It is a routing decision, made per task, and the people getting the most out of AI in 2026 are making it dozens of times a day.

#### Where Claude is genuinely better

**Prose a person reads end to end.** Claude remains the model people consistently prefer for writing with a voice: memos that sound like you, editing that preserves intent, drafts that need one pass of cleanup instead of three. DeepSeek's English output is competent and noticeably flatter, and on nuanced tone work the gap is visible immediately.

**Restraint in code.** On real codebases, the expensive failure is not a wrong answer; it is an eager one, the model rewriting things you did not ask about. Claude's coding behavior is unusually disciplined about scope, which is why it anchors so many agentic coding tools. DeepSeek writes strong code but needs a tighter leash.

**Long-context care.** Both models accept large inputs; Claude is better at using them faithfully, keeping caveats attached to the claims they qualify across a long document. For contract review and multi-document synthesis, that reliability is the product.

**Judgment under ambiguity.** Vague brief, conflicting constraints, a decision that needs weighing rather than computing: Claude asks the better clarifying question or makes the more defensible call. This is the hardest quality to benchmark and the easiest to feel.

#### Where DeepSeek is genuinely better

**Cost, by an order of magnitude.** This is not a small edge to be traded off against polish. It changes what is economical: running the same job three ways and picking the best, processing every document instead of a sample, giving every row of a dataset its own model call. Workflows that are irresponsible at Claude prices are routine at DeepSeek prices.

**Reasoning value.** On hard math and logic, DeepSeek's reasoning models perform far above their price class. Not the strongest reasoning available, but the strongest per dollar by a wide margin, and for internal analytical work that is usually the right maximum.

**Open weights.** DeepSeek publishes its model weights, so the same models can be run or hosted anywhere. If data location rules out a hosted flagship, this is a capability Claude does not offer at all.

**Volume tolerance.** Rate limits and message caps bind much later when each call costs a fraction of a cent. For bulk work, the cheap model is often the only one you can actually run at the scale the job needs.

#### Head to head by task

| Task | Better choice | Why |
| --- | --- | --- |
| Client-facing writing and editing | Claude | Voice, intent, one-pass cleanup |
| Internal summaries and notes | DeepSeek | Good-enough at a tenth of the cost |
| Code changes on a real codebase | Claude | Scope discipline is the feature |
| First-pass code and boilerplate | DeepSeek | Strong output, price invites iteration |
| Contract and long-document review | Claude | Faithful use of long context |
| Bulk extraction and classification | DeepSeek | The arithmetic is the answer |
| Hard math and logic, internal | DeepSeek | Best reasoning per dollar |
| Hard judgment calls, external stakes | Claude | The tiebreaker you pay for |

If you route by this table for a week, the pattern becomes automatic: DeepSeek is the default, Claude is the escalation, and the escalation earns its price exactly where the stakes justify it.

#### Hosting, privacy, and the part people skip

The models are not the whole comparison. Claude is a hosted service from Anthropic with enterprise-grade data commitments. DeepSeek is three things at once: a free app and API hosted on DeepSeek's servers, and open weights anyone else can host.

If your organization's concern is where data goes, the free DeepSeek app is the wrong answer, and the right DeepSeek answer is the same models served by a provider you have vetted. That distinction, covered properly in [DeepSeek vs ChatGPT](https://whizi.io/resources/deepseek-vs-chatgpt), applies unchanged here: judge the hosting separately from the weights.

Guardrails also differ: DeepSeek's hosted products carry content restrictions on politically sensitive topics that the open weights served elsewhere may not, while Claude's constraints are Anthropic's own and apply everywhere it runs. For most professional work neither set binds; know which one you are getting regardless.

#### How to decide in one afternoon

Take three real tasks from your week: one piece of writing someone else will read, one coding task on your actual code, one bulk-ish job like summarizing five documents.

Run each through both models with the same prompt. Score three things: how much cleanup the output needed, whether anything was confidently wrong, and what the run cost. The cost callout above gives the per-answer baseline; the [Model Cost Index](https://whizi.io/tools/model-cost-index) has the full table.

Most people land in the same place: DeepSeek takes the volume, Claude takes the judgment, and the real productivity gain came from having both available in the same conversation instead of picking a side. That routing workflow, one thread where the cheap model does the bulk work and Claude handles the escalations, is exactly what [Whizi](https://whizi.io/pricing) is built around, with both models in one plan.

Key points:

- The price gap is an order of magnitude; the quality gap depends entirely on the task
- Claude's edge is judgment-heavy work: prose a person will read, and code changes that need restraint
- DeepSeek's edge is volume: everywhere quality above a threshold stops mattering, it wins on arithmetic

Checklist:

- Route by task, not by loyalty: DeepSeek as default, Claude as escalation
- Never ship DeepSeek prose to a client without a read; never pay Claude prices for bulk extraction
- Judge DeepSeek's hosting separately from its weights when privacy is the question
- Run the three-task test with real work before believing anyone's benchmark, including this page
- Recompute the routing when either side ships a major release; the timeline tracks those dates

FAQ:

Q: Is DeepSeek better than Claude?
A: At its price, frequently yes; in absolute terms, usually no. DeepSeek is remarkably capable for roughly a tenth of the cost per answer, and for summaries, extraction, and first-pass code the gap in quality is smaller than the gap in price. For judgment-heavy writing and disciplined code changes, Claude is still clearly ahead.

Q: Is Claude worth the higher price?
A: For the tasks where its edge lives, yes: client-facing prose, careful editing, and code changes with real stakes. For volume work the honest answer is no, and the efficient setup uses Claude only where a person will judge the output.

Q: Which is better for coding, Claude or DeepSeek?
A: Claude for changes to a codebase you maintain, because it stays inside the requested scope. DeepSeek for generation volume: boilerplate, scripts, first drafts, and experiments, where its price lets you iterate freely. Many developers use exactly that split.

Q: Can I use Claude and DeepSeek together?
A: Yes, and it is the setup this comparison keeps arriving at. In Whizi both are in the same model picker from the Pro plan up, so you can draft in one and escalate to the other inside a single conversation, on one bill, without a DeepSeek account.

### Claude vs Gemini: which is better for long-context work?

URL: https://whizi.io/resources/claude-vs-gemini/
Updated: August 2026

Quick answer: Claude is often the better first choice when the output is a human-facing deliverable: memos, edits, briefs, and careful synthesis. Gemini is worth testing first when the input is large or highly structured, such as long PDFs, transcripts, research packets, and schema-shaped extraction. Many teams use both, one drafting and one auditing.

Compare Claude vs Gemini for writing quality, tool use, long documents, PDFs, structured outputs, and research workflows.

#### Capabilities summary

The practical Claude vs Gemini question is not "which company has the better model?" It is "which model should handle this workflow first?" Claude and Gemini can both be useful for writing, research, documents, tool use, and structured analysis. The difference shows up when you compare real tasks: a messy strategy memo, a 90-page PDF, a research pack, a support policy, a product requirements document, or a structured extraction job.

Claude is often a strong fit when the work depends on tone, nuance, careful synthesis, and readable long-form output. Anthropic positions Claude across models with different speed, intelligence, and cost tradeoffs, and its documentation emphasizes practical workflows such as tool use and PDF support. For teams, Claude can feel especially useful when the output must become a memo, brief, response, review, or polished recommendation.

Gemini is especially worth testing when the work depends on long context and structured handling of large inputs. Google documents Gemini long-context workflows for using large bodies of relevant material, and its structured output guidance is useful when the result needs to become JSON, a table, a classification, or another predictable format. If your task starts with "read this whole packet," Gemini deserves a serious test.

| Decision factor | Claude is often a strong fit | Gemini is often a strong fit |
| --- | --- | --- |
| Writing quality | Executive memos, sensitive edits, synthesis, natural prose | Writing from large source packs, summaries from broad context |
| Long documents | PDF review, nuanced synthesis, careful caveats | Very large context windows, large document sets, long transcripts |
| Research | Turning evidence into a readable brief | Processing large source collections before synthesis |
| Tool use | Agent-style workflows with clear tool instructions | Structured extraction and schema-shaped outputs |
| Best first test | Human-facing deliverable | Large input or structured output task |

The table is a starting point, not a verdict. Claude may outperform Gemini on one PDF summary. Gemini may outperform Claude on the next research packet. The right habit is to define the workflow, test the same prompt in both, and judge the answer by accuracy, coverage, format control, and cleanup time.

#### Writing and editing

For writing and editing, Claude often has the clearest appeal. It can be a strong first stop for turning rough notes into a thoughtful memo, rewriting a customer-facing response, editing a long article, tightening an executive update, or converting research into a narrative recommendation. If your success metric is "does this sound like something a capable person would actually send?", Claude should be in the test set.

Gemini can also be useful for writing, especially when the writing depends on a lot of source material. For example, if you have a long transcript, a folder of notes, a product spec, and a competitive analysis, Gemini may be a strong first pass for identifying the source material that needs to shape the final draft. Then Claude may be the better second pass for tone, structure, and reader empathy.

Use this two-model writing workflow when quality matters: first, ask Gemini to map the source pack. Have it identify themes, claims, contradictions, missing evidence, and reusable examples. Second, ask Claude to turn the mapped material into the final deliverable. Third, ask Gemini or Claude to review the final draft against the original sources and flag unsupported claims.

Copy-paste prompt for Claude: `Rewrite this into a polished executive memo. Preserve the facts, remove vague claims, keep the tone direct, and return: 1) the revised memo, 2) five edits you made, 3) claims that need evidence, and 4) questions a skeptical reader may ask.`

Copy-paste prompt for Gemini: `Use the source material below to create a writing brief. Extract the key facts, evidence, contradictions, useful examples, and open questions. Do not write the final draft yet. Return the brief as a table plus a short synthesis.`

#### Tool use patterns

Tool use changes the Claude vs Gemini comparison because the model is no longer just writing an answer. It may need to call a function, use a document, produce arguments, extract records, or follow a schema that another system depends on. Anthropic documents tool use for Claude as a way to give the model external capabilities through defined tools. Google documents structured outputs for Gemini so responses can follow a schema, which matters for extraction, classification, and application workflows.

For business users, the practical lesson is simple: the more structured the workflow, the more explicit the prompt needs to be. Do not ask "analyze this." Ask for fields, constraints, output shape, verification rules, and what to do when information is missing. A good prompt says: "If a field is not present, write Not found. Do not infer. Separate observed evidence from interpretation."

Claude is a strong candidate when the tool workflow depends on multi-step reasoning and a careful explanation of what happened. For example, you might ask Claude to review customer feedback, decide which tool should be used next, and explain its routing decision. Gemini is a strong candidate when the workflow depends on schema-shaped extraction from a large or mixed input pack, such as turning a long document into a table of claims, owners, dates, risks, and follow-up questions.

Here is a simple function calling Claude vs Gemini test: give both models the same imaginary tool list and ask them to choose the next action. Score them on whether they selected the right tool, filled arguments correctly, asked for missing information, and avoided inventing data. Then run a structured extraction test: give both the same source material and ask for a strict table or JSON-like output. The winner may be different for each task.

#### Docs/long-context workflow

Long-context work is where Gemini should always be considered. Google documents Gemini models and workflows around long context, including use cases where the model receives a large amount of source material up front. That makes Gemini a natural first test for long transcripts, research archives, technical specs, market maps, contracts, large PDFs, and document packets where the answer depends on seeing the whole picture.

Claude is still highly relevant for long documents. Anthropic documents PDF support, and Claude can be strong at turning dense material into a readable synthesis with caveats. If the task is "understand this large document and write the memo a human needs," Claude may outperform a model that merely extracts more data. Long context is not only about how much text can fit. It is about whether the model can prioritize, reason, and communicate the answer usefully.

Use this workflow for serious document work:

1. Inventory the source pack. Ask for sections, tables, entities, dates, claims, and missing context.
2. Extract evidence into a structured table. Require source locations when available.
3. Separate facts from interpretation. Make the model mark uncertain items.
4. Ask the second model to challenge the extraction and identify omissions.
5. Produce the final output: memo, decision table, research brief, requirements doc, or action list.
6. Manually verify numbers, quotes, legal claims, financial claims, medical claims, and anything that affects customers.

Prompt for long documents: `Analyze this document packet for a decision. First create a document map. Then extract claims, numbers, risks, decisions, open questions, and evidence. Do not infer missing facts. Put uncertain items in a separate section. Finish with a decision table: recommendation, evidence, confidence, owner, and what must be verified manually.`

This workflow keeps both models honest. Gemini gets a fair chance to use long context. Claude gets a fair chance to produce the human-facing synthesis. Whizi is useful here because you can run the same prompt across models without rebuilding the context in separate tools.

#### Decision rules

Use Claude first when the output is a human-facing deliverable: a memo, edit, brief, customer response, policy explanation, investor update, or long-form synthesis. Use Gemini first when the input is large, mixed, or highly structured: long PDFs, transcripts, research packets, source collections, tables, specs, or document sets. Use both when the decision is important enough that a second model can catch gaps.

A fast scoring rubric helps remove brand bias. Give each output 1 to 5 points for source accuracy, coverage, clarity, format compliance, reasoning quality, and cleanup time. If Gemini covers more evidence but Claude writes the better deliverable, use Gemini for extraction and Claude for the final draft. If Claude gives a stronger synthesis but misses source details, ask Gemini to audit the evidence.

Decision rules you can actually use:

- Choose Claude for final prose, nuanced editing, careful review, and reader-ready synthesis.
- Choose Gemini for large context, source inventory, structured extraction, and document-heavy analysis.
- Use Claude as the critic when Gemini produces a dense table that needs judgment.
- Use Gemini as the auditor when Claude produces a polished narrative that needs source checking.
- Compare both before buying another standalone subscription.
- Save the winning prompt as a repeatable workflow instead of debating the model again next week.

For the broader comparison across the three major assistants, read [ChatGPT vs Claude vs Gemini](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini), or see how [Gemini stacks up against ChatGPT](https://whizi.io/resources/gemini-vs-chatgpt) head to head. When the buying decision becomes real, compare your expected usage against [Whizi pricing](https://whizi.io/pricing), then create your account at [register](https://whizi.io/register). The goal is not to crown one permanent winner. The goal is a unified workspace where Claude, Gemini, and other models can each handle the work they are best suited for.

Key points:

- Claude is often a strong first choice for nuanced writing, editing, synthesis, and careful review
- Gemini is especially worth testing when the task depends on long context, large source packs, or structured extraction
- Most teams should choose by workflow: run the same task in both, then keep the model that produces the most usable output

Checklist:

- Use Claude first for polished writing, nuanced editing, and careful synthesis.
- Use Gemini first for long-context document packets, source inventory, and structured extraction.
- Run the same prompt in both models before choosing a default workflow.
- Score outputs on accuracy, coverage, clarity, format compliance, reasoning quality, and cleanup time.
- Use one model to draft and the other to audit when the work affects customers, strategy, or revenue.

FAQ:

Q: Is Claude or Gemini better for writing?
A: Claude is often a strong first choice for polished prose, editing, and nuanced synthesis. Gemini can be a strong writing partner when the draft depends on large source material that needs to be mapped first.

Q: Is Gemini better than Claude for long documents?
A: Gemini is especially worth testing for large-context workflows because Google documents long-context use cases for Gemini. Claude can still be very strong for PDF review and readable synthesis, so the best choice depends on the document and output.

Q: Which is better for structured output?
A: Gemini has official structured output guidance that is useful for schema-shaped extraction. Claude is also useful in tool workflows where reasoning, review, and explanation matter. Test both with your exact schema and validation rules.

### Context window comparison: every major AI model in one table

URL: https://whizi.io/resources/context-window-comparison/
Updated: August 2026

Quick answer: The largest context window in the current catalogue is 1M tokens, held by DeepSeek V4 Flash 0731 from DeepSeek. The tables below compare every major AI model in tokens and in plain pages, generated from the same live catalogue as the AI Model Cost Index. One token is about three quarters of an English word.

A living context window comparison across every major AI model: GPT, Claude, Gemini, Grok, DeepSeek, Llama, and more, in tokens and in plain pages of text.

#### The quick answer

The largest context window in the current catalogue is **1M tokens**, held by DeepSeek V4 Flash 0731 from DeepSeek. As a rough physical intuition, one token is about three quarters of an English word, and a manuscript page is about 400 words, so the table below translates every window into pages.

One row per major provider, showing its largest-context model:

| Provider | Largest-context model | Context window | Fits roughly |
| --- | --- | --- | --- |
| **DeepSeek** | DeepSeek V4 Flash 0731 | 1M tokens | about 2,500 pages |
| **OpenAI** | GPT-5.4 | 1M tokens | about 2,000 pages |
| **Google** | Gemini 2.5 Flash | 1M tokens | about 2,000 pages |
| **Meta** | Llama 4 Maverick | 1M tokens | about 2,000 pages |
| **Qwen** | Qwen3.8 2.4T A95B | 1M tokens | about 2,000 pages |
| **Moonshot** | Kimi K3 | 1M tokens | about 2,000 pages |
| **Z.ai** | GLM 5.2 | 1M tokens | about 2,000 pages |
| **MiniMax** | MiniMax M3 | 1M tokens | about 2,000 pages |
| **Anthropic** | Claude Fable 5 | 1M tokens | about 1,900 pages |
| **xAI** | Grok 4.3 | 1M tokens | about 1,900 pages |
| **Mistral** | Mistral Large 3 2512 | 262K tokens | about 492 pages |

If the question behind your search is "what is a context window in the first place", start with the plain-English explainer at [What is a context window?](https://whizi.io/resources/what-is-a-context-window) and come back for the numbers.

#### Every model in the index (100 models)

The full catalogue, sorted by context window. The data comes from the same source as the [AI Model Cost Index](https://whizi.io/tools/model-cost-index) and was last fetched 2026-08-20; when the catalogue updates, this table updates with it. The credits column shows what each model costs per message inside Whizi, where a plan includes it.

| Model | Provider | Context window | Fits roughly | Credits in Whizi |
| --- | --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 | DeepSeek | 1M | about 2,500 pages | 1 |
| GPT-5.4 | OpenAI | 1M | about 2,000 pages | 15 |
| GPT-5.5 | OpenAI | 1M | about 2,000 pages | 20 |
| GPT-5.6 Luna | OpenAI | 1M | about 2,000 pages | 1 |
| GPT-5.6 Luna Pro | OpenAI | 1M | about 2,000 pages | Not metered |
| GPT-5.6 Sol | OpenAI | 1M | about 2,000 pages | 20 |
| GPT-5.6 Sol Pro | OpenAI | 1M | about 2,000 pages | Not metered |
| GPT-5.6 Terra | OpenAI | 1M | about 2,000 pages | 4 |
| GPT-5.6 Terra Pro | OpenAI | 1M | about 2,000 pages | Not metered |
| LongCat 2.0 | Meituan | 1M | about 2,000 pages | 1 |
| DeepSeek V4 Pro 0813 | DeepSeek | 1M | about 2,000 pages | 2 |
| Gemini 2.5 Flash | Google | 1M | about 2,000 pages | 2 |
| Gemini 2.5 Flash Lite | Google | 1M | about 2,000 pages | 1 |
| Gemini 3.1 Flash Lite | Google | 1M | about 2,000 pages | 1 |
| Gemini 3.1 Pro Preview | Google | 1M | about 2,000 pages | 10 |
| Gemini 3.5 Flash | Google | 1M | about 2,000 pages | 8 |
| Gemini 3.5 Flash Lite | Google | 1M | about 2,000 pages | 2 |
| Gemini 3.6 Flash | Google | 1M | about 2,000 pages | 3 |
| Gemini 3.7 Flash | Google | 1M | about 2,000 pages | 2 |
| GLM 5.2 | Z.ai | 1M | about 2,000 pages | 3 |
| GLM 5.3 | Z.ai | 1M | about 2,000 pages | 5 |
| Inkling | Thinkingmachines | 1M | about 2,000 pages | 4 |
| Kimi K3 | Moonshot | 1M | about 2,000 pages | 10 |
| Laguna S 2.1 | Poolside | 1M | about 2,000 pages | 1 |
| Llama 4 Maverick | Meta | 1M | about 2,000 pages | 1 |
| MiniMax M3 | MiniMax | 1M | about 2,000 pages | 1 |
| Muse Spark 1.1 | Meta | 1M | about 2,000 pages | 5 |
| Muse Spark 1.2 | Meta | 1M | about 2,000 pages | 5 |
| Qwen3.8 2.4T A95B | Qwen | 1M | about 2,000 pages | 8 |
| Claude Fable 5 | Anthropic | 1M | about 1,900 pages | 50 |
| Claude Opus 4.7 (Fast) | Anthropic | 1M | about 1,900 pages | Not metered |
| Claude Opus 4.8 | Anthropic | 1M | about 1,900 pages | 20 |
| Claude Opus 4.8 (Fast) | Anthropic | 1M | about 1,900 pages | Not metered |
| Claude Opus 5 | Anthropic | 1M | about 1,900 pages | 20 |
| Claude Opus 5 (Fast) | Anthropic | 1M | about 1,900 pages | Not metered |
| Claude Sonnet 4.5 | Anthropic | 1M | about 1,900 pages | 10 |
| Claude Sonnet 4.6 | Anthropic | 1M | about 1,900 pages | 10 |
| Claude Sonnet 5 | Anthropic | 1M | about 1,900 pages | 10 |
| Fugu Ultra | Sakana | 1M | about 1,900 pages | 25 |
| Grok 4.3 | xAI | 1M | about 1,900 pages | 4 |
| Qwen3.5 Plus 2026-04-20 | Qwen | 1M | about 1,900 pages | 1 |
| Qwen3.6 Flash | Qwen | 1M | about 1,900 pages | 1 |
| Qwen3.7 Flash | Qwen | 1M | about 1,900 pages | 1 |
| Qwen3.7 Max | Qwen | 1M | about 1,900 pages | 5 |
| Qwen3.7 Plus | Qwen | 1M | about 1,900 pages | 1 |
| Qwen3.8 27B | Qwen | 1M | about 1,900 pages | 3 |
| Qwen3.8 Max | Qwen | 1M | about 1,900 pages | 6 |
| Inkling Small | Thinkingmachines | 524K | about 983 pages | 1 |
| Solar Pro 4 | Upstage | 524K | about 983 pages | 1 |
| Nemotron 3 Ultra | NVIDIA | 512K | about 961 pages | 3 |
| Grok 4.5 | xAI | 500K | about 938 pages | 6 |
| Grok 4.6 | xAI | 500K | about 938 pages | 6 |
| GPT Chat Latest | OpenAI | 400K | about 750 pages | 25 |
| GPT-5.4 Mini | OpenAI | 400K | about 750 pages | 4 |
| GPT-5.4 Nano | OpenAI | 400K | about 750 pages | 2 |
| Nova Pro 1.0 | Amazon | 300K | about 563 pages | 3 |
| Hy3 | Tencent | 262K | about 492 pages | 1 |
| Kimi K2 Thinking | Moonshot | 262K | about 492 pages | 10 |
| Kimi K2.7 Code | Moonshot | 262K | about 492 pages | 10 |
| Laguna XS 2.1 | Poolside | 262K | about 492 pages | 1 |
| Ling-3.0-flash | inclusionAI | 262K | about 492 pages | 1 |
| Mistral Large 3 2512 | Mistral | 262K | about 492 pages | 2 |
| Mistral Medium 3.5 | Mistral | 262K | about 492 pages | 6 |
| Nemotron 3.5 Lightning | NVIDIA | 262K | about 492 pages | 1 |
| Nex-N2-Mini | Nex Agi | 262K | about 492 pages | 1 |
| Nex-N2-Pro | Nex Agi | 262K | about 492 pages | 1 |
| Qwen3 Coder Next | Qwen | 262K | about 492 pages | 1 |
| Qwen3.6 35B A3B | Qwen | 262K | about 492 pages | 1 |
| Ring-2.6-1T | inclusionAI | 262K | about 492 pages | 1 |
| Sakana Namazu | Sakana | 262K | about 492 pages | 4 |
| Seed 2.1 Turbo | Bytedance Seed | 262K | about 492 pages | 2 |
| Seed-2.0-Code | Bytedance Seed | 262K | about 492 pages | 3 |
| Step 3.7 Flash | Stepfun | 262K | about 492 pages | 1 |
| Codestral 2508 | Mistral | 256K | about 480 pages | 1 |
| Command A | Cohere | 256K | about 480 pages | 10 |
| Grok Build 0.1 | xAI | 256K | about 480 pages | 3 |
| KAT-Coder-Air V2.5 | Kwaipilot | 256K | about 480 pages | 1 |
| KAT-Coder-Pro V2.5 | Kwaipilot | 256K | about 480 pages | 3 |
| GLM 4.6 | Z.ai | 205K | about 384 pages | 1 |
| GLM 4.7 | Z.ai | 205K | about 384 pages | 2 |
| GLM 5 | Z.ai | 205K | about 384 pages | 2 |
| MiniMax M2 | MiniMax | 205K | about 384 pages | 1 |
| Claude Haiku 4.5 | Anthropic | 200K | about 375 pages | 4 |
| DeepSeek V3.1 | DeepSeek | 164K | about 307 pages | 1 |
| DeepSeek V3.2 | DeepSeek | 164K | about 307 pages | 1 |
| Aion-3.0 | Aion Labs | 131K | about 246 pages | 10 |
| Aion-3.0-Mini | Aion Labs | 131K | about 246 pages | 2 |
| Granite 4.1 8B | Ibm Granite | 131K | about 246 pages | 1 |
| Kimi K2 0711 | Moonshot | 131K | about 246 pages | 10 |
| Llama 3.3 70B Instruct | Meta | 131K | about 246 pages | 1 |
| Mistral Medium 3.1 | Mistral | 131K | about 246 pages | 2 |
| Muse Glimmer 30B | Meta | 131K | about 246 pages | 2 |
| Nano Banana 2 (Gemini 3.1 Flash Image) | Google | 131K | about 246 pages | Not metered |
| Nano Banana Pro (Gemini 3 Pro Image) | Google | 131K | about 246 pages | Not metered |
| Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) | Google | 66K | about 123 pages | Not metered |
| R1 | DeepSeek | 64K | about 120 pages | 2 |
| Perceptron Mk1 | Perceptron | 33K | about 61 pages | 1 |
| Phi 4 | Microsoft | 16K | about 31 pages | 1 |
| Hy-MT2-1.8B | Tencent | 8K | about 15 pages | Not metered |
| Hy-MT2-30B-A3B | Tencent | 8K | about 15 pages | Not metered |

#### How to read these numbers honestly

**The window is shared.** Everything counts against it at once: your prompt, every earlier message in the conversation, any documents you attached, and the answer being written. A model with a 128K window that has already read a 100K-token contract does not have 128K left for the discussion; it has 28K.

**Advertised is not effective.** Long-context research keeps finding the same pattern: models recall the start and end of a huge context better than the middle. A model can accept a million tokens and still gloss over the clause buried at token 400,000. For high-stakes work on long documents, ask specifically about the middle sections, or feed the document in labelled parts.

**Bigger is not automatically better.** Large windows cost more to run and can dilute attention across irrelevant text. If your work is emails and short drafts, the window size will never be the thing you notice. It starts to matter at whole contracts, codebases, transcripts, and books.

**Tokens are not words.** The three-quarters rule is a good English average, but code, numbers, and non-English languages often tokenize less efficiently, so the page estimates in the table are deliberately rough.

#### Using the right window for the job

The practical conclusion from the table is that context window is a per-task choice, not a per-subscription choice. The model with the biggest window is rarely the best writer, and the best writer is rarely the cheapest way to summarize 800 pages.

That is the argument for having the catalogue in one workspace. In Whizi you can hand the 200-page PDF to a large-context model, then switch to Claude or GPT in the same conversation to draft from what it found, without the document leaving the thread. The [how to choose a model](https://whizi.io/resources/how-to-choose-ai-model) guide covers the same decision for quality and cost.

For citing this page: the anchor-free URL is stable, the data date is printed above the full table, and the token figures come from the provider catalogue rather than from marketing pages.

Key points:

- The single largest window in the current catalogue is 1M tokens (DeepSeek V4 Flash 0731)
- Every figure is shown in tokens and in plain pages, because nobody thinks in tokens
- The table is generated from the same live catalogue as the AI Model Cost Index, not maintained by hand

Checklist:

- Subtract your documents and history from the window before trusting the headline number
- For long documents, test recall from the middle, not just the start
- Translate tokens to pages with the three-quarters rule before deciding anything
- Pick the window per task instead of per subscription
- Check the data date above the full table when citing a figure

FAQ:

Q: Which AI model has the largest context window?
A: In the current catalogue it is DeepSeek V4 Flash 0731 from DeepSeek at 1M tokens. The full table above lists all 100 models sorted by window size, with the data date printed above it.

Q: What is the largest Claude context window?
A: The largest Claude context window in the index is 1M tokens (Claude Fable 5), which fits about 1,900 pages of text. Smaller Claude variants are listed in the full table above.

Q: What is the largest GPT context window?
A: The largest GPT context window in the index is 1M tokens (GPT-5.4), which fits about 2,000 pages of text. Smaller GPT variants are listed in the full table above.

Q: What is the largest Gemini context window?
A: The largest Gemini context window in the index is 1M tokens (Gemini 2.5 Flash), which fits about 2,000 pages of text. Smaller Gemini variants are listed in the full table above.

Q: What is the largest Grok context window?
A: The largest Grok context window in the index is 1M tokens (Grok 4.3), which fits about 1,900 pages of text. Smaller Grok variants are listed in the full table above.

Q: Is a bigger context window always better?
A: No. Larger windows cost more, and recall degrades toward the middle of a very large context. Bigger is decisive for whole books, codebases, and long transcripts, and irrelevant for everyday chat. Match the window to the task rather than paying for the biggest number.

Q: How many pages fit in a context window?
A: Divide the token count by roughly 530 to get manuscript pages: one token is about three quarters of an English word, and a page is about 400 words. So 128K tokens is about 240 pages, and a million tokens is about 1,900 pages. Code and non-English text usually fit less.

### Copilot vs ChatGPT: same models, very different products

URL: https://whizi.io/resources/copilot-vs-chatgpt/
Updated: August 2026

Quick answer: Choose Copilot if your day runs through Word, Excel, Outlook, and Teams, because it can act on your actual files, mail, and calendar. Choose ChatGPT for everything else: research, code, long conversations, and model choice. Both run substantially on OpenAI models, so the comparison comes down to packaging and data access rather than intelligence.

A practical Copilot vs ChatGPT comparison: what Microsoft actually sells, where the Office integration earns its price, and when a general assistant beats an in-app one.

#### The short answer

Use **Copilot** if your working day is Word, Excel, Outlook, PowerPoint, and Teams, and what you want is AI inside those windows: draft this email, summarize this thread, build this slide from this document. The integration is the product, and inside Microsoft 365 it does things a browser tab structurally cannot.

Use **ChatGPT** for everything that is not an Office document: open-ended thinking, research, code, long conversations, projects, and any work where you want to pick the model and control the prompt. As a general assistant it is simply the deeper product.

The twist that makes this comparison different from Grok vs ChatGPT or Claude vs ChatGPT: Copilot is substantially built on OpenAI models. You are rarely comparing intelligence; you are comparing packaging, context, and control over the same underlying capability.

#### First, which Copilot do you mean?

Microsoft ships several products under one name, and most disappointment with "Copilot" starts with buying the wrong one.

**Microsoft 365 Copilot** is the in-app assistant inside Word, Excel, Outlook, PowerPoint, and Teams, sold per seat to organizations. **Copilot Pro** is the consumer subscription that adds in-app features to personal Microsoft 365 apps. **Copilot chat** is the free-to-try general assistant in the browser, on Windows, and in Edge. **GitHub Copilot** is a different product entirely, for code in your editor; if that is what you meant, the comparison you want is against coding assistants, not ChatGPT.

This article compares the Microsoft 365 and Pro versions, plus Copilot chat, against ChatGPT, because that is what people searching this phrase are choosing between.

#### Where Copilot genuinely earns it

**It can see your files, mail, and calendar.** The honest advantage. "Summarize the email thread with the client and draft a reply consistent with what we promised in the attached proposal" is a prompt Copilot can execute against your actual mailbox and files, with permissions handled by the same Microsoft identity your company already manages. Recreating that in ChatGPT means copying content out by hand.

**It works where the document already is.** Rewriting a paragraph in place in Word, or generating formulas against the spreadsheet that is already open, removes the copy-paste loop entirely. For high-volume Office work that loop is most of the friction.

**IT departments can actually approve it.** Tenant boundaries, data residency, audit, and admin controls ride on existing Microsoft agreements. In many organizations Copilot is not competing with ChatGPT; it is competing with "no AI allowed", and winning.

**The free Copilot chat is a real product.** For someone fully inside the Microsoft ecosystem who wants a general assistant without another account, it is a legitimate starting point.

#### Where ChatGPT is clearly ahead

**Model control.** In ChatGPT you choose the model, including the reasoning tiers, and you know what you are getting. Copilot routes your request through whatever Microsoft has provisioned for that feature, and the in-app experiences often feel a step behind what the same underlying models do when prompted directly.

**Everything outside Office.** Long-form thinking, research with browsing, data analysis on arbitrary files, image work, voice, custom instructions, and projects. The general-assistant surface is far larger and improves faster.

**Prompt depth.** In-app Copilot features are largely buttons and short boxes wrapped around templates. ChatGPT rewards the detailed, iterative prompting that produces genuinely good output. Power users hit Copilot's ceiling quickly.

**Consistency.** Copilot's quality varies noticeably by app: strong in Teams summaries, weaker in Excel, uneven in PowerPoint. ChatGPT is one product with one quality bar.

#### Head to head by task

| Task | Better choice | Why |
| --- | --- | --- |
| Email triage and drafting in Outlook | Copilot | It can read the thread and your calendar |
| Meeting summaries in Teams | Copilot | Native access to the transcript |
| Rewriting inside a Word document | Copilot | In-place editing beats copy-paste |
| Open-ended research and thinking | ChatGPT | Deeper prompting, browsing, reasoning tiers |
| Code you will ship | ChatGPT | Reasoning models plus a real conversation |
| Excel formulas and quick analysis | Split | Copilot for in-sheet basics, ChatGPT for real analysis on exported data |
| Slides | Split | Copilot drafts from your files; the thinking is better done in a chat first |
| Anything confidential under company policy | Copilot, usually | Tenant controls are the whole point |

The pattern: Copilot wins when the context is your Microsoft data, ChatGPT wins when the value is the model's reasoning. Neither wins both.

#### Pricing, and the stacking problem

Copilot Pro and ChatGPT Plus sit at similar per-month prices, around $20 each, while Microsoft 365 Copilot is sold per seat to organizations at a higher rate on top of the 365 license. The individual buyer's trap is paying for both because each covers what the other cannot.

If that is where you have landed, do the arithmetic before renewing: two assistants at around $20 each is roughly $480 a year for one model family in two wrappers. The [savings calculator](https://whizi.io/tools/calculator) will put a number on your actual stack.

The alternative worth knowing about: a multi-model workspace covers the general-assistant half with GPT, Claude, and Gemini together, below the price of either subscription. It will not write inside your Outlook window; nothing but Copilot will. But for everyone whose Office usage is light, it replaces the general half and adds the model choice Copilot never offers. That is the case for [Whizi against Copilot](https://whizi.io/alternatives/copilot) specifically.

#### How to decide in one afternoon

Track one ordinary workday and count two things: how many times you wanted AI inside an Office file with its context, and how many times you wanted AI for something free-standing: a question, a draft, a plan, code.

Mostly the first: Copilot, and stop feeling guilty about ChatGPT headlines. Mostly the second: ChatGPT or a multi-model workspace, and use the free Copilot chat for the occasional Office moment.

Genuinely both, every day: that is the one profile where paying for two products is defensible. For everyone else the split test above settles it in a day.

Key points:

- Copilot runs on OpenAI models, so this is not really a model comparison; it is a product comparison
- The Office integration is the entire case for Copilot, and it is a real case if your work lives in Word, Excel, and Outlook
- ChatGPT wins on model quality control, breadth, and everything that happens outside a Microsoft window

Checklist:

- Name which Copilot you are actually evaluating before comparing anything
- Count one day's AI moments: inside Office files versus free-standing
- Test the free Copilot chat before paying for Pro
- Remember both products run substantially on the same model family
- Price the two-subscription stack annually before renewing both

FAQ:

Q: Is Copilot as good as ChatGPT?
A: On raw capability they draw from the same OpenAI model family, so the ceiling is similar. In practice ChatGPT usually feels stronger because you control the model and the prompt, while Copilot wraps the models in narrower in-app features. Copilot's advantage is context: it can see your files, mail, and meetings, which ChatGPT cannot.

Q: Is Copilot Pro worth it over ChatGPT Plus?
A: Only if the in-app Office features are the thing you will use daily. For general assistant work, ChatGPT Plus is the deeper product at the same price. If your Office usage is occasional, the free Copilot chat plus one good general assistant covers more for less.

Q: Does Copilot use GPT models?
A: Yes, Copilot is substantially built on OpenAI models, with Microsoft adding its own orchestration and, in some features, other models. That is why this choice is mostly about packaging, data access, and admin control rather than model intelligence.

Q: Can I replace both with one subscription?
A: You can replace the general-assistant half. A multi-model workspace like Whizi gives you GPT, Claude, and Gemini in one plan below the price of either subscription. The in-app Office features are the one thing only Microsoft sells, so keep Copilot if that is where your value is.

### DeepSeek vs ChatGPT: cost, quality, and the part people miss

URL: https://whizi.io/resources/deepseek-vs-chatgpt/
Updated: August 2026

Quick answer: DeepSeek is genuinely competitive with ChatGPT on text reasoning, mathematics and code, at roughly a tenth of the API cost. ChatGPT is clearly ahead on multimodal work, tool use, reliability and everything built around the model. The choice is whether you are buying a model or a product.

How DeepSeek compares to ChatGPT on reasoning, coding, and cost, why the answer depends on which DeepSeek you mean, and what the privacy tradeoffs actually are.

#### The short answer

On text reasoning, mathematics, and code, DeepSeek is genuinely competitive with frontier Western models, and it costs a fraction as much to run through an API. That is not hype, it is the reason the industry paid attention when R1 landed and why every provider's pricing has been under pressure since.

On everything that is not the raw model, ChatGPT is ahead and not close. Image and audio understanding, voice, code execution on your files, connectors, custom assistants, the developer tooling, the reliability under load, the documentation, the ecosystem. If you are choosing a daily assistant rather than a model to build on, that gap matters more than the benchmark scores do. The same product-versus-model split shows up in the [Grok vs ChatGPT](https://whizi.io/resources/grok-vs-chatgpt) matchup.

So the honest framing is not "which is better". It is "are you buying a model or a product?" DeepSeek sells a very good model very cheaply. OpenAI sells a product with a very good model inside it. Those are different purchases.

#### First, which DeepSeek are you asking about?

Almost every confused DeepSeek argument comes from two people discussing different things. There are three distinct ways to use it, and they differ on cost, on privacy, and on what you are allowed to do with the output.

| | The app and website | The DeepSeek API | The open weights |
| --- | --- | --- | --- |
| Cost | Free | Very low per token | Your own compute |
| Where your data goes | DeepSeek's servers | DeepSeek's servers | Wherever you run it |
| Can you self-host | No | No | Yes |
| Can you fine-tune | No | No | Yes |
| Topic restrictions | Yes, on politically sensitive subjects | Yes | Reduced, but training-level behaviour remains |
| Good for | Trying it out | Cheap production text and code | Privacy-sensitive or regulated work |

That third column is the one people forget, and it is the most interesting. DeepSeek publishes model weights under permissive licensing, which means the model can be downloaded, run on your own hardware, hosted by a provider in your own jurisdiction, and adapted. No frontier Western lab does that with its best models.

The practical consequence: a team that cannot send data to a Chinese company can still use DeepSeek by running the weights themselves or through a provider that hosts them elsewhere. If you read that a government banned DeepSeek, they almost always mean the app and the hosted API, not the weights.

#### Where DeepSeek is genuinely competitive

**Reasoning and mathematics.** The reasoning line is DeepSeek's strongest work. On hard multi-step problems, competition mathematics, and logic puzzles, it performs in the same band as models that cost far more to run. If you have a problem with a verifiable right answer, it is worth testing there first.

**Code.** Solid at implementation, refactoring, debugging, and explaining unfamiliar code. Not the best at large-scale codebase reasoning, where a very large context window helps more, but strong per dollar by a wide margin.

**Visible reasoning.** The reasoning models show their work in a way that is genuinely useful for review. When the answer is wrong, you can usually see which step went wrong, which is worth more than it sounds when you are checking output rather than trusting it.

**Cost.** This is the headline. Per million tokens, the DeepSeek API has consistently sat around an order of magnitude below flagship Western pricing, sometimes further with caching. For anything running at volume, that is not a small saving, it is the difference between a feature being viable and not.

**Efficiency as a design goal.** The interesting part of DeepSeek's research is architectural: mixture-of-experts routing, sparse attention, and training techniques that get frontier results without frontier budgets. That work has been published, and it has visibly influenced how everyone else prices and builds.

#### Where ChatGPT is clearly ahead

**Multimodal work.** Give ChatGPT a screenshot, a whiteboard photo, a chart, or a scanned invoice and it handles it well. DeepSeek's text models are text models. If your work involves images, this is decisive on its own.

**Tool use and code execution.** ChatGPT can run code against your uploaded data, produce a chart, and hand you the file. That workflow, plus connectors into workplace tools, is a product capability, and no model on its own replaces it.

**Reliability and capacity.** DeepSeek's hosted service has had visible periods of degradation and signup pauses under load. If you are building something customers depend on, uptime history matters as much as quality.

**Ecosystem.** SDKs, well-documented function calling, integrations, community answers to obscure problems, and a hiring pool that already knows the tooling. When you hit a strange edge case at 2am, the odds someone has written about it are much higher.

**Consistency of behaviour.** ChatGPT follows complex formatting and role instructions more reliably across long conversations. DeepSeek drifts more, which shows up quickly in agent-style workloads where the same prompt runs a thousand times.

**Voice, images out, and everything else.** Generation beyond text is a large part of what people pay for and is not really part of DeepSeek's offering.

#### The cost comparison, done properly

If you use a chat interface, this section barely matters: DeepSeek's app is free and ChatGPT's subscription is around $20. Skip ahead.

If you are building anything, the per-token difference compounds fast. Take a realistic support-assistant workload: 2,000 input tokens and 500 output tokens per request, 50,000 requests a month. That is 100 million input tokens and 25 million output tokens. Run that on flagship Western pricing and you are looking at a meaningful monthly bill. Run it on DeepSeek and the same workload can land closer to the cost of a lunch. Even after the price cuts that DeepSeek itself triggered across the industry, the gap has stayed wide.

Three things that change the maths, and they are the ones people miss:

- **Reasoning models generate a lot of tokens you never see.** Extended thinking can be several times the visible output. Cheap per token times many more tokens is not always cheap. Measure total tokens, not price per token.
- **Cache pricing matters more than headline pricing** for anything with a stable system prompt. A large discount on repeated input can beat a smaller sticker price.
- **The cheapest model that passes your evaluation wins.** Not the best model. Build a small test set of real inputs with known good outputs, and run it against both. Most production workloads do not need a frontier model, and you only find that out by measuring.

A sensible pattern for teams: route the bulk of traffic to the cheap model, escalate to the expensive one when confidence is low or the task is high-stakes. That is much easier to build when you are not locked to one vendor's API.

#### Privacy, jurisdiction, and censorship

This deserves to be handled factually rather than as a scare story, because the details determine whether it affects you at all.

**Where data goes.** DeepSeek's hosted service processes and stores data on servers in China, and its privacy policy says so. Chinese law gives authorities broader access to data held domestically than most Western frameworks do. Whether that matters depends entirely on what you are sending. A recipe question is not a risk. Client contracts, patient data, unreleased code, or anything covered by GDPR or a customer data processing agreement are a different matter.

**Regulatory response.** Several governments and regulators have restricted the DeepSeek app on official devices or in their jurisdictions, citing data handling. If your organisation has a procurement or security review process, expect the hosted service to fail it, and expect that decision to be about jurisdiction rather than about model quality.

**Content restrictions.** The hosted service declines or deflects on politically sensitive subjects, particularly ones sensitive within China. Running the open weights reduces this, since the refusal layer is partly in the serving stack, but some behaviour is trained in and does not fully disappear. If your work touches history, geopolitics, or anything comparative about governments, test it before you rely on it.

**The workaround that actually works.** Use the open weights through a provider hosting them in your own jurisdiction, or run them yourself. You keep the cost and quality advantages, and the data-residency question goes away. This is exactly why open weights are strategically important and not just a nice gesture.

For a fuller comparison of how the major providers handle chat data, retention, and training opt-outs, our [AI subscription cost guide](https://whizi.io/resources/ai-subscription-costs) covers the buying side and each vendor's current policy pages are the authoritative source for the rest.

#### Who should use which

| If you are | Use | Because |
| --- | --- | --- |
| A general daily user | ChatGPT | The product around the model is most of the value |
| Cost-sensitive and building at volume | DeepSeek API | Order-of-magnitude cheaper for text and code |
| In a regulated industry | DeepSeek open weights, self-hosted, or a Western model | Data residency is the whole question |
| Doing heavy math or logic | Test both | DeepSeek is genuinely strong here and cheap to try |
| Working with images, audio, or files | ChatGPT | DeepSeek's text models do not cover this |
| Running an agent loop in production | ChatGPT, then optimise | Instruction-following consistency beats price at first |
| Researching or experimenting | Both | The cost of trying DeepSeek is close to zero |

The pattern most teams settle into is not one or the other. It is a cheap capable model for the volume work, a frontier model for the hard cases and anything customer-facing, and a routing rule between them. That is a sensible architecture, and it is why single-vendor lock-in is worth avoiding even when you like the vendor.

If you want to try DeepSeek's reasoning against GPT, Claude, and Gemini without opening four accounts and without sending anything to a hosted Chinese endpoint, [Whizi](https://whizi.io/pricing) includes all of them on one subscription. Put the same prompt through two models, keep the better answer, and see [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side) for how that works in practice.

Key points:

- DeepSeek is competitive on reasoning and code at roughly a tenth of the API cost
- The app, the API, and the open weights are three different products with three different privacy stories
- ChatGPT remains clearly ahead on multimodal work, tool use, and everything around the model

Checklist:

- Decide whether you mean the app, the API, or the open weights, since they differ on every axis
- Check whether your data can legally leave your jurisdiction before using the hosted service
- Measure total tokens including hidden reasoning, not just the price per token
- Test both against a small set of real inputs with known good answers
- Rule DeepSeek out immediately if your work involves images, audio, or voice
- Consider routing cheap traffic to one model and hard cases to another instead of picking one

FAQ:

Q: Is DeepSeek better than ChatGPT?
A: On text reasoning, mathematics, and code, DeepSeek is competitive at a fraction of the API cost. On multimodal work, tool use, reliability, and everything built around the model, ChatGPT is clearly ahead. Which one wins depends on whether you are buying a model or a product.

Q: Is DeepSeek safe to use?
A: The hosted app and API process data on servers in China, which several governments have restricted for official use. For personal, non-sensitive questions that is a minor concern. For client data, regulated data, or proprietary code it is a real one, and the answer is to run the open weights yourself or through a provider in your own jurisdiction.

Q: Is DeepSeek really open source?
A: The model weights are published under permissive licensing, so you can download, run, host, and adapt them. The training data and the full training pipeline are not published, so "open weights" is the accurate term rather than open source in the strict sense. It is still far more open than any frontier Western model.

Q: Why is DeepSeek so much cheaper?
A: Architectural efficiency, mainly. Mixture-of-experts routing activates only part of the model per token, and later releases added sparse attention that cuts the cost of long inputs. The lab has published this research, and the resulting price pressure has pushed the whole industry down.

Q: Can I use DeepSeek without sending data to China?
A: Yes. Because the weights are published, you can run the model on your own hardware or use a provider that hosts it in your region. You keep the cost and quality advantages and the data-residency issue goes away.

### Free ChatGPT alternatives: what’s actually free?

URL: https://whizi.io/resources/free-chatgpt-alternatives/
Updated: August 2026

Quick answer: Free ChatGPT alternatives are real, but the best one depends on your highest-frequency task: writing, coding, research, documents, or images. Free plans cap messages, model access, file uploads, and speed, and privacy terms vary widely. Test candidates with a real prompt from your own week, then compare accuracy, limits, and cleanup time.

Compare free ChatGPT alternatives by use case, limits, privacy tradeoffs, and upgrade signals so you can choose the right AI chatbot without tool sprawl.

#### What "free" really means: limits, privacy, and caps

Free ChatGPT alternatives are real, but "free" rarely means unlimited, private, always-fast access to every best model. In 2026, most free AI chat options are designed as an entry point: enough access to learn the product, handle light tasks, or test a workflow before upgrading. That can be perfectly useful. It just needs to be understood clearly.

The first limit is usually usage. A free AI chatbot may cap messages, restrict peak-time access, throttle speed, or route you to a smaller model when demand is high. The second limit is capability: file uploads, long-context work, image understanding, browsing, memory, projects, collaboration, or access to the newest model family. The third limit is workflow. A tool can feel free until your research notes live in one app, your coding prompts in another, and your writing drafts across five chat histories.

Privacy is the part people skip too quickly. A "chatgpt alternative no sign up" search can be useful for a quick test, but no-sign-up tools deserve extra caution. Before you paste customer data, contracts, source code, financials, medical details, or private strategy notes, check what the tool says about data use, retention, account controls, and whether conversations may be used to improve systems. If the privacy terms are vague, treat the tool as public enough that you would not paste anything sensitive.

Model names also matter. Official OpenAI, Anthropic, and Gemini docs all describe multiple models rather than one universal model. A free product may advertise access to a familiar brand or model family, while the exact model, quota, context size, tool access, and feature set vary by product surface. That is why this guide avoids fake pricing and permanent rankings. Free-plan limits change often. Your best move is to evaluate the current plan against your actual workflow.

Use this free-plan evaluation checklist before you commit your daily work to any free ChatGPT alternative:

- What model or model family am I actually using?
- How many messages, files, images, or documents can I process before a cap appears?
- Does the free plan support the task I care about: writing, coding, research, PDFs, images, or long context?
- Can I export or recover important work later?
- What does the privacy policy say about training, retention, deletion, and sensitive data?
- Are outputs easy to verify, or does the tool hide sources and assumptions?
- Will this still be efficient if I use it every day for a month?

#### Best free picks by use case

There is no single best free AI chatbot for everyone. The right free ChatGPT alternative is the one that handles your highest-frequency task with acceptable limits and a review process you trust. A marketer, developer, student, founder, and operations lead may all choose differently. If you are brand new to AI tools, start with the [best AI apps for beginners](https://whizi.io/resources/best-ai-apps-for-beginners) comparison.

| Use case | What to look for in a free option | Common free-plan limits to check | Best test prompt |
| --- | --- | --- | --- |
| Writing drafts | Clear tone control, rewrites, outlines, email and article structures | Message caps, weaker long-form consistency, limited saved style memory | "Rewrite this for [audience] in [tone]. Keep facts unchanged and list every assumption." |
| Coding help | Strong explanations, debugging steps, test suggestions, code review discipline | Smaller context, limited file handling, hallucinated APIs, no repository awareness | "Given this bug report and code, identify likely causes before proposing a minimal fix." |
| Research | Source discipline, extraction tables, uncertainty notes, synthesis quality | Weak citations, no browsing, limited document upload, shallow summaries | "Use only these sources. Return claim, evidence, confidence, and what to verify." |
| PDF or document summaries | File support, structured extraction, page-aware notes if available | Upload caps, context limits, missing citations, vague summaries | "Extract decisions, risks, dates, owners, and unresolved questions from this document." |
| Image or multimodal work | Image input, OCR-like extraction, structured visual analysis | Limited uploads, lower resolution, no image generation, weak layout understanding | "Turn this image into a table of visible text, objects, issues, and uncertainties." |
| Open-source experimentation | Local control, model variety, community tooling, customization | Setup time, hardware needs, weaker convenience, less polished UX | "Compare two local model outputs on accuracy, speed, and edit time." |

For free AI for writing, test whether the model preserves your facts and improves specificity, and see the [ChatGPT alternatives for writing](https://whizi.io/resources/chatgpt-alternatives-for-writing) comparison for the tools that hold a voice best. The danger is a smooth sentence that quietly changes meaning. Give the model your audience, goal, source notes, voice examples, and required format. Then ask it to list the edits it made.

For free AI for coding, avoid treating any chatbot as an automatic patch machine. The useful workflow is repro first, then diagnosis, then smallest fix, then tests. Free models can explain code, draft unit tests, compare approaches, and spot edge cases. They can also invent libraries, misunderstand constraints, or over-refactor.

For free AI for research, the core question is traceability. A good answer separates what the source says from what the model infers. Ask for tables, evidence, confidence levels, and open questions. If a free AI chat online tool gives confident summaries without sources, use it for brainstorming, not final research.

Open-source ChatGPT alternatives are worth considering when you care about local control, customization, or experimentation. Meta's Llama family is the usual starting point, and [what Llama 3 is in plain English](https://whizi.io/resources/llama-3) explains what downloading a model actually gets you. They are not automatically easier. You may need to manage hardware, model downloads, inference speed, updates, and interface quality. For technical users, that tradeoff can be worth it. For a team trying to ship daily work, the setup burden may outweigh the subscription savings.

#### When paid plans are worth it

Paid AI plans are worth considering when free limits interrupt the work itself. If you only ask a few lightweight questions each week, a free plan may be enough. If AI is part of your writing pipeline, code review loop, research workflow, support process, or document analysis stack, free-plan friction can become the real cost.

The upgrade signal is not "I want the newest model." The upgrade signal is "I keep losing time to caps, switching, missing features, or cleanup." A founder may hit document upload limits during market research. A developer may hit context limits when a bug crosses multiple files. A marketer may spend more time fixing generic copy than drafting from scratch.

Use this decision rule: stay free when the task is low-risk, low-volume, easy to verify, and not blocked by caps. Upgrade when the task is frequent, revenue-adjacent, private, collaborative, long-context, multimodal, or expensive to redo. The point is not to pay for AI because it feels advanced. The point is to remove bottlenecks from work you already know is valuable.

Before you upgrade any single tool, define the paid feature you are buying: higher volume, better model access, longer context, file upload, faster responses, team controls, or stronger privacy settings. If you cannot name the bottleneck, wait. If you can, compare the paid option against a consolidated plan.

A simple upgrade worksheet helps keep the decision honest:

| Question | Stay free if... | Consider paid if... |
| --- | --- | --- |
| Volume | You rarely hit caps | Caps interrupt weekly work |
| Quality | Outputs are good enough after quick edits | Cleanup time is becoming the real cost |
| Context | Tasks fit in short prompts | You need long docs, files, or multi-step history |
| Risk | Work is casual or easy to verify | Work affects customers, revenue, code, or strategy |
| Privacy | You only paste public information | You need stronger controls for private work |
| Workflow | One tool is enough | You are juggling multiple free products |

#### Consolidate tools vs stacking subscriptions

The hidden cost of free ChatGPT alternatives is sprawl. First you use one free tool for writing, another for coding, another for research, another for PDFs, and another for images. Then one hits a cap, one loses a chat, one changes its free tier, and one produces a better answer but only for a narrow task. You still have not paid much money, but you are paying with attention.

Stacking free tools makes sense during exploration. It is a cheap way to learn what different models are good at. But once you know your core workflows, comparison should become structured. Run the same prompt across models, score the outputs, and keep the winner for that task.

Here is a practical consolidation workflow you can use this week. First, choose three real tasks: one writing task, one coding or analysis task, and one research task. Second, write one strong prompt for each task with context, constraints, and output format. Third, run each prompt in the free tools you are considering. Fourth, score the answers on accuracy, usefulness, format compliance, edit time, and verification quality. Fifth, decide whether the winning tools justify separate accounts or whether a consolidated workspace is cleaner.

Whizi is built around that second path. Instead of treating free vs paid as a moral victory or loss, Whizi helps you compare model outputs in one place, keep reusable prompts, and reduce the sprawl that builds when every task sends you to a different AI app. If AI is already part of your daily work, compare that time cost against a plan designed to keep the workflow together.

For a broader paid-and-free shortlist, read the main [ChatGPT alternatives guide](https://whizi.io/resources/chatgpt-alternatives), compare the consolidation options in [the best all-in-one AI platforms](https://whizi.io/resources/best-all-in-one-ai-platforms), or jump straight to the [best ChatGPT alternative with Claude and Gemini built in](https://whizi.io/alternatives/chatgpt) if you already know you want more than one model. When you are ready to stop bouncing between tabs, create your account at [register](https://whizi.io/register) or compare options on the [pricing page](https://whizi.io/pricing).

Key points:

- Free AI tools are useful for testing, light work, and learning, but limits can appear in usage caps, model access, speed, file uploads, privacy terms, and history retention.
- The best free AI chatbot depends on the job: writing drafts, debugging code, summarizing sources, analyzing documents, or trying open-source models.
- If you keep adding free tools to dodge limits, a consolidated workspace can become cleaner than juggling separate accounts, tabs, and upgrade prompts.

Checklist:

- Check message caps, model access, file limits, speed, and export options before relying on a free AI tool.
- Do not paste sensitive data into no-sign-up tools unless you have reviewed their privacy and retention terms.
- Test free AI tools with real writing, coding, research, document, or image tasks instead of toy prompts.
- Upgrade only when a specific limit is costing you time, quality, privacy, or workflow consistency.
- Consolidate when managing separate free tools becomes more expensive than one focused workspace.

FAQ:

Q: What is the best free ChatGPT alternative?
A: The best free ChatGPT alternative depends on your task. Test options with real writing, coding, research, document, or image prompts, then compare accuracy, limits, privacy terms, and cleanup time.

Q: Are free AI chatbots really free?
A: Many are free for light use, but limits often apply. Check message caps, model access, file uploads, speed, data controls, export options, and whether advanced features require a paid plan.

Q: Can I use free AI tools for work?
A: Yes, for low-risk tasks and public information. Be careful with private customer data, confidential documents, source code, contracts, financials, or anything governed by company policy.

Q: When should I stop using only free AI tools?
A: Consider upgrading or consolidating when caps, scattered histories, missing features, or verification problems slow down work you rely on every week.

### What is Google Gemini AI and How to Use It (A Beginner's Guide)

URL: https://whizi.io/resources/gemini/
Updated: August 2026

Quick answer: Google Gemini is Google's AI assistant. You type a question in ordinary English and it answers. Two things set it apart: it is tied to Google Search, so it can look up what happened last week and show you the links, and it can read a 300 page report in one go.

Discover what Google Gemini AI is, how it works, and practical ways beginners can use it to save time, brainstorm ideas, and summarize long documents.

#### What Gemini is, and what makes it different

Gemini is Google's AI assistant. You type a question in ordinary English and it answers. If you have used ChatGPT, this will feel familiar within about ten seconds.

Two things genuinely set it apart, and both are more useful than they sound.

**It knows what is happening now.** Because Google built it, it is closely tied to Google Search. Ask about something from last week and it can actually look, give you links, and tell you where the answer came from. Most assistants answer from what they learned during training, which means a confident answer about last month that is simply out of date.

**It can read enormous amounts at once.** This is the underrated one. Every assistant has a limit on how much text it can hold in mind during a conversation, and Gemini's is much higher than most. A 300 page report, a full contract set, or a semester of lecture slides can go in as whole files.

#### Why "reads more at once" changes what you can do

When a document is bigger than an assistant can hold, it has to work through it in pieces. That is fine for a question about one section and unreliable for a question about the whole thing.

With everything genuinely in view, these work:

- `Do any two sections of this contract contradict each other?`
- `Across these fifteen documents, where do the findings disagree, and why?`
- `Is this term defined anywhere in the document, and does the definition match how it is used later?`
- `What is in the detail that the executive summary does not mention?`

None of those work well when a tool is reading a chunk at a time, because each one needs the whole document present simultaneously. If you regularly deal with long PDFs, reports, or transcripts, this single capability is likely to matter more to you than anything else about the model.

#### Prompts worth copying

**Reading a long document**

`Read the attached document and tell me: the five things that matter most, anything that creates an obligation or a cost for me, anything left vague, and anything important that the summary section leaves out. Quote the exact text and give the page for each point.`

Always ask for the quote. It takes seconds to check and it is the only reliable way to know the answer came from your document rather than from general knowledge about documents like yours.

**Researching something current**

`Find out what changed about [topic] in the last six months. Give me a source link and a date for each finding. Where sources disagree, show both rather than picking one. Say clearly if you cannot verify something.`

**Comparing documents**

`Compare the two attached files. Give me a table of every meaningful difference: topic, what the first says, what the second says. Ignore formatting differences. List separately anything present in one and missing from the other.`

**Understanding an image**

`Here is a photo of [thing]. Describe exactly what you can see before interpreting anything. Then tell me [your question].` Asking for the description first catches misreadings before they become a wrong answer.

#### What it is not the best at

Being honest here saves you disappointment later.

**Writing that a person reads end to end.** Gemini writes competently. Claude writes better, and for emails, messages, and anything with a tone requirement the difference is noticeable within one prompt.

**Rigid formatting at scale.** If you need two hundred rows in exactly the same shape, ChatGPT holds a strict format more reliably.

That is not a weakness so much as a specialisation. Gemini reads and researches; other models write and structure. Most real tasks involve both halves, which is why the useful pattern is to research in Gemini and then hand the result to something else.

| What you are doing | Best choice | Why |
| --- | --- | --- |
| Anything from last week | Gemini | Tied to Google Search, so it can look and give you the links |
| A 300 page report or a full contract set | Gemini | Holds far more text at once than most assistants |
| A question spanning fifteen documents | Gemini | Everything is present at once rather than read in pieces |
| Reading a photo, chart, or screenshot | Gemini | Handles images alongside text |
| Writing a person reads end to end | Claude | Writes better, and the difference shows within one prompt |
| Two hundred rows in exactly the same shape | ChatGPT | Holds a strict format more reliably |

#### What it costs

There is a free version and it is genuinely capable, including web access, with limits on how much you can use the more powerful models.

The paid plan is around $20 a month, and it comes bundled as Google One AI Premium with 2TB of cloud storage. That bundle deserves an honest look. If you would buy 2TB of Google storage anyway, because you have a big photo library or a household on Drive, the AI is arriving cheaply and this is good value. If you are on the free storage tier using a fraction of it, part of your subscription is buying space you will never fill.

The other thing the paid plan adds is Gemini inside Gmail, Docs, and Drive. If your work lives in those, that is a real advantage no external tool can copy.

And the alternative worth knowing: Whizi includes the Gemini models alongside ChatGPT and Claude for less than Google's plan costs, without the storage. See [Whizi vs Gemini Advanced](https://whizi.io/vs/whizi-vs-gemini-advanced) for the direct comparison.

#### Things to be careful about

**Web access does not mean always correct.** Having sources reduces invented answers substantially. It does not remove the harder error, which is a real source attached to a claim it does not actually support. Open the links.

**It does not always search.** Depending on how you ask, it may answer from memory instead of looking. For anything current, check that it actually cited something.

**It agrees too readily.** Push back and it will often abandon a correct answer. Ask what it thinks before saying what you think.

**Long does not mean thorough.** Ask a very long document for "everything important" and you get a flat list. Ask specific questions and you get useful answers.

**Be careful what you upload.** Not the place for other people's personal information or anything confidential your employer has a policy about. Removing names rarely affects the analysis.

Key points:

- The two things Gemini is genuinely best at, and why they matter more than they sound
- Copy-ready prompts for documents, research, and images
- What the paid plan actually includes, including the storage nobody mentions

Checklist:

- Use Gemini for anything recent, and open the links it gives you
- Upload long documents as files rather than pasting extracts
- Always ask for the exact quote supporting an answer
- Ask specific questions of a long document rather than for a summary
- Use a different model for writing that someone will read end to end
- Check whether you would pay for the bundled 2TB of storage on its own
- Ask for a description before interpretation when working with images

FAQ:

Q: Do I need to know how to code to use Google Gemini?
A: No, none at all. You type in plain English exactly as you would message a colleague, and there is nothing to install or configure. The only real skill is being specific: saying who an answer is for, what format you want, and what to leave out will improve your results far more than any technical knowledge.

Q: Is Google Gemini free to use?
A: There is a free version that is genuinely capable, including web access, with limits on how heavily you can use the more powerful models. The paid plan is around $20 a month and arrives bundled with 2TB of Google One storage plus integration into Gmail and Docs. Whether that is good value depends mostly on whether you would buy the storage anyway.

Q: What is Gemini best at compared to ChatGPT?
A: Two things clearly. Anything recent, because it can search and give you links you can check, while most assistants answer from a fixed training date. And long documents, since it can hold far more text at once, which makes questions spanning a whole report or a set of files actually work. ChatGPT is better at structure and formatted output, and Claude is better at writing.

Q: Can Gemini read PDFs and images?
A: Yes to both, and long PDFs are where it is strongest, since you can upload whole files rather than extracts. For images it can read photos, charts, screenshots, and diagrams. A useful habit with images is to ask it to describe exactly what it sees before interpreting anything, which catches misreadings before they turn into a confident wrong answer.

Q: Should I use Gemini or something else?
A: Both, if you can, since their strengths barely overlap. Gemini for reading and research, Claude for writing, ChatGPT for structure and formatted output. The common pattern is to research in Gemini and hand the findings to another model to write up, which works better than either alone and is one step instead of two tabs when both live in the same conversation.

### Gemini vs ChatGPT: which is better for multimodal work?

URL: https://whizi.io/resources/gemini-vs-chatgpt/
Updated: August 2026

Quick answer: Gemini is the stronger first test when the task centers on long context and multimodal input: large PDFs, transcripts, research packets, screenshots, and charts. ChatGPT is the stronger first stop for drafting, planning, editing, and everyday productivity. Both document structured outputs, so run the same file through each and compare.

Compare Gemini vs ChatGPT for images, PDFs, long documents, structured outputs, and daily productivity workflows without picking a model on hype.

#### Multimodal capability

The short version: Gemini vs ChatGPT is not a simple winner-takes-all comparison. Both model families can be useful for text, images, files, reasoning, and productivity. The better question is narrower: "Which model is the better fit for this exact input and output?" For multimodal work, that means judging how well the assistant handles text plus images, screenshots, charts, PDFs, tables, and long background material.

Gemini has a clear advantage in the way Google positions the Gemini API around multimodal and long-context work. Google says Gemini models can understand text, video, audio, and images, and its long-context guide emphasizes use cases where you provide a large body of relevant material up front. That makes Gemini especially worth testing when the job is "read this large packet and answer questions from it" or "combine visual evidence with written context."

ChatGPT remains very strong for practical productivity: drafting, analysis, planning, code help, data cleanup, and turning messy notes into useful outputs. OpenAI documents a broad model lineup, including models that support text and image inputs, structured outputs, tools, and reasoning-oriented workflows. In practice, ChatGPT is often the smoother first stop when the task is primarily language, strategy, editing, or step-by-step execution.

The mistake is treating multimodal as a checkbox. "Can accept images" is a very different claim from "can reliably pull details out of a screenshot, connect them to a PDF, and give you a table you can use." For anything that matters, run the same input through both and score the results on accuracy, missing details, format control, and how much cleanup is left. [Comparing models side by side](https://whizi.io/docs/compare-models-side-by-side) shows how to do that with one prompt instead of two open tabs.

#### Long-document workflows

Long context is where Gemini deserves special attention. Google states that many Gemini models include context windows of 1 million or more tokens, and its long-context documentation gives concrete examples like large codebases, many documents, long transcripts, and extended reference material. That does not mean every long prompt should be dumped into a model without thinking. It means Gemini is a strong candidate when the core problem is keeping a large amount of source material available at once.

Use Gemini first when you have a dense document packet: an investor memo, legal summary, research report, product requirements document, competitive teardown, or a folder of notes you want analyzed together. Use ChatGPT first when the document task quickly turns into a communication task: writing the recommendation, creating an executive summary, building a launch plan, or turning findings into client-ready language.

A practical PDF workflow looks like this:

1. Upload the PDF, report, or document packet.
2. Ask for a source-grounded inventory: sections, tables, charts, dates, claims, and unanswered questions.
3. Ask the model to extract only what is explicitly present, with page or section references when available.
4. Ask a second model to review the extraction for missing items or overconfident claims.
5. Convert the final answer into the format you need: briefing memo, spreadsheet-style table, decision doc, FAQ, or task list.
6. Manually verify any number, quote, legal detail, medical detail, or financial claim before using it.

Here is a prompt you can reuse: "Analyze this PDF for a business decision. First create a document map. Then extract the claims, numbers, risks, and recommendations. Do not infer missing facts. Put uncertain items in a separate section. Finish with a decision table that includes evidence, confidence, and what I should verify manually."

For image work, use a similar process: "Review this screenshot or image. Describe only visible evidence first. Then extract structured information into a table. Then list possible interpretations separately from confirmed observations. Flag anything too small, cropped, blurry, or ambiguous to read." This helps prevent the model from blending visual evidence with assumptions.

#### Structured outputs

Structured outputs matter when the answer needs to become data. A nice paragraph is not enough if you are extracting invoice fields, tagging customer feedback, turning a PDF into a CSV-like table, classifying research notes, or generating JSON for an app. This is one of the cleanest ways to compare Gemini API vs ChatGPT API use cases.

Google documents Gemini structured outputs as a way to make model responses follow a provided JSON Schema, with use cases including data extraction, classification, and agentic workflows. Google also notes that structured outputs do not guarantee the values are semantically correct, so application-level validation still matters. That warning is important: valid JSON can still contain a wrong number.

OpenAI documents Structured Outputs for ensuring responses follow a supplied JSON Schema, with support in current large language models and guidance on function calling versus response formatting. OpenAI also distinguishes Structured Outputs from basic JSON mode, where output may be valid JSON without necessarily matching the schema you intended.

For non-developers, the same principle applies in plain language: tell the model exactly what fields you want. For example: "Return a table with columns for claim, source location, evidence quote, risk level, owner, and follow-up question. If a field is missing, write Not found." Whether you use Gemini, ChatGPT, or both, the goal is predictable output you can review quickly.

#### Best by scenario

The best AI for images and text depends on the workflow. Use this scenario table as a starting point, then test with your real material.

| Scenario | Start with | Why | Verification step |
| --- | --- | --- | --- |
| Long PDF or research packet | Gemini | Long-context workflows are a major Gemini strength, especially when the source material is large | Ask ChatGPT to critique the summary and list missing evidence |
| Screenshot, chart, or image plus written instructions | Gemini or ChatGPT | Both are worth testing; quality depends on image clarity and the requested output | Require a visible-evidence section before interpretation |
| Executive memo from messy notes | ChatGPT | Strong fit for structure, tone, prioritization, and polished drafting | Ask Gemini to check whether the memo missed document details |
| Data extraction into JSON or table | Both | Gemini and OpenAI both document structured output workflows | Validate fields, enums, dates, and numbers before using the result |
| Product requirements or specs | Gemini first for large context, ChatGPT for final plan | Gemini can process more source material; ChatGPT can shape the execution plan | Compare assumptions and ask for test cases or acceptance criteria |
| Everyday productivity | ChatGPT | Often quick for emails, planning, rewriting, brainstorming, and task breakdowns | Use Gemini when the task includes large docs or visual context |

The most reliable workflow is not model loyalty. It is model comparison. Run the same prompt in Gemini and ChatGPT, then choose the answer that is more accurate, more useful, and easier to verify.

A simple scoring rubric: give each output 1 to 5 points for source accuracy, coverage, structure, actionability, and cleanup required. If the difference is small, use the model that is faster or cheaper for that job. If the difference is large, save the winning prompt as your default workflow.

#### CTA section

The cleanest way to decide between Gemini vs ChatGPT is to compare them on the same task inside one workspace. Pick one PDF, screenshot, chart, or document packet from your actual week. Run the prompt in both models. Look for what each model notices, misses, invents, and formats well.

Try this inside Whizi: upload the same PDF or image task, run it through Gemini and ChatGPT, then turn the winning output into a reusable workflow. You do not need to decide that one model is your permanent favorite. You need a repeatable way to pick the right model for the job.

For a broader model decision framework, read [ChatGPT vs Claude vs Gemini](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini). When you are ready to compare plans, see [Whizi pricing](https://whizi.io/pricing), or create your account at [Whizi registration](https://whizi.io/register).

Key points:

- Use Gemini when long-context, document-heavy, and multimodal inputs are central to the task
- Use ChatGPT when you want strong drafting, planning, tool workflows, and structured reasoning across everyday work
- Run the same PDF or image prompt in both models before deciding which answer is good enough to use

Checklist:

- Use Gemini first for large document packets, long-context analysis, and multimodal source review
- Use ChatGPT first for drafting, planning, rewriting, and turning analysis into polished productivity outputs
- Ask for visible evidence before interpretation when analyzing images or screenshots
- Use structured fields when extracting data from PDFs, charts, invoices, research notes, or customer feedback
- Compare the same prompt across models before adopting a workflow for repeat use

FAQ:

Q: Is Gemini better than ChatGPT for documents?
A: Gemini is especially worth testing for long-document and large-context workflows because Google emphasizes very large context windows in the Gemini API. ChatGPT can still be excellent for summarizing, rewriting, and turning findings into polished work. The best answer is to test both on the same document.

Q: Is ChatGPT or Gemini better for images?
A: Both can be useful for image-and-text tasks. For reliable work, ask the model to separate visible evidence from interpretation, then compare outputs. Image clarity, crop quality, and prompt specificity often matter as much as the model choice.

Q: Which is better for structured outputs?
A: Both Gemini and OpenAI document structured output capabilities. Use structured outputs when you need JSON, tables, classifications, or extracted fields, and always validate the values before using them in production or decision-making.

### Gemini vs Grok: two different bets on knowing things first

URL: https://whizi.io/resources/gemini-vs-grok/
Updated: August 2026

Quick answer: Gemini is the stronger choice for the work itself: documents, images, long context, and anything inside Gmail, Drive, or Docs. Grok is stronger on the last few hours of public conversation, because it pulls from X posts as they happen. Very few tasks are genuinely contested between the two.

A practical Gemini vs Grok comparison: Google's ecosystem and multimodal depth against live X data and an unhedged voice, with a task table and a one-afternoon test.

#### The short answer

Use **Gemini** for the work itself: documents, spreadsheets, images, long reports, video, and anything touching Gmail, Drive, or Docs. Its multimodal range and long context make it the stronger general workhorse, and Google's search grounding covers ordinary freshness well.

Use **Grok** when the question is about the last few hours of public conversation: how a launch is landing, sentiment around a company, a developing story where the signal is in posts rather than articles. That firehose is Grok's structural advantage, and Gemini does not have an equivalent.

The interesting thing about this matchup is that both are selling recency, from different wells. Google indexes the web faster than anyone; X hosts the conversation before the web writes it up. Which well matters depends entirely on the question you ask most.

#### Where Gemini is genuinely stronger

**Multimodal work as a first-class feature.** Photos, screenshots, PDFs, charts, audio, and video in one model, and generation on the image side. If your inputs are ever not plain text, Gemini handles more of them, more natively.

**Long context that gets used.** Gemini's 1M-token windows made whole-book and whole-codebase prompts routine, and the [context window comparison](https://whizi.io/resources/context-window-comparison) shows it still setting the pace. For feeding a model everything and asking questions, it is the default choice.

**The Google ecosystem.** Summarize this Gmail thread, pull from this Drive folder, draft in Docs. For anyone whose work lives in Google's apps, the integration story mirrors what Copilot offers Office users, with a stronger underlying model family.

**Grounded search as a default.** Answers can carry Google-backed citations, which for verifiable factual work is a steadier foundation than any model's memory.

#### Where Grok is genuinely stronger

**The live firehose.** Grok pulls from X posts as they happen. Earnings reactions, developer sentiment on a release, what is actually being said in a controversy right now: for these queries Grok answers something Gemini structurally cannot, because the indexed web runs hours behind the conversation.

**An unhedged voice.** Grok commits to positions, makes jokes, and skips the safety preamble. For competitive analysis, satire, and questions where you want a take rather than a survey, it reads like a sharp colleague instead of a careful institution.

**Fewer refusals on legitimate edge cases.** Blunt questions get blunt answers. The same trait has produced Grok's public incidents, so the pairing rule stands: enjoy the directness, proofread anything that leaves your desk.

**Reasoning per subscription dollar.** Recent Grok releases post strong numbers on hard reasoning benchmarks, and for X Premium+ subscribers the model is effectively bundled, which makes its price hard to argue with for people already paying for X.

#### Head to head by task

| Task | Better choice | Why |
| --- | --- | --- |
| Long documents and PDFs | Gemini | Context length plus faithful document handling |
| Photos, charts, video inputs | Gemini | Widest multimodal range |
| Breaking news and live sentiment | Grok | The firehose is the feature |
| Verifiable factual research | Gemini | Search grounding with citations |
| Opinionated first drafts and satire | Grok | The voice is the feature |
| Gmail, Drive, Docs workflows | Gemini | Native integration |
| Hard math and logic | Split | Both post strong reasoning numbers; test on your own problems |
| Client-facing deliverables | Gemini | Predictability wins where a colleague reads it |

The split is unusually clean: Gemini owns the documents-and-data day, Grok owns the what-is-happening-right-now hour. Very few tasks are genuinely contested between them. If you are weighing either against OpenAI instead, the [Gemini vs ChatGPT](https://whizi.io/resources/gemini-vs-chatgpt) and [Grok vs ChatGPT](https://whizi.io/resources/grok-vs-chatgpt) comparisons cover those matchups directly.

#### Pricing and access

Both sit at the standard flagship price point, around $20 a month, with heavier tiers above. Grok has the bundle quirk: X Premium+ subscribers effectively get it included, which for that group settles the price question outright. Gemini's subscription increasingly bundles into Google One tiers, which does something similar for Google-ecosystem households.

The stacking trap applies here as everywhere: this pairing is popular precisely because the two cover different hours of the day, and two subscriptions at around $20 each is roughly $480 a year. If both are on your card, run your stack through the [savings calculator](https://whizi.io/tools/calculator) before the next renewal.

In Whizi, Gemini models are available across plans and Grok models on the higher tiers, in one conversation. The one thing that does not travel is Grok's live X access, which stays with the native product; what travels is the reasoning and most of the voice.

#### How to decide in one afternoon

Write down your last ten AI questions. Sort them into two piles: "about my documents, data, or work product" and "about what is happening right now".

A lopsided first pile says Gemini, and the free tier will confirm it. A lopsided second pile says Grok, especially if you already pay for X. A genuine mix is the multi-model case: route each question to the model built for it, and let the two check each other on the questions where recency and depth both matter.

Whichever way it lands, resist deciding from launch benchmarks. Both families ship majors every few months, the [release timeline](https://whizi.io/resources/ai-model-timeline) shows how quickly leads change hands, and your ten real questions are a better benchmark than any chart.

Key points:

- Both sell freshness, from opposite sources: Google's indexed web versus X's live conversation
- Gemini wins the working day: documents, images, long context, and the Google ecosystem
- Grok wins the last four hours: sentiment, breaking situations, and an unhedged read on them

Checklist:

- Sort your last ten AI questions: work product versus what-is-happening-now
- Test Gemini free on one long document before paying for anything
- If you pay for X Premium+, price Grok as free in the comparison
- Proofread Grok output that leaves your desk; enjoy the voice internally
- Recheck the matchup when either ships a major; leads change hands in weeks

FAQ:

Q: Is Grok better than Gemini?
A: For live public conversation and unhedged takes, yes: Grok pulls from X posts as they happen, which Gemini structurally cannot. For documents, multimodal inputs, long context, and Google-ecosystem work, Gemini is clearly ahead. They are strong at different hours of the day.

Q: Which is better for research, Gemini or Grok?
A: Depends on the research. Verifiable factual work favors Gemini's search grounding and citations. Sentiment, reception, and developing stories favor Grok's live X access. Serious research often needs both: Grok for what is being said, Gemini for what is documented.

Q: Do Gemini and Grok cost the same?
A: Both flagship subscriptions sit around $20 a month, but the bundles change the math: Grok is effectively included with X Premium+, and Gemini increasingly bundles into Google One tiers. Price the bundle you already pay for before adding either as a standalone.

Q: Can I use Gemini and Grok together without two subscriptions?
A: In Whizi, yes: Gemini models across plans and Grok models on the higher tiers, in one conversation. You get both models' reasoning side by side; Grok's live X feed is the one feature that stays exclusive to the native product.

### Grok vs ChatGPT: what each one is actually good at

URL: https://whizi.io/resources/grok-vs-chatgpt/
Updated: August 2026

Quick answer: Use Grok for questions about what is happening right now, since it draws on posts from X as they are published. Use ChatGPT for work that has to be dependable: client deliverables, documents, integrations, and code you will ship. Both are strong general models, so the differences sit at the edges.

A practical Grok vs ChatGPT comparison covering real-time data, reasoning quality, tone and guardrails, pricing, and which one fits which kind of work.

#### The short answer

Use Grok when the question is about what is happening right now, especially anything moving through public conversation: a breaking story, a product launch reaction, sentiment on a company, a developing situation where the useful information is in posts rather than articles. Nothing else does that as directly, because Grok is wired into X.

Use ChatGPT when the job needs to be dependable: client deliverables, structured workflows, connected tools, documents, code you will ship, and anything where a colleague might read the output. The ecosystem around ChatGPT is years deeper, and for most professional work that ecosystem matters more than a few benchmark points.

Both are strong general models. If you put a normal question to each, you will usually get two good answers, and which one you prefer will come down to taste. The interesting differences sit at the edges, so that is where this comparison spends its time.

#### Where Grok is genuinely different

**Live access to public conversation.** This is the real feature, and it is not a small one. Ask about a company's earnings reaction, a controversy from four hours ago, or how developers are responding to a framework release, and Grok can pull from posts as they happen. ChatGPT browses the web, but the web indexes slowly and articles get written hours later. For anything where the signal is in real-time chatter, Grok answers a question the others structurally cannot.

One caveat worth knowing before you build a workflow on it: that live feed is a feature of xAI's own apps and API search options, not something that automatically travels with the model. If you use Grok through a third-party workspace, you are getting the model's reasoning, which may or may not come with live search attached depending on how that product is configured. Check before you rely on it for time-sensitive work.

**A push on reasoning.** xAI has spent heavily on test-time compute, the approach where a model thinks longer on hard problems before answering. Recent Grok releases have posted strong numbers on the harder reasoning and math benchmarks, and the heavy tiers run multiple reasoning passes in parallel. On genuinely difficult analytical problems it is competitive with anything else available.

**A different default personality.** Grok is deliberately less hedged. It will give a direct opinion, make a joke, and skip the disclaimers that other assistants attach to anything remotely contested. Plenty of people find this refreshing after years of being told that a topic is complex and they should consult a professional.

**Fewer refusals on edgy but legitimate work.** Comedy writing, fiction involving conflict, blunt competitive analysis, and frank medical or legal questions all tend to get answered rather than deflected. If you have been irritated by an assistant refusing something obviously reasonable, this is the difference you will feel first.

#### Where ChatGPT is still ahead

**Ecosystem depth.** Custom GPTs, connectors into common workplace tools, a mature developer API with well-documented function calling, a huge base of tutorials and community answers, and integrations built into products you already use. When something breaks at 11pm, someone has written about your exact problem. That is worth more than it sounds.

**Predictability.** ChatGPT's output is more consistent in tone across sessions, which matters when the text goes to a client, a customer, or a manager. Grok's personality is a feature until it turns up in a deliverable you did not proofread carefully.

**Document and file work.** Uploading a stack of PDFs, spreadsheets, and images and asking for analysis across them is a well-worn path in ChatGPT, with the code execution environment doing real work on your data. Grok handles files, but the workflow is younger.

**Enterprise readiness.** Data controls, admin tooling, compliance documentation, and procurement paperwork are all more established. If your company needs a signed data processing agreement before you can use a tool, that difference decides it for you.

**Reliability of the record.** Grok has had several public incidents where the model produced offensive or badly wrong output at scale. xAI patched each one, and every major lab has had bad days, but the pattern is worth weighing if the tool will speak on behalf of your business.

#### Head to head by task

Benchmarks change with every release, and by the time you read a leaderboard it is usually out of date. Task fit is more stable, so use this as your starting point rather than a scoreboard.

| Task | Better starting point | Why |
| --- | --- | --- |
| Breaking news and live sentiment | Grok | Direct access to posts as they happen, not indexed articles |
| Client-facing writing | ChatGPT | More consistent register, less chance of an unwanted joke |
| Hard reasoning and math | Roughly tied | Both run extended thinking modes, test on your own problems |
| Everyday coding help | Roughly tied | Grok's fast coding models are strong, ChatGPT has better tooling around them |
| Long document analysis | ChatGPT | More mature file handling, though Gemini beats both here |
| Sourced research with citations | Neither | A dedicated research tool does this better, see [Perplexity vs ChatGPT](https://whizi.io/resources/perplexity-vs-chatgpt) |
| Comedy, satire, blunt takes | Grok | Fewer refusals, more willing to commit to an angle |
| Automations and integrations | ChatGPT | Larger connector ecosystem and more mature API patterns |
| Regulated or enterprise settings | ChatGPT | Better established compliance and admin story |

One pattern shows up repeatedly in real use: people reach for Grok when they want to know something, and ChatGPT when they want to make something. That is not a rule, but it predicts behaviour better than benchmark tables do.

#### Pricing and how you actually get access

Both have usable free tiers with daily caps, and both put their best modes behind a subscription. The access paths differ in a way that catches people out.

ChatGPT sells you a plan directly: a free tier, Plus at around $20 a month, a much heavier Pro tier, and Team seats for small groups. Simple to reason about. Our [is ChatGPT Plus worth it](https://whizi.io/resources/is-chatgpt-plus-worth-it) breakdown covers whether the upgrade is justified.

Grok is sold in several overlapping ways: free with limits, a dedicated subscription usually priced above ChatGPT Plus, access bundled into X's premium tiers, and a very expensive top tier for the heaviest reasoning modes. Which one you want depends on whether you already pay for X, since the bundle can make Grok effectively free for someone who was subscribing anyway.

Verify current pricing on both sites before deciding, because these tiers have been restructured more than once. The important budgeting point is not the sticker price, it is that these are two more monthly charges in a category where people already have two or three. If ChatGPT Plus, Claude Pro, and a Grok subscription all end up on the same card, run the [AI subscription savings calculator](https://whizi.io/tools/calculator) and see the total written down.

#### The guardrail question, answered honestly

This is the part of the comparison most articles either avoid or turn into a political argument. It is worth being plain about, because it genuinely affects which tool fits which job.

Grok is tuned to refuse less. That is a real advantage when you are writing satire, analysing a hostile competitor, asking a blunt question about a medical symptom, or writing fiction where characters behave badly. Assistants that treat every sharp edge as a risk are frustrating to work with, and Grok is a relief in those moments.

The same tuning is a liability when the output is unsupervised or public. A model that will commit to a strong opinion will sometimes commit to a wrong one, in your brand voice, in front of your customers. If you are putting AI text in front of an audience without a human reading it first, the more conservative model is the safer default, and that is true regardless of which lab you prefer.

The practical approach most people land on: use the less hedged model for drafting, exploration, and anything private, and the more predictable one for anything with your name on it. That is only awkward if switching between them means switching between subscriptions and browser tabs.

#### How to decide in one afternoon

Pick four prompts from your actual work. Make one of them time-sensitive, something where the answer changed this week. Make one of them a hard analytical problem where you already know the right answer. Make one a piece of writing you would send to someone else. Make the last one whatever you do most often.

Run all four through both models. Score them on three things only: was it correct, how much editing did it need, and did it tell you something you did not already know. Ignore how impressive the response looked, since both models are good at looking impressive.

Most people find the time-sensitive prompt goes to Grok, the deliverable goes to ChatGPT, and the other two are close enough that either would do. Which is a slightly awkward conclusion, because it means the honest recommendation is often "both", and paying two subscriptions for occasional access to one strength each is exactly the spending pattern nobody plans for.

That is the case for a multi-model workspace. [Whizi](https://whizi.io/pricing) gives you Grok, GPT, Claude, Gemini, and DeepSeek behind one subscription, so you can send the same prompt to two of them and keep the better answer without opening a second account. If you want the direct comparison against your current plan, start with [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus), or read [how to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) for the workflow that makes it worth doing.

Key points:

- Grok's real advantage is live access to what people are posting right now, not raw model quality
- ChatGPT still wins on ecosystem, integrations, and predictability for client-facing work
- The tone and guardrail differences matter more in practice than most benchmark charts do

Checklist:

- Decide whether you need live public conversation data, since that is Grok's clearest edge
- Check whether your Grok access includes live search or only the underlying model
- Test both on one hard problem where you already know the correct answer
- Use the more predictable model for anything that goes out unsupervised
- Check whether X premium already includes the Grok tier you were about to buy separately
- Total your AI subscriptions before adding a third general assistant

FAQ:

Q: Is Grok better than ChatGPT?
A: Not overall. Grok is better for real-time questions about public conversation and for work where a less hedged model helps. ChatGPT is better for integrations, document workflows, consistency, and enterprise settings. On general reasoning quality they trade places with every release.

Q: Does Grok have real-time information?
A: Yes, and it is its strongest differentiator. Grok can draw on posts from X as they are published, which is faster than waiting for articles to be written and indexed. Note that this depends on the access route: using the Grok model through a third-party product does not automatically include the live feed.

Q: Is Grok good for coding?
A: It is competitive. xAI ships fast coding-focused variants that handle everyday development work well. ChatGPT has more mature tooling around code, and dedicated in-editor assistants beat both for day to day work inside a repository.

Q: How much does Grok cost compared to ChatGPT Plus?
A: Grok's standalone subscription generally sits above ChatGPT Plus, with a much more expensive heavy tier above that, and access is also bundled into X premium plans. Check both pricing pages for current figures, since these tiers have been restructured several times.

Q: Can I use Grok and ChatGPT without paying for both?
A: Yes. A multi-model workspace like Whizi includes Grok alongside GPT, Claude, Gemini, and DeepSeek on one plan, which is usually cheaper than two separate subscriptions if you only need each one for part of your work.

### How to cancel ChatGPT Plus (web, iPhone, and Android)

URL: https://whizi.io/resources/how-to-cancel-chatgpt-plus/
Updated: August 2026

Quick answer: Cancel ChatGPT Plus wherever you subscribed. On the web, open Settings, then Subscription, then Manage my subscription, and choose Cancel plan. If Apple or Google billed you, cancel in iOS Settings under Subscriptions or in the Play Store. You keep Plus until the period ends, and your chat history stays.

Step by step instructions to cancel ChatGPT Plus wherever you are billed, plus what happens to your chats, when access ends, and how to request a refund.

#### First, find out who is actually charging you

This is the step people skip, and it is the reason most "I cannot find the cancel button" problems happen. ChatGPT Plus can be billed three different ways, and you can only cancel it where it was purchased.

- **Billed by OpenAI** if you subscribed on the website through a card. Cancel on the web.
- **Billed by Apple** if you subscribed inside the iPhone or iPad app. Cancel in iOS Settings. The website will not let you.
- **Billed by Google** if you subscribed inside the Android app. Cancel in the Play Store.

To check: open ChatGPT, go to **Settings**, then **Subscription**. If your plan was bought through an app store, the panel usually says so directly, with wording like "your subscription is managed by Apple". If you see a normal "Manage my subscription" button that opens a billing portal, you are billed by OpenAI.

If ChatGPT is not loading or you cannot get in at all, check your card statement or your email receipts. A receipt from Apple or Google Play tells you where to go without needing to log in anywhere.

One more case worth ruling out: if your access came through a Team or Enterprise workspace, you do not have a personal subscription to cancel. A workspace admin has to remove your seat, and you should ask them rather than hunting for a button that does not exist for you.

#### Cancel on the web (billed by OpenAI)

This is the most common path and it takes about a minute.

1. Go to **chatgpt.com** and sign in with the account that is being charged. If you have more than one account, check the email address on your receipt first, because cancelling the wrong one is easy and confusing.
2. Open **Settings**. It sits behind your profile picture or your name, depending on which layout you are seeing.
3. Choose **Subscription** (in some versions this is called **My plan**).
4. Click **Manage my subscription**. This opens the billing portal.
5. Choose **Cancel plan**, then confirm. You will usually be shown a downgrade offer or a short survey on the way out. Answering is optional.
6. Look at the panel afterwards. It should now show a date your access ends rather than a date you will next be charged.

Keep the confirmation email. If a charge appears later, that email is the fastest way to resolve it.

#### Cancel on iPhone or iPad (billed by Apple)

Apple subscriptions are managed entirely by Apple. Deleting the ChatGPT app does nothing to the billing, which is the single most expensive misunderstanding in this whole process.

1. Open the **Settings** app on your iPhone or iPad. This is the iOS Settings app, not the settings inside ChatGPT.
2. Tap **your name** at the very top of the screen.
3. Tap **Subscriptions**.
4. Select **ChatGPT** from the list.
5. Tap **Cancel Subscription** and confirm.

If ChatGPT is not in the list, you are not billed by Apple, so go back and try the web method. If you have several Apple Accounts, the subscription may live on a different one, so check the account you were signed into when you first subscribed.

You can also reach the same screen from the App Store: tap your profile picture, then **Subscriptions**.

#### Cancel on Android (billed by Google Play)

Same principle as Apple: Google handles the billing, so Google handles the cancellation.

1. Open the **Google Play Store** app.
2. Tap your **profile icon** in the top right.
3. Choose **Payments and subscriptions**, then **Subscriptions**.
4. Select **ChatGPT**.
5. Tap **Cancel subscription**, pick a reason if asked, and confirm.

You can do the same thing in a browser at play.google.com/store/account/subscriptions if you would rather use a computer. Make sure the Google account in the top corner is the one that pays for the subscription.

#### What happens after you cancel

**You keep Plus until the end of the current period.** Cancelling stops the next charge, it does not cut you off immediately. If you cancel on the 3rd and your period runs to the 27th, you have Plus features until the 27th. There is no reason to wait until the last day to click cancel, and several reasons not to.

**Your chats stay.** Cancelling downgrades your plan, it does not delete your account or your conversation history. Your chats, custom instructions, and saved memory remain on the free tier. What changes is access: lower message limits, restricted access to the reasoning models, smaller allowances for research and image generation, and less priority at busy times.

**Some things become read-only or unavailable.** Custom GPTs you built stay in your account but the paid capabilities behind them may stop working. Projects and larger uploads fall back to free tier limits. If you have work you care about, export or copy it while you still have full access.

**Deleting your account is a separate action.** If your goal is to remove your data rather than stop paying, cancel first, then go to **Settings**, then **Data controls**, then **Delete account**. That is permanent and it takes effect after a short grace period. Do not delete the account as a way of cancelling an Apple or Google subscription, because the billing lives outside OpenAI and will continue.

**If you want to stop training on your chats without cancelling**, that is also in **Data controls**, as a separate toggle. Plenty of people who cancel over privacy concerns did not realise the setting existed.

#### Refunds, and when you can get one

Subscriptions are normally not prorated. Cancelling in the middle of a month usually means you keep access to the end of it rather than getting money back for the unused days. That is the default across all three billing routes.

There are exceptions worth trying if the charge genuinely should not have happened:

| Situation | Where to ask | Realistic odds |
| --- | --- | --- |
| Charged after you thought you cancelled | OpenAI support via the help centre, with your confirmation email | Good, if you have the email |
| Renewed unexpectedly within the last day or two | OpenAI support, or Apple / Google if they billed you | Reasonable, ask politely and quickly |
| Billed by Apple and want a refund | reportaproblem.apple.com, choose the charge and request a refund | Mixed, Apple decides case by case |
| Billed by Google Play | Play Store order history, request a refund | Mixed, faster within 48 hours |
| Consumer cooling-off period in the EU or UK | OpenAI support, cite the right to withdraw | Depends on whether you have used the service |

Ask once, be specific, and include the date and amount. Support handles a lot of these, and a clear message with a receipt attached does better than a long complaint.

#### Before you go: is cancelling actually the fix?

It is worth being honest about why you are here, because three of the four common reasons have a better solution than going back to the free tier.

**"It is too expensive for what I use."** Then the free tier probably is the right answer, and you should cancel without guilt. Come back if your usage changes. Our [is ChatGPT Plus worth it](https://whizi.io/resources/is-chatgpt-plus-worth-it) breakdown has a two-week test that tells you when it is time to subscribe again.

**"I am paying for too many AI tools."** Cancelling one of four is a partial fix. The full fix is working out which subscriptions overlap, and the [AI subscription savings calculator](https://whizi.io/tools/calculator) does that in a couple of minutes. Most people find that two of their plans do the same job.

**"Another model is better for what I do."** That is usually true for a specific task rather than for everything, which is why people end up subscribing to two products and using each for half their work. Read [ChatGPT Plus vs Claude Pro vs Gemini Advanced](https://whizi.io/resources/chatgpt-plus-vs-claude-pro-vs-gemini-advanced) before you replace one $20 plan with a different $20 plan.

**"I want the good models without the single-vendor lock-in."** This is the case where a multi-model workspace makes more sense than any single subscription. [Whizi](https://whizi.io/pricing) puts GPT, Claude, Gemini, Grok, and DeepSeek behind one plan, so you can switch models inside a conversation instead of switching products. [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) is the honest side by side, including where ChatGPT still wins.

And if none of those apply and you simply do not need it right now, cancel, keep your chat history, and enjoy the free tier. It is genuinely good.

#### Common problems

**There is no cancel button.** You are billed through Apple or Google. Go to the store, not the website.

**You cancelled but were charged again.** Check whether you cancelled on the right account, and whether you have a second subscription through an app store. Two subscriptions on two routes is more common than you would think, especially if you signed up on your phone and later logged in on a laptop. If you also want your card off the account entirely, see [how to remove your card from ChatGPT](https://whizi.io/resources/how-to-remove-card-from-chatgpt).

**You cancelled and lost access immediately.** That should not happen at the end of a normal cancellation. If it did, you may have been on a trial, or the billing period may have ended on the same day. Contact support with your receipt.

**The billing portal will not load.** Try a different browser or disable extensions that block third-party scripts, since the portal is hosted by the payment provider. If it still fails, support can cancel from their side.

**You cannot log in to cancel.** Use the password reset first. If the account email is one you no longer control, contact support with the card's last four digits and a receipt, since that is how they verify ownership.

Key points:

- The most common reason a cancellation fails is billing through Apple or Google rather than OpenAI
- Cancelling never deletes your chat history, and deleting your account is a separate action
- You keep Plus features until the end of the period you already paid for

Checklist:

- Check whether OpenAI, Apple, or Google is charging you before hunting for a cancel button
- Cancel in the same place you subscribed, since deleting the app does not stop billing
- Confirm the subscription panel now shows an end date instead of a renewal date
- Save the confirmation email until the final date has passed
- Export anything you need while you still have full access
- Decide whether the real problem is the price, the tool, or the number of tools you pay for

Steps:

1. Find out where you are billed: Open ChatGPT settings and look at your subscription. If it says your plan is managed by Apple or Google, you have to cancel in that store, not on the website.
2. Cancel on the web: In ChatGPT, open Settings, then Subscription, then Manage my subscription. In the billing portal that opens, choose Cancel plan and confirm.
3. Cancel on iPhone or iPad: Open the iOS Settings app, tap your name at the top, tap Subscriptions, select ChatGPT, then tap Cancel Subscription.
4. Cancel on Android: Open the Google Play Store, tap your profile icon, choose Payments and subscriptions, then Subscriptions, select ChatGPT and tap Cancel subscription.
5. Confirm the cancellation: Check that your subscription now shows an end date rather than a renewal date, and keep the confirmation email until that date passes.

FAQ:

Q: Do I lose my chat history when I cancel ChatGPT Plus?
A: No. Cancelling downgrades your plan to the free tier and leaves your account, conversations, custom instructions, and memory in place. Deleting your history or your account is a separate action under Data controls in settings.

Q: Will I get a refund if I cancel mid-month?
A: Usually not. Subscriptions are not normally prorated, so you keep Plus features until the end of the period you already paid for. If you were charged in error or renewed unexpectedly, contact OpenAI support, or Apple or Google if they billed you.

Q: Why is there no cancel option in my ChatGPT settings?
A: Because you subscribed inside the iPhone or Android app, which means Apple or Google handles the billing. Cancel in iOS Settings under Subscriptions, or in the Google Play Store under Payments and subscriptions.

Q: Does deleting the ChatGPT app cancel my subscription?
A: No, and this is the most expensive mistake people make. App store subscriptions keep renewing after the app is deleted. You have to cancel in the store itself.

Q: How do I know my cancellation actually worked?
A: Your subscription panel should show a date when access ends rather than a date when you will next be charged, and you should receive a confirmation email. Keep that email until the end date passes.

Q: Can I keep access to good models after cancelling?
A: The free tier still gives you the flagship model with a daily cap. If you want broader model access without a single-vendor plan, a multi-model workspace like Whizi includes GPT, Claude, Gemini, Grok, and DeepSeek on one subscription.

### How to cancel Perplexity Pro (web, iPhone, and Android)

URL: https://whizi.io/resources/how-to-cancel-perplexity-pro/
Updated: August 2026

Quick answer: Cancel Perplexity Pro wherever you subscribed. On the web, open your account settings, find your subscription, and choose to cancel in the billing portal. If Apple or Google billed you, cancel in iOS Settings under Subscriptions or in the Play Store. You keep Pro until the period ends, and your threads and library stay on the free tier.

Step by step instructions to cancel Perplexity Pro wherever you are billed, what happens to your threads and spaces, when access ends, and how to ask for a refund.

#### First, work out who is actually charging you

Perplexity Pro can be billed three ways, and you can only cancel it where it was purchased. This is the step that explains almost every "there is no cancel button" complaint.

- **Billed by Perplexity** if you subscribed on the website with a card. Cancel on the web.
- **Billed by Apple** if you subscribed inside the iPhone or iPad app. Cancel in iOS Settings; the website cannot do it.
- **Billed by Google** if you subscribed inside the Android app. Cancel in the Play Store.

The fastest way to check is your receipt: a charge or invoice from Perplexity points at the web route, a receipt from Apple or Google Play points at the store. If you subscribed long ago and cannot remember, your card statement carries the same answer.

One more case worth ruling out: Pro access that came bundled through a partner offer, a carrier deal, or an employer does not always have a personal subscription behind it. If you never entered a payment method, there may be nothing to cancel, and the bundle owner controls the access.

#### Cancel on the web (billed by Perplexity)

The whole path lives in your account settings and takes about a minute.

1. Go to **perplexity.ai** and sign in with the account that is being charged. If you have several accounts, match the email on your receipt first.
2. Open your **account settings** from your profile menu.
3. Find the **subscription** section. It shows your plan and a button to manage it.
4. The manage button opens a **billing portal** hosted by the payment provider. Choose the cancel option there and confirm.
5. Check the panel afterwards: it should now show a date your access ends rather than a date you will next be charged.

The exact menu names move around as the product gets updated, but the structure is stable: profile menu, settings, subscription, manage, cancel. If the billing portal refuses to load, try another browser or disable extensions that block third party scripts, since the portal runs on the payment provider's domain.

Keep the confirmation email. If a charge appears later, that email is the fastest way to resolve it with support.

#### Cancel on iPhone, iPad, or Android (billed by the store)

Store subscriptions are managed entirely by the store. Deleting the Perplexity app changes nothing about the billing.

**On iPhone or iPad:** open the iOS **Settings** app, tap **your name** at the top, tap **Subscriptions**, select **Perplexity**, and tap **Cancel Subscription**. You can reach the same screen from the App Store via your profile picture.

**On Android:** open the **Google Play Store**, tap your **profile icon**, choose **Payments and subscriptions**, then **Subscriptions**, select **Perplexity**, and tap **Cancel subscription**. The same list exists in a browser at play.google.com/store/account/subscriptions.

If Perplexity is not in the store's subscription list, you are not billed by that store: go back and use the web method, and check whether a second account holds the subscription.

#### What happens after you cancel

**You keep Pro until the end of the paid period.** Cancelling stops the next charge rather than cutting you off, so there is no reason to wait until the last day.

**Your threads, library, and spaces stay.** Cancelling downgrades the plan, it does not delete the account or your history. What changes is capacity: the free tier gives you a limited number of the enhanced Pro searches per day, and the paid tier's model selection and heavier usage allowances step back to free-tier levels.

**Uploads and heavy research workflows feel the downgrade first.** If your work depends on large file uploads or long research sessions, export or finish what matters while you still have full access.

**Deleting the account is a separate action** in account settings, and it is permanent. Do not delete the account as a way of cancelling a store subscription: Apple and Google billing lives outside Perplexity and keeps renewing until cancelled in the store.

#### Refunds, and when asking is worth it

Subscriptions are normally not prorated: cancelling mid-period keeps your access to the end of it rather than returning money for unused days. That is the default on all three billing routes.

The exceptions worth pursuing are the same as for any AI subscription: a charge after you thought you cancelled, an unexpected renewal caught within a day or two, a store purchase you can dispute at reportaproblem.apple.com or in Play Store order history, and consumer withdrawal rights in the EU and UK. Ask once, be specific, attach the receipt.

The general playbook, including template wording and realistic odds per route, is in our [AI subscription refunds guide](https://whizi.io/resources/ai-subscription-refunds).

#### Before you go: is cancelling actually the fix?

Three common reasons bring people to this page, and they have different right answers.

**"I use it, but not $20 worth."** The free tier of Perplexity remains genuinely useful for search-style questions, so cancelling and staying on free is a reasonable answer, and you can resubscribe any month your usage spikes.

**"I have Perplexity for research and another subscription for everything else."** That stack is exactly the overlap the [AI subscription savings calculator](https://whizi.io/tools/calculator) exposes. Two $20 plans that each cover half your work usually lose to one plan that covers both halves. See [Whizi vs Perplexity Pro](https://whizi.io/vs/whizi-vs-perplexity-pro) for the honest comparison, including the research features where Perplexity still wins.

**"I want stronger general models, not just search."** Then compare before you resubscribe to anything: [Whizi](https://whizi.io/pricing) puts GPT, Claude, Gemini, Grok, and DeepSeek behind one subscription with web search included, so the research workflow and the writing workflow stop needing separate bills.

Key points:

- The usual reason a cancellation fails is billing through Apple or Google rather than Perplexity itself
- Cancelling never deletes your threads or your library, and you keep Pro until the end of the paid period
- Decide first whether the problem is the price, the tool, or how many AI subscriptions you are stacking

Checklist:

- Check whether Perplexity, Apple, or Google is charging you before hunting for a cancel button
- Cancel in the same place you subscribed, since deleting the app does not stop billing
- Confirm the subscription panel shows an end date instead of a renewal date
- Finish or export heavy research work while you still have Pro access
- Run your remaining AI subscriptions through the savings calculator before replacing one $20 plan with another

Steps:

1. Find out where you are billed: Check your receipt email. A charge from Perplexity means you cancel on the website; a receipt from Apple or Google means you cancel in that store, and the website cannot do it for you.
2. Cancel on the web: Sign in at perplexity.ai, open your account settings, find the subscription section, and choose the cancel option in the billing portal that opens.
3. Cancel on iPhone or iPad: Open the iOS Settings app, tap your name at the top, tap Subscriptions, select Perplexity, then tap Cancel Subscription.
4. Cancel on Android: Open the Google Play Store, tap your profile icon, choose Payments and subscriptions, then Subscriptions, select Perplexity and tap Cancel subscription.
5. Confirm the cancellation: Check that your subscription now shows an end date rather than a renewal date, and keep the confirmation email until that date passes.

FAQ:

Q: Do I lose my threads when I cancel Perplexity Pro?
A: No. Cancelling downgrades you to the free tier and leaves your account, threads, and library in place. Deleting your account is a separate, permanent action in account settings.

Q: Will I get a refund if I cancel Perplexity Pro mid-month?
A: Usually not. You keep Pro until the end of the period you paid for instead of getting money back. Charges after a cancellation or unexpected renewals are worth raising with support, or with Apple or Google if they billed you.

Q: Why is there no cancel option in my Perplexity settings?
A: Almost always because you subscribed inside the iPhone or Android app, so Apple or Google handles the billing. Cancel in iOS Settings under Subscriptions or in the Play Store under Payments and subscriptions.

Q: Does deleting the Perplexity app cancel my subscription?
A: No. Store subscriptions keep renewing after the app is deleted. Cancel in the store itself, then delete the app if you want to.

Q: What does the free tier of Perplexity keep after downgrading?
A: Standard search stays, along with your history. The enhanced Pro searches drop to a small daily allowance, and the paid model selection and heavier usage limits step back to free levels.

Q: What should I replace Perplexity Pro with if I still need research?
A: If search-with-sources is all you need, the free tier may be enough. If you also pay for a general AI subscription, compare a consolidated option first: Whizi includes web search plus GPT, Claude, Gemini, Grok, and DeepSeek on one plan.

### How to cancel SuperGrok or X Premium (and know which one you actually have)

URL: https://whizi.io/resources/how-to-cancel-supergrok/
Updated: August 2026

Quick answer: Cancel SuperGrok wherever you bought it: in your grok.com account settings, or in the app store that billed you. Grok access through X Premium is a separate subscription billed by X, so cancel that one in your X subscription settings. Each keeps running until you cancel it where it was purchased.

Step by step instructions to cancel SuperGrok on grok.com or Grok access through X Premium, why the two are billed separately, and where each cancel button lives.

#### The confusion this page exists to fix

Grok is unusual among AI assistants: two different subscriptions from two different companies unlock it, and they know nothing about each other.

- **SuperGrok** is xAI's own subscription, bought on grok.com or inside the standalone Grok app. It exists purely to give you more Grok: higher limits and the heavier modes.
- **X Premium and Premium+** are subscriptions to the X platform that include Grok access alongside the platform features. They are billed by X, on their own price and their own renewal date.

Both can be active at once, and that is exactly how people end up paying twice for the same assistant. If you have ever "cancelled Grok" and watched a charge arrive anyway, the odds are you cancelled one of the two and the other kept renewing.

So before touching any cancel button, look at your receipts or card statement and name what you are actually paying for: a charge referencing xAI or grok.com, a charge from X, a receipt from Apple, a receipt from Google Play, or more than one of these. Each one has its own door, covered below.

#### Cancel SuperGrok on the web (billed by xAI)

1. Go to **grok.com** and sign in with the account that is being charged.
2. Open your **account settings** from the profile menu.
3. Find the **subscription** section showing your SuperGrok plan.
4. Choose **manage**, which opens a billing portal hosted by the payment provider, and select the cancel option there.
5. Confirm, then check the panel shows an end date rather than a renewal date.

The exact menu wording shifts as the product updates, but the structure holds: profile, settings, subscription, manage, cancel. If the billing portal will not load, try another browser or disable extensions that block third party scripts.

Keep the confirmation email until the end date passes. If you were charged through the standalone Grok app on a phone rather than the website, the store method below is your route instead.

#### Cancel a store-billed Grok subscription (Apple or Google)

Subscriptions started inside an iPhone or Android app are managed entirely by the store, and deleting the app changes nothing about the billing.

**On iPhone or iPad:** open the iOS **Settings** app, tap **your name**, tap **Subscriptions**, select the Grok or X subscription in the list, and tap **Cancel Subscription**.

**On Android:** open the **Google Play Store**, tap your **profile icon**, choose **Payments and subscriptions**, then **Subscriptions**, select the entry, and tap **Cancel subscription**.

If nothing Grok-related appears in the store list, the store is not billing you: go back to the receipts and use the grok.com or X route instead.

#### Cancel Grok access that came with X Premium

If your Grok access rides on X Premium, the subscription belongs to X, and cancelling it is an X account action: open your **subscription settings under Premium** on the X website or app and manage the plan there. If you subscribed to Premium through the iPhone or Android app, the store holds the billing and the store method above applies.

Worth deciding before you cancel: X Premium bundles platform features with the Grok access. If you value the platform features, cancelling Premium to stop using Grok throws both away, and the right move might be keeping Premium and cancelling only a separate SuperGrok plan if you also have one.

And the reverse: if you only ever subscribed to Premium for Grok, check whether the tier you pay for is the one you need, since Grok limits differ between Premium tiers and SuperGrok.

#### What happens after you cancel, and refunds

**Access runs to the end of the paid period.** Cancelling stops the next charge rather than cutting you off immediately, on every route.

**Your conversation history stays** on the free tier, and free Grok access with lower limits remains available. Deleting an account, on grok.com or on X, is a separate and permanent action; do not use it as a cancellation shortcut while a store subscription is active, because store billing lives outside the account and keeps renewing.

**Refunds follow the usual subscription rules:** normally not prorated, worth requesting only for charges after a cancellation, surprise renewals caught quickly, store purchases disputed at reportaproblem.apple.com or in Play Store order history, and EU or UK withdrawal rights. The playbook with template wording is in the [AI subscription refunds guide](https://whizi.io/resources/ai-subscription-refunds).

And the audit question that saves the most money: if you were paying for Grok twice, what else in your AI stack overlaps? The [savings calculator](https://whizi.io/tools/calculator) answers that in two minutes, and [Whizi vs Grok](https://whizi.io/vs/whizi-vs-grok) shows what one consolidated plan covers, Grok included, alongside GPT, Claude, Gemini, and DeepSeek.

Key points:

- SuperGrok and X Premium are two different subscriptions that both unlock Grok, and each must be cancelled where it was bought
- People who "cancelled Grok" and keep getting charged usually cancelled one of the two while the other kept renewing
- Cancelling keeps your access until the end of the paid period, and your conversation history stays

Checklist:

- Name the subscription first: SuperGrok from xAI, X Premium from X, or a store-billed version of either
- Check your card statement for more than one Grok-related charge before assuming there is only one
- Cancel in the place that billed you, since the other services cannot do it for you
- Confirm an end date replaced the renewal date, and keep the confirmation email
- Audit the rest of your AI subscriptions while you are at it

Steps:

1. Identify which subscription you have: Check your receipt. A charge from xAI or grok.com means SuperGrok. A charge from X means X Premium. A receipt from Apple or Google means the store bills you for whichever app you subscribed in.
2. Cancel SuperGrok on the web: Sign in at grok.com, open your account settings, find the subscription section, and choose the cancel option in the billing portal that opens.
3. Cancel a store-billed subscription: On iPhone, open iOS Settings, your name, Subscriptions, and cancel there. On Android, open the Play Store, Payments and subscriptions, Subscriptions, and cancel there.
4. Cancel Grok access that came with X Premium: Open X, go to your subscription settings under Premium, and manage or cancel the plan there, or in the app store if you subscribed through the X app.
5. Confirm the cancellation: Check that the subscription now shows an end date rather than a renewal date, and keep the confirmation email until that date passes.

FAQ:

Q: What is the difference between SuperGrok and X Premium?
A: SuperGrok is xAI's own subscription bought on grok.com or in the Grok app, and it only buys more Grok. X Premium is a subscription to the X platform that includes Grok access among its features and is billed by X. They are separate products with separate bills.

Q: I cancelled Grok but I am still being charged. Why?
A: Almost always because two subscriptions were active: a SuperGrok plan and an X Premium plan, or a web subscription plus a store-billed one. Check your card statement for the source of the remaining charge and cancel on that route.

Q: Do I lose my Grok chats when I cancel?
A: No. Cancelling downgrades you to the free tier and your history stays. Deleting the account is a separate, permanent action.

Q: Does deleting the Grok app or the X app stop the billing?
A: No. Store subscriptions keep renewing after an app is deleted, and web subscriptions are unaffected by apps entirely. Cancel in the store or on the website that billed you.

Q: Can I keep using Grok for free after cancelling?
A: Yes, within the free limits. Free Grok access exists with lower usage caps, and your history remains available.

Q: Is there a way to keep Grok without a separate subscription for it?
A: If you want Grok alongside other models rather than on its own bill, a multi-model workspace covers that: Whizi includes Grok together with GPT, Claude, Gemini, and DeepSeek on one subscription.

### Which AI model should I use? How to choose in 10 minutes

URL: https://whizi.io/resources/how-to-choose-ai-model/
Updated: August 2026

Quick answer: Choose an AI model by defining the task first: input, action, output, and review standard. Name the constraint that matters most, cost, speed, privacy, accuracy, or edit time, then run the same prompt and source material through two or three candidates and score the results. Keep a routing rule per workflow.

Use a 10-minute decision framework, scorecard, and A/B test protocol to choose the best AI model for writing, coding, research, docs, and mixed work.

#### Define your task

The fastest way to answer "which ai model should i use?" is to stop asking it in the abstract. AI models are not equally good at every job. A model that writes a crisp email may not be the best choice for a long PDF extraction, and a model that explains code well may not be the one you want for a polished executive memo. Define the job first.

Use this task frame before comparing models: **input, action, output, review standard**. Input is what the model receives: notes, code, screenshots, a PDF, a data table, or a blank prompt. Action is the work you need done: summarize, rewrite, debug, extract, compare, classify, brainstorm, plan, or synthesize. Output is the deliverable: email, table, code patch, research memo, checklist, outline, JSON-like fields, or decision recommendation. Review standard is how you will decide whether the answer is good enough.

Here is the practical version: "I need to [action] using [input] and produce [output]. The answer is good if it is [review standard]." Example: "I need to summarize a 30-page investor memo into a risk table. The answer is good if every risk is traceable to the source and grouped by severity." That definition points you toward a long-context and verification-friendly workflow instead of a generic chatbot preference.

Do this before reading another model ranking. Official OpenAI, Anthropic, and Gemini model docs describe model families and capabilities, but they cannot know your audience, source material, cost constraints, or tolerance for mistakes. Your task definition turns a vague model choice into a small experiment.

#### Constraints: cost, speed, privacy

After the task, define the constraints. Most model decisions are tradeoffs between quality, speed, cost, privacy, and workflow friction. If you do not name the constraint upfront, you may pick the most impressive answer instead of the most useful one.

Cost matters when you are paying for multiple subscriptions or team seats. Speed matters when the task is part of support, sales, operations, or engineering review. Privacy matters when the input includes customer data, internal strategy, credentials, employee information, financial information, or anything your organization would not want pasted into an unapproved tool.

Use this checklist: What editing time can you accept? Does the answer need to be correct, or only useful as a draft? Can you paste the source material into the tool? Do you need citations or traceability? Will this run once, weekly, or hundreds of times? Is the task reversible if the AI gets it wrong?

A cheap model can be expensive if it creates cleanup work. A powerful model can be wasteful if the task is a simple rewrite. A fast model can be risky if the output needs careful source handling. The right AI model is the one that clears the constraint that matters most for the job in front of you.

#### Capabilities: vision, tools, long docs

Now check capabilities. The big categories are text quality, reasoning, coding, long context, vision, structured outputs, tool use, and file handling. A model does not need to win every category. It needs to support the capabilities your task requires.

For writing, evaluate voice control, specificity, structure, and edit time. For coding, evaluate whether the model can reason from a reproduction, propose a small fix, and name protective tests. For research, evaluate source discipline and uncertainty. For document work, look for long-context handling and structured outputs. For image, screenshot, and mixed-media tasks, choose a multimodal model and ask for extraction before interpretation.

The text-only versus multimodal decision is straightforward: if the input is only notes, prose, code, or structured text, a strong text model may be enough. If the input includes screenshots, charts, images, scanned documents, PDFs, or mixed visual context, test a multimodal model. The long-context decision is similar: if the important information is spread across many pages or files, use a model and workflow designed to handle longer inputs, then verify the answer against the original source.

Do not treat capability as a yes-or-no checkbox. Treat it as a test requirement. If the task needs vision, test with a real image. If it needs long context, test with a long source. If it needs tools, ask the model what data it would need before it answers.

#### A/B test protocol

You do not need to pick one model forever. Run a small A/B test when the task matters, then save the model choice that wins for that workflow. Whizi is built for this habit: run the same prompt across models, compare the outputs side by side, and keep the routing rule that works.

Here is the protocol, and it genuinely fits in ten minutes.
1. **Define the task.** Input, action, output, and the standard you will review it against.
2. **Pick two or three candidates** based on the capability the task needs, not on which one you like.
3. **Paste the same prompt and the same source material into each.** Identical inputs or the test proves nothing.
4. **Read and score the outputs** using the scorecard below. Give this three or four minutes, not thirty seconds.
5. **Challenge each one** with the same follow-up: "What could be wrong with this answer, and what should I verify?" The quality of that reply is often more revealing than the original answer.
6. **Pick a winner for this workflow**, not for everything.
7. **Write it down.** Save the prompt, the winning model, and one note about when you would use a different one.

Copy-paste test prompt: "I am choosing an AI model for this workflow. Complete the task using only the context provided. Follow the output format exactly. After the answer, include assumptions, risks, and a verification checklist. Task: [task]. Context: [source material]. Output format: [format]. Quality bar: [how I will judge success]."

Use this scorecard from 1 to 5 for each output. A perfect score is rare. The winner is the model that gives you the best usable answer under the constraint that matters most.

| Scorecard item | What to look for | Red flag |
| --- | --- | --- |
| Accuracy | Claims match the source or your known facts | Confident details you did not provide |
| Usefulness | The output moves the work forward | Polished prose with no decision value |
| Format compliance | It follows the requested table, memo, list, or schema | It ignores required fields |
| Specificity | It uses your context, examples, and constraints | Generic advice that could fit anyone |
| Edit time | You can use it with light revision | You need to rewrite the whole answer |
| Speed | It returns fast enough for the workflow | Quality is fine but too slow for routine use |
| Cost fit | The model is appropriate for task value | Premium effort on a low-stakes task |
| Context handling | It uses the full source without losing key details | It misses important sections or mixes facts |
| Verification risk | It surfaces assumptions and checks | It hides uncertainty |

#### Decision table

Use this table as a starting point, not a permanent ranking. The best ai model for writing, coding, research, or long documents depends on your exact task and review standard. The table simply tells you where to start the test.

| Task | Start by testing | Challenger | Decision rule |
| --- | --- | --- | --- |
| Email, outline, or quick first draft | A fast general-purpose text model | A stronger writing model | Pick the one with the least cleanup and most specific context use |
| Long-form editing or tone-sensitive copy | A writing-focused model | A general-purpose model | Pick the one that improves structure without flattening voice |
| Debugging or implementation planning | A coding-capable reasoning model | A careful review-oriented model | Pick the one that proposes the smallest safe change and tests |
| Code review or refactor planning | A careful long-context model | A coding-focused model | Pick the one that catches real risks, not style noise |
| Research from provided sources | A model strong at synthesis | A model strong at long-context extraction | Pick the one that separates claims, sources, and uncertainty |
| Large PDF or document analysis | A long context ai model | A model known for careful summarization | Pick the one that extracts before summarizing and flags gaps |
| Screenshot, image, chart, or mixed media | A multimodal model | Another multimodal-capable model | Pick the one that returns structured observations before conclusions |
| Daily mixed work | Whizi side-by-side testing | Two or three major models | Pick a routing rule instead of one permanent winner |

The most mature AI workflows use routing rules: one model for quick drafts, another for careful editing, another for long documents, and another for image or screenshot tasks. That is why "ChatGPT vs Claude vs Gemini which is best" is usually the wrong final question. Ask which model should handle this task first, and when you should compare.

For a deeper comparison of the major model families, read [ChatGPT vs Claude vs Gemini](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini). When you are ready to test your own prompts, [create a Whizi account](https://whizi.io/register), run the same prompt across models, and compare plans at [pricing](https://whizi.io/pricing) if you want one workspace for the whole routing system.

Key points:

- A 10-minute protocol for choosing a model without turning the decision into a research project
- A practical scorecard for accuracy, format, speed, cost, context fit, and verification risk
- A decision table that helps you route writing, coding, research, long-doc, and multimodal tasks

Checklist:

- Define the task as input, action, output, and review standard
- Name the constraint that matters most: cost, speed, privacy, accuracy, or edit time
- Choose candidate models based on required capabilities, not brand preference
- Use the exact same prompt and source material for every model test
- Score outputs before revising the prompt
- Ask each model what could be wrong with its answer
- Save a routing rule for repeatable workflows
- Use Whizi when a task matters enough to compare models side by side

FAQ:

Q: Which AI model should I use?
A: Define the task first: input, action, output, and review standard. Then pick the constraint that matters most (cost, speed, privacy, accuracy, or edit time) and test two or three candidate models on the same prompt. The right model is the one that clears your most important constraint with the least cleanup, not the one that tops a generic ranking.

Q: How do I compare AI models quickly?
A: Run the 10-minute A/B protocol: paste the same prompt and source material into each model, score the outputs on accuracy, format compliance, specificity, edit time, and verification risk, then ask each model what could be wrong with its answer. Save the winner as your routing rule for that workflow.

Q: Do I need a multimodal model or a text-only model?
A: If your input is only notes, prose, code, or structured text, a strong text model is usually enough. If it includes screenshots, charts, images, scanned documents, or PDFs with visual context, test a multimodal model and ask it to extract observations before interpreting them.

Q: Is one AI model best for everything?
A: No. Mature workflows use routing rules: one model for quick drafts, another for careful editing, another for long documents, and another for image or screenshot tasks. Instead of picking one model forever, decide which model handles each kind of task and re-test when a workflow matters.

### How to remove your card from ChatGPT (and actually stop the charges)

URL: https://whizi.io/resources/how-to-remove-card-from-chatgpt/
Updated: August 2026

Quick answer: To remove your card from ChatGPT, open Settings, then Subscription, then Manage my subscription, and delete it under payment methods in the billing portal. If no delete option appears, the subscription is still active: cancel it first, or add a replacement card. If Apple or Google bills you, the card is in their settings, not ChatGPT's.

Step by step instructions to remove your credit card from ChatGPT, why the delete option is sometimes missing, and how to make sure the billing really stops.

#### Before anything: know which of two problems you are solving

People search for how to remove a card from ChatGPT for two different reasons, and the right steps depend on which one is yours.

- **You want to stop being charged.** Then the action that matters is cancelling the subscription, not deleting the card. A removed card does not end a plan; the provider simply asks for a new payment method, retries, or suspends the account while the plan technically continues.
- **You want your card details out of the account.** Maybe you used a work card, a card that is about to expire, or you just do not want it on file after cancelling. Then the billing portal is where you go, and this guide walks through it.

If you want both, do them in this order: cancel first, then remove the card. Both live in the same portal, so it takes two minutes.

And one disclaimer that prevents a lot of confusion: if you subscribed inside the iPhone or Android app, **ChatGPT has never seen your card**. Apple or Google holds it, and nothing in ChatGPT settings will show or remove it. Skip to the app store section below.

#### Remove a card on the web (billed by OpenAI)

1. Go to **chatgpt.com** and sign in with the account that is being charged. Check the email on your receipt if you have more than one account.
2. Open **Settings** behind your profile picture or name.
3. Choose **Subscription** (sometimes labelled **My plan**).
4. Click **Manage my subscription**. This opens the billing portal, which is hosted by the payment provider rather than on the ChatGPT site itself.
5. Find the **payment methods** section. Your saved cards are listed by their last four digits.
6. Choose the card and select the option to **delete** or **remove** it, then confirm.

The exact wording moves around as the portal gets updated, but the structure stays the same: subscription status at the top, payment methods and billing history below it.

If the portal will not load at all, try another browser or temporarily disable extensions that block third party scripts, since the portal runs on the payment provider's domain. If it still fails, OpenAI support can update payment details from their side.

#### Why the delete option is missing, and what to do about it

This is the step where most people get stuck, so here is the honest explanation: **while a subscription is active, the portal usually refuses to remove the only card on file**, because it needs something to charge at renewal. The delete option is either greyed out or simply not shown.

You have three ways through, pick the one that matches your goal:

- **Cancel the plan first.** Once the subscription is set to end, the card stops being required and the delete option appears. Follow our [how to cancel ChatGPT Plus](https://whizi.io/resources/how-to-cancel-chatgpt-plus) walkthrough, then come back and remove the card.
- **Add a replacement, then delete the old one.** If you want to keep the subscription and only swap cards, add the new card first, set it as the default, and the old one becomes deletable.
- **Wait out the end date.** If you already cancelled, the card can be removed as soon as the paid period ends, and often immediately after the cancellation is confirmed.

What we do not recommend: cancelling the card at your bank as a way of cancelling the subscription. It works eventually, but it can drag your account through failed-payment retries and support emails, and it does nothing Apple- or Google-side if that is where the billing lives. Cancel properly, it is genuinely a two minute job.

#### Subscribed on iPhone or Android? Your card is not in ChatGPT

App store billing keeps your payment details entirely inside your Apple or Google account. ChatGPT only learns that you paid, never how.

**On iPhone or iPad:** your cards live in the iOS **Settings** app, under your name, then **Payment & Shipping**. Subscriptions are managed separately under your name, then **Subscriptions**. Cancelling the ChatGPT subscription there stops the charges; removing a card from Payment & Shipping affects everything billed to your Apple Account, not just ChatGPT.

**On Android:** cards live in the **Google Play Store** under your profile icon, then **Payments and subscriptions**, then **Payment methods**. The same caveat applies: a card removed there disappears for every Play Store purchase, not only ChatGPT.

Because store cards are shared across all your apps, the cleaner move on mobile is almost always to cancel the ChatGPT subscription and leave the card alone, unless you want it gone from the store entirely.

#### After removal: what actually happens to your card data

A few facts that answer the questions people are usually really asking:

**Removing the card stops nothing by itself.** An active plan continues and the provider will ask for a new payment method. If your goal was to stop paying, confirm the subscription shows an end date. That is the signal that matters.

**Billing history keeps the last four digits.** Past invoices still show which card paid them. That is normal accounting retention across the industry, not a sign the full card is still on file.

**Charges after removal usually mean a second billing route.** The classic case: you subscribed on the website and, months earlier, also started a subscription in the app. Check both the billing portal and your phone's subscription list before assuming the removal failed. Our [refunds guide](https://whizi.io/resources/ai-subscription-refunds) covers what to do about a charge that should not have happened.

**Deleting your account removes the billing profile too.** If your goal is full removal of your data, cancel the subscription, then use Settings, then Data controls, then Delete account. Do not delete the account as a shortcut while an app store subscription is active, because that billing lives outside OpenAI and keeps renewing.

#### While you are in there: is the $20 doing its job?

Most people cleaning up their ChatGPT billing are really auditing their AI spend. Two minutes of that audit pays for the visit:

If you are paying for more than one AI subscription, run the stack through the [AI subscription savings calculator](https://whizi.io/tools/calculator). Overlapping plans are the norm, not the exception, and most people find two of their tools doing the same job.

If you are removing the card because you are switching tools, read [ChatGPT Plus vs Claude Pro vs Gemini Advanced](https://whizi.io/resources/chatgpt-plus-vs-claude-pro-vs-gemini-advanced) before replacing one $20 single-vendor plan with another. And if the real complaint is vendor lock-in rather than price, [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) is the honest side by side: one subscription that includes GPT, Claude, Gemini, Grok, and DeepSeek, including the places where ChatGPT still wins.

Key points:

- Removing your card and cancelling your subscription are two different actions, and doing only one of them is how surprise charges happen
- The delete option is often hidden while a subscription is active, which is the number one reason people think it is impossible
- If you subscribed through the iPhone or Android app, your card is not in ChatGPT at all, it is in your Apple or Google account

Checklist:

- Decide whether you are stopping charges, removing card data, or both
- Cancel the subscription first if the delete option is missing
- Check both the web billing portal and your phone for a second subscription
- Confirm the panel shows an end date, not a renewal date
- Keep the confirmation email until the final billing date passes

Steps:

1. Open the billing portal: In ChatGPT, open Settings, then Subscription, then Manage my subscription. This opens the billing portal where your saved cards live.
2. Find your payment methods: In the billing portal, look for a payment methods or payment information section. Your saved cards are listed there with their last four digits.
3. Remove or replace the card: Delete the card you want gone. If the portal refuses because a subscription is active, either cancel the plan first or add a replacement card and then delete the old one.
4. Cancel the plan if your goal is to stop paying: Removing a card does not cancel a subscription. Choose Cancel plan in the same portal so the panel shows an end date instead of a renewal date.
5. Confirm nothing is left behind: Reopen the portal and check that the card list and the subscription status both look the way you expect, then keep the confirmation email.

FAQ:

Q: Why can I not remove my card from ChatGPT?
A: Almost always because the subscription is still active and it is the only card on file. Cancel the plan first, or add a replacement card and then delete the old one. Once nothing needs charging, the delete option appears.

Q: Does removing my card cancel ChatGPT Plus?
A: No. The subscription stays active and the provider will ask for a new payment method or retry billing. To stop paying, cancel the plan in the billing portal so it shows an end date.

Q: I removed my card but was charged again. How?
A: Usually a second billing route: a website subscription plus an app store subscription on the same account, or the charge going to a different card or account than you checked. Look at the receipt to see who billed you, then cancel on that route.

Q: Where is my card if I subscribed in the iPhone or Android app?
A: In your Apple or Google account, not in ChatGPT. Apple keeps it under Settings, your name, Payment & Shipping. Google keeps it in the Play Store under Payments and subscriptions. ChatGPT never receives card details for app store billing.

Q: Does ChatGPT keep my card details after I remove the card?
A: The saved card is removed from your payment methods, while past invoices keep the last four digits for record keeping. If you want your whole billing profile gone, cancel first and then delete the account under Data controls.

Q: Can I use a different card for ChatGPT without losing my plan?
A: Yes. In the billing portal, add the new card first, make it the default, then remove the old one. The subscription continues without interruption.

### How to talk to AI: what to type when the box is empty

URL: https://whizi.io/resources/how-to-talk-to-ai/
Updated: August 2026

Quick answer: Talk to AI by writing a request in full sentences instead of typing keywords. Say what you want done, who it is for, how long the answer should be, and the facts only you know. Treat the first reply as a draft and correct it in the same chat, because the conversation remembers everything you already said.

The chat window is open and the cursor is blinking. Here is what to type first, the four details worth including, ten openers to copy, and the follow-ups that fix a bad answer.

#### You are not typing into a search box

The cursor blinks. The placeholder text says something like "Message ChatGPT" and you sit there, because nothing you can think of feels like a proper question. This is the most common first experience people have with AI, and it says nothing about you. The box looks exactly like a search bar and behaves like nothing you have used before.

Almost everyone types keywords the first time: three or four words, no verb, no context. That is a reflex trained by twenty years of search engines, and it is the one habit worth dropping on day one. A search box wants keywords because it is matching pages. A chat model wants sentences because it is working out what job you are handing it. Take a real need, cancelling a gym membership without an argument, written both ways:
- **As a search query:** cancel gym membership letter template
- **As a request:** I want to cancel my gym membership. Write a short, polite message I can send by email. I have been a member two years, my contract needs 30 days notice, and I do not want to give a reason.

The first version returns a generic template with brackets to fill in. The second returns a message you can send, with the notice period already in it and the reason deliberately left out. Same need, about fifteen seconds more typing, a completely different level of usefulness. Try it now on something small from your own week, an email you owe someone or a form you have been putting off. It will also, at some point, tell you something untrue in a confident voice, which the last section deals with. If all of this is new, [what AI actually is](https://whizi.io/resources/what-is-ai) covers the ground underneath it, and [AI versus Google search](https://whizi.io/resources/ai-vs-google-search) explains which questions still belong in a search engine. The two boxes reward completely different habits.

| Difference | A search box | A chat model |
| --- | --- | --- |
| What it is doing | Matching keywords against pages that already exist | Working out what job you are handing it, then producing something new |
| What to type | Three or four words with no verb | Sentences that say what you want, for whom, and how long |
| What comes back | A generic template with brackets for you to fill in | A message you can send, with your own details already in it |
| When it is the right tool | A page somewhere already holds your answer | You want something written, compared, planned, or explained for your case |

#### Four things worth putting in a first message

There is no formula to memorize. What matters is including the things only you know, because the model cannot see your inbox, your deadline, your job, or the person you are writing to. Think of it as briefing a capable temp on their first morning. They can do the work, but nobody has told them who the work is for.

Four details carry almost all of the weight. Two of them is usually enough to change the answer, and all four is better.
1. **What you want.** The verb and the thing you want back: rewrite, summarize, plan, compare, explain, draft.
2. **Who it is for.** Your manager, a client, your landlord, a nine-year-old, yourself in six months.
3. **How long.** One sentence, a paragraph, five bullets, one page. Left unsaid, most models pick a length nobody asked for.
4. **Facts only you know.** The date, the budget, the history, the thing you cannot say out loud in the message itself.

Here is one real message, twice. Before: "Write an email to my team about the delay." After: "Write a short email to my six-person team explaining that the client launch moved from March 3 to March 17. Three sentences, calm, no apology spiral. They already know the client changed the brief, so do not re-explain that part." The extra detail is not politeness. It is the six people, the two dates, and the one thing they already know, and it removes about three rounds of correcting afterwards.

#### The follow-up is where the value is

Most people send one message, read the answer, decide AI is either amazing or useless, and close the tab. That single-shot habit is responsible for most of the disappointment. The first answer is a first draft written by someone who has never met you, and the whole point of a chat is that you get to reply.

The conversation remembers everything you have already said, so you never rewrite the original request to fix it. You can type three words. These are the corrections that handle the usual problems, and you can copy them as they are.
- **Too long:** Half the length. Cut the wind-up, keep the specifics.
- **Too formal:** Say it the way I would say it out loud to a colleague I like.
- **Wrong angle:** You wrote it as a [X]. I need it as a [Y]. Try again with that framing.
- **It made something up:** Which parts of that did I tell you, and which did you fill in yourself? List the ones you filled in.
- **It does not sound like me:** Here are two things I wrote myself. Match this voice instead of a corporate one. [paste two of your own paragraphs]

The fourth line is worth turning into a habit. These systems produce fluent text whether or not they know the answer, so asking which parts were filled in surfaces the invented ones faster than checking every sentence yourself. It works on figures too: ask where a number came from and you will often get an honest "I estimated that." You will notice none of these corrections are polite. Politeness is optional here. What decides the next answer is whether the correction names something to change, which is why "make it better" does nothing and "cut the first paragraph" works every time.

#### Ten opening lines you can copy

Pick the one that matches something actually sitting on your desk today. Each is written to be pasted as it is, with your own material added underneath.

1. Rewrite this email so it is shorter and warmer, and keep the request clear: [paste]
2. Explain this document in plain language, then tell me what I should be worried about: [paste]
3. Plan a [event] for [number] people on [date] with a budget of [amount]. Give me a timeline, not ideas.
4. I am deciding between [A] and [B]. Build a table of tradeoffs, then tell me what I have not considered.
5. Teach me [topic] in five minutes starting from zero, and stop twice to check I followed you.
6. I need to tell [person] [awkward thing]. Draft it kindly, three sentences, no over-apologizing.
7. Summarize this in five bullets for someone who will only read the first two: [paste]
8. Translate this into [language] and keep the tone friendly but professional: [paste]
9. Here is my reasoning for [decision]. Argue the strongest case against it: [paste]
10. I need to write [thing] and I do not know how to start. Ask me four questions first, then draft it.

Number ten is the one for a cursor that is still blinking. Handing the blank page back as a short interview is the fastest way out of it, and the four questions it asks will usually be the four things you had not decided yet. If you want a longer supply organized by the kind of day you are having, [AI for everyday tasks](https://whizi.io/resources/ai-for-everyday-tasks) has a week of them.

#### When the answer is bad, and what to do about it

Bad answers are not random noise. They arrive in a few recognizable shapes, and each one has a repair that takes a single message.

- **The confident invention.** A statistic, a citation, a policy, or a product feature that does not exist, delivered in exactly the tone it uses for things it knows. The repair: ask it to mark every claim it is not certain about, then check those yourself, and never pass on a number you have not seen at its source. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) explains where the invented ones come from.
- **The bland average.** Technically correct, entirely generic, reads like a brochure for nothing in particular. The repair: give it something specific to react to. Your draft, your constraints, your own sentences. Vague input produces vague output, reliably.
- **The agreeable mirror.** You asked why X is better than Y, and it argued for X, because that is how you framed it. The repair: ask the opposite question in a new message and read both answers side by side before you decide.

If you have repaired the same answer twice and it is still wrong, stop rephrasing. Two rewrites usually fix a wording problem. A fourth rarely fixes a model problem, and models fail in different directions: one invents sources while another refuses to commit to anything, one writes warmly while another stays stiff no matter what you tell it. Sending the identical message to a different model is often a five-second fix for something you were about to spend twenty minutes on. That is worth arranging even if it just means keeping a free account on a second provider; Whizi is the version of that where the switch is a dropdown instead of a second 20 dollars a month. [How to choose an AI model](https://whizi.io/resources/how-to-choose-ai-model) covers which one to reach for first, and once correcting answers feels routine, [prompt engineering for beginners](https://whizi.io/resources/prompt-engineering-beginners) turns these repairs into a structure you can reuse.

Key points:

- Most bad first experiences come from typing a search query into something that is waiting for a request written in sentences
- You never have to get the first message right, because the chat remembers what you said and "make it shorter" is a complete instruction
- When two corrections have not fixed an answer, the problem is usually the model rather than your wording, so send the same message somewhere else

Checklist:

- Write your first message as a request in full sentences, not as keywords.
- Say who the output is for and how long you want it to be.
- Include the facts only you know, such as dates, budgets, and history.
- Treat the first answer as a draft and reply with one specific correction.
- Ask which parts of an answer were filled in before you rely on any number.
- Paste two paragraphs of your own writing when the output needs to sound like you.
- Send the same message to a different model instead of rephrasing a fourth time.

FAQ:

Q: What do I say to ChatGPT the first time?
A: Bring something real from your day rather than a test question. Paste an email you need to answer or a document you need to understand, then say what you want done with it and who it is for. A first message like "rewrite this so it is shorter and friendlier, it is going to a client" will teach you more in thirty seconds than any tutorial.

Q: What should I ask AI if I have no specific question?
A: Turn the blank page around: type "I need to write [thing] and I do not know how to start. Ask me four questions first, then draft it." The questions it asks are usually the decisions you had not made yet. This also works for planning, deciding between options, and preparing for a difficult conversation.

Q: Do I need to be polite to an AI?
A: No. Please and thank you cost nothing and change little about the quality of the answer. What actually changes the output is specificity: the audience, the length, and the facts only you have. Blunt instructions like "half the length, keep the numbers" are perfectly good messages.

Q: How do I fix a bad answer without starting over?
A: Reply in the same conversation with one correction. The chat already holds everything you said, so short messages like "too formal, say it the way I would say it out loud" work without repeating your original request. Starting a new chat throws away context you already paid for in typing.

Q: Is talking to AI different from searching Google?
A: Yes, and mixing them up is the most common beginner mistake. Search matches keywords against existing pages, so short queries work. A chat model produces something new for your situation, so it needs sentences, context, and a clear job. Use search when a page somewhere already has your answer, and chat when you want something written, compared, planned, or explained for your case.

### How to use AI: a practical beginner's guide

URL: https://whizi.io/resources/how-to-use-ai/
Updated: August 2026

Quick answer: Use AI by giving it four things in every prompt: a persona, a specific task, the context it needs, and the output format you want. Then verify the result before you act on it, checking facts against sources and watching for invented details. Match the model to the job rather than picking one forever.

Learn how to use AI for work with a simple prompt framework, 25 copy-paste prompts, verification checklist, and model-choice guide.

#### What AI can and cannot do

Learning how to use AI starts with a healthier expectation: AI is not a magic answer machine. It is a drafting, reasoning, summarizing, brainstorming, classifying, and transformation tool. It can turn rough notes into a clean outline, compare options, rewrite copy in a specific tone, explain a confusing concept, extract fields from a document, generate code examples, or help you pressure-test a plan. Used well, it gives you a faster first pass and a sharper second pass.

The catch is that generative AI is still probabilistic. It may sound confident while missing context, inventing details, flattening nuance, or choosing a format that does not match your real goal. That is why beginners should avoid asking, "What should I do?" as a stand-alone prompt. Better prompts give the model a role, task, context, constraints, and output format. Better workflows also include verification before anything important gets shipped.

A useful beginner ai workflow has four moves: define the job, give the model enough context, ask for a structured output, then review the result against sources or your own standards. This guide gives you the exact framework, examples, and checks to do that without becoming a prompt engineering specialist.

#### The beginner prompt template

Use this four-part template whenever you are learning how to use ai chatbots: Persona, Task, Context, Format. It is simple enough to remember and specific enough to improve almost every output.

Persona: Tell the model what perspective to use. Examples: "Act as a senior product marketer," "Act as a careful research assistant," or "Act as a patient tutor." Persona should shape judgment, not create fake authority. If the output needs legal, medical, financial, or safety review, say so and keep a human expert in the loop.

Task: State the action clearly. Weak task: "Help with this." Strong task: "Turn these notes into a 600-word client update with three sections: progress, risks, and next steps." The task should contain the verb: summarize, compare, rewrite, extract, critique, plan, debug, classify, or generate.

Context: Give the model the material it needs to avoid guessing. Include audience, goal, source text, background, constraints, examples of good output, and anything it should not do. For long inputs, put instructions before the source material and label the source material clearly so the model can separate your request from the content.

Format: Specify the output shape. Ask for a table, checklist, email, outline, JSON-like fields, decision memo, bullets, or a draft with headings. Format is one of the easiest ways to get better results because it turns vague assistance into a deliverable.

Copy-paste master template: "Act as [persona]. Your task is to [specific task]. Context: [audience, goal, source material, constraints, examples]. Requirements: [must include, must avoid, tone, length]. Output format: [table/checklist/email/outline/memo]. Before finalizing, list any assumptions or missing information."

#### 25 prompts by scenario

Use these ai prompts for beginners as working starters. Replace the bracketed details, paste your source material where needed, and run the same prompt across more than one model when the decision matters.

1. Work plan: "Act as an operations lead. Turn this goal into a one-week execution plan. Goal: [goal]. Context: [team, deadline, constraints]. Output a table with day, priority, owner, deliverable, and risk."

2. Meeting notes: "Act as an executive assistant. Summarize these meeting notes into decisions, open questions, owners, and deadlines. Keep action items specific and flag anything ambiguous. Notes: [paste notes]."

3. Email draft: "Act as a concise business writer. Draft an email to [audience] about [topic]. Goal: [outcome]. Tone: warm, direct, not salesy. Include a subject line and a clear next step."

4. Email rewrite: "Rewrite this email so it is clearer, shorter, and more confident without sounding harsh. Keep all factual details unchanged. Email: [paste email]."

5. Brainstorm: "Generate 20 ideas for [project]. Constraints: [budget, audience, channel]. Group ideas by low effort, medium effort, and bold bets. Add one sentence explaining why each could work."

6. Decision memo: "Act as a strategy partner. Compare these options: [options]. Use criteria: cost, speed, risk, user impact, reversibility. End with a recommendation and the strongest counterargument."

7. Research questions: "I am researching [topic]. Create a research plan with the 10 questions I should answer, the evidence needed for each, and likely source types to check."

8. Source summary: "Summarize this source for a busy operator. Include thesis, key evidence, useful statistics, caveats, and what I should verify independently. Source: [paste text]."

9. Market scan: "Act as a market research assistant. Build a competitor scan for [category]. Columns: company, target customer, positioning, pricing signal, strongest claim, weakness, source needed."

10. Customer language: "Extract customer language from these reviews or calls. Group phrases by pain, desired outcome, objection, and buying trigger. Do not invent quotes. Text: [paste text]."

11. Content outline: "Create an SEO-friendly outline for an article targeting [keyword]. Audience: [audience]. Include search intent, H2s, examples to include, and what would make the article genuinely useful."

12. First draft: "Write a first draft of [asset] for [audience]. Use this outline: [outline]. Make it specific, practical, and free of generic filler. Ask for missing details before making claims."

13. Editing pass: "Edit this draft for clarity, structure, and usefulness. Do not change the meaning. Return: biggest issue, revised draft, and five suggested cuts. Draft: [paste draft]."

14. Tone match: "Rewrite this in the style of the example below. Preserve facts and structure, but match sentence length, directness, and level of detail. Example: [paste example]. Draft: [paste draft]."

15. Slide outline: "Turn this idea into a 7-slide outline. Each slide needs a title, one key point, supporting evidence needed, and speaker note. Idea: [paste idea]."

16. Explain simply: "Explain [concept] to a smart beginner. Use an analogy, then a practical example, then three common mistakes to avoid."

17. Learning plan: "Create a 14-day learning plan for [skill]. I can spend [time] per day. Include daily practice, a checkpoint, and a small project by day 14."

18. Code explanation: "Explain what this code does, where it may fail, and what tests would increase confidence. Keep the explanation accessible to a junior developer. Code: [paste code]."

19. Debugging: "Act as a careful debugging partner. Given this error, code, and expected behavior, list the most likely causes, the smallest next diagnostic step, and a safe fix. Error/context: [paste]."

20. Code review: "Review this change for correctness, security, edge cases, and missing tests. Prioritize real risks over style preferences. Diff: [paste diff]."

21. Data cleanup: "Turn this messy list into a clean table. Infer categories only when obvious, mark uncertain fields as unknown, and list cleanup rules used. Data: [paste data]."

22. Spreadsheet formula: "I need a spreadsheet formula for [goal]. Columns are [columns]. Explain the formula and include one example row."

23. Personal productivity: "Plan my day from this task list. Constraints: [meetings, energy, deadlines]. Group tasks into deep work, admin, quick wins, and defer. Task list: [paste]."

24. Responsible use check: "Review this AI-assisted output for risks. Check for unsupported claims, privacy concerns, bias, missing caveats, and places that need human review. Output: [paste]."

25. Model comparison: "I am testing AI models for [task]. Score this output from 1-5 on accuracy, completeness, clarity, usefulness, and risk. Explain the score and suggest one better follow-up prompt. Output: [paste]."

#### How to verify outputs

The most important skill in how to use generative ai is not prompt cleverness. It is verification. Treat AI output like a strong draft from a fast assistant: useful, but not automatically true. The higher the stakes, the more verification you need.

Run this checklist before you publish, send, or act on AI output. It takes a couple of minutes and it catches almost everything that goes wrong.
1. **Check the facts.** Identify every factual claim and check it against a primary source or a trusted internal document.
2. **Hunt for hidden assumptions.** Dates, prices, policies, customer segments, technical constraints, and legal requirements are where models quietly guess.
3. **Check the format.** Did the answer actually follow the structure you asked for, or did it drift back to prose?
4. **Remove invented certainty.** Add caveats wherever the evidence is incomplete, because the model will not.
5. **Check privacy.** Do not paste customer, employee, health, financial, credential, or proprietary data into a tool your organization has not approved.
6. **Get a second opinion.** Ask a different model, or the same model with a critique prompt, to find the weakest claim.
7. **Read it yourself.** A final human pass for tone, judgment, and context, which is the part no model can do for you.

For research, require citations or source notes and verify the cited source yourself. For writing, compare the draft to your voice and remove generic claims. For code, run tests and review the diff. For strategy, ask for counterarguments. For long documents, ask the model to quote or point to the section that supports each important conclusion, then check the source material directly.

#### Picking the right model for the job

You do not need to pick one AI model forever. The better habit is to match the model to the job, then compare outputs when quality matters. OpenAI, Anthropic, and Google all publish prompting guidance, and it points at the same practical idea. Different models suit different tasks, and the gap widens as you move between fast drafting, hard reasoning, long documents, coding, and multimodal work.

| The job | What to reach for | What to judge it on |
| --- | --- | --- |
| Writing, outlining, email, ideation | A general-purpose model | Specificity, tone control, and whether it followed your format |
| Coding | A model that explains tradeoffs | Does it ask for missing context, propose tests, and prefer a small fix to a rewrite |
| Research | Whichever workflow keeps sources traceable | Can you follow every claim back to something you can check |
| Long documents | A large-context model | Ask for extraction before synthesis, then verify against the original |
| Images, PDFs, screenshots | A multimodal model | Ask for structured extraction before interpretation |

A simple decision table: use a fast model for drafts and brainstorming, a stronger reasoning model for decisions and debugging, a long-context model for large source material, and a multimodal model for images or PDFs. Then A/B test the same prompt. If one answer is more accurate but another has better structure, ask the stronger model to revise using the better structure. That is often faster than trying to force one model to do every job perfectly. If you are still choosing where to start, the [best AI apps for beginners](https://whizi.io/resources/best-ai-apps-for-beginners) guide compares the main options by task.

#### Your first 15 minutes in Whizi

The easiest way to learn how to use AI responsibly is to run small, concrete comparisons. Start in Whizi with five prompts from this guide: one email rewrite, one meeting summary, one research plan, one decision memo, and one verification check. Use the same prompt across models and compare the outputs side by side.

Minute 1-3: choose one real task from your day, not a fake demo. Minute 4-6: paste the master prompt template and fill in persona, task, context, and format. Minute 7-10: run the prompt across models and mark which output is clearest, most useful, and least risky. Minute 11-13: ask the best output for a revision with one specific constraint, such as shorter, more evidence, or more direct. Minute 14-15: save the prompt pattern so you can reuse it. [AI workflow templates](https://whizi.io/resources/ai-workflow-templates) has seven of these routines already written out if you would rather start from one than build your own.

For research-heavy work, try the Founder Research Stack at /templates/founder-research-stack. For more learning paths, browse /resources. When you are ready to turn experimentation into a repeatable workflow, create an account at /register and compare plans at /pricing.

Key points:

- A beginner prompt template you can reuse for writing, research, planning, and analysis
- 25 copy-paste prompts grouped by real work scenarios
- A practical verification checklist so AI output becomes usable instead of risky

Checklist:

- Define the exact task before opening the chatbot
- Use Persona, Task, Context, and Format in every serious prompt
- Paste examples of good output when tone or structure matters
- Ask the model to list assumptions and missing information
- Verify factual claims against primary sources or internal documents
- Check privacy before pasting sensitive or proprietary information
- Run important prompts across more than one model
- Use long-context models for large source material and ask for extraction before synthesis
- Use multimodal models when the input includes images, screenshots, PDFs, or mixed media
- Save the prompts that work so AI becomes a repeatable workflow, not a one-off trick

FAQ:

Q: What is the easiest way to start using AI?
A: Start with one real work task, use the Persona, Task, Context, Format template, and ask for a structured output you can review. Good beginner tasks include rewriting an email, summarizing notes, outlining a document, or creating a decision table.

Q: How do I write better prompts?
A: Better prompts are specific about the role, task, background, constraints, and output format. Include examples when you care about tone, and ask the model to identify assumptions before it finalizes the answer.

Q: Can I trust AI answers?
A: You can use AI answers as drafts, summaries, and reasoning aids, but you should verify important claims. Check sources, review assumptions, protect private data, and use human judgment before publishing or making high-stakes decisions.

Q: Which AI model should a beginner use?
A: Beginners should use the model that performs best on their actual task. A practical approach is to run the same prompt across models, compare accuracy and usefulness, then save the prompt and model pairing that works best.

### How to use ChatGPT: a practical guide for real work

URL: https://whizi.io/resources/how-to-use-chatgpt/
Updated: August 2026

Quick answer: Use ChatGPT by opening chatgpt.com or the mobile app and typing a task-shaped prompt: say what you want done, paste the context it needs, and name the output format. Read the answer critically, then refine with a follow-up instead of starting over. The free plan covers most beginners.

How to use ChatGPT properly: setup, the free and paid tiers compared, a prompt pattern that works, 18 prompts by scenario, and how to check what it tells you.

#### Open it and send one real task

Go to chatgpt.com in any browser, or install the ChatGPT app on iOS, Android, Mac, or Windows. Signing in with an email, Google, Microsoft, or Apple account takes about a minute and gets you chat history, memory, and the ability to pick up a conversation on another device. If signup refuses you outright, [ChatGPT is not available in your country](https://whizi.io/resources/chatgpt-not-available-in-your-country) explains which of the two separate checks you hit. There is a logged-out mode with fewer features if you would rather not create an account, covered separately in [how to use ChatGPT without an account](https://whizi.io/resources/use-ai-without-an-account).

Once you are in, the interface is one text box. That simplicity is why people underuse it: with nothing to click, most beginners type something like "help me with my email" and get back something generic, then conclude the tool is overrated. The problem is almost never the model. It is that a vague request has no correct answer, so it returns the average of every possible answer.

Do this instead for your first message. Pick something you actually have to do today, not a demo. Paste the raw material into the chat, describe what you want done with it in one sentence, and say what shape the result should take. A first prompt that works looks like this: "Here are my notes from a client call. Turn them into a short follow-up email with a subject line, three bullet points on what we agreed, and one clear next step. Notes: [paste]."

Then read the answer and reply to it. This is the part beginners skip. ChatGPT keeps the whole conversation in view, so "shorter, and drop the compliment in the opening" gets you a better second draft in seconds. Starting a fresh chat and rewriting your prompt from scratch throws away the context you already gave it.

#### Free, Go, and Plus: what each tier gets you

You can use ChatGPT for free, and for a lot of people that is where the story ends. The free plan gives you the flagship model for a capped number of messages before dropping you to a lighter fallback model, plus small daily allowances for image generation, file uploads, and web search. For a student writing a few essays a month or someone drafting occasional emails, nothing above it is necessary.

The paid tiers are about headroom and access rather than raw intelligence, with one real exception: the deeper reasoning and multi-step research tools. Prices below are United States list prices as of August 2026, and OpenAI localizes them elsewhere.

| What you get | Free | Go ($8/mo) | Plus ($20/mo) |
| --- | --- | --- | --- |
| Everyday fast model | Capped, then a lighter fallback model | Roughly ten times the free allowance | Same reported cap as Go |
| Reasoning models for hard problems | Very limited | Limited allowance | Included, higher allowance |
| Deep research style reports | Small monthly allowance | Thin allowance, well below Plus | Significantly more |
| Agent style multi-step tasks | Not included | Not included | Included |
| Image generation | A few per day | Around ten times free | Higher caps, faster |
| File uploads and analysis | Capped | Around ten times free | Highest cap |
| Memory and context length | Shortest | Longer than free | Longest |
| Ads | Shown in test markets | Shown in test markets | Not shown |
| New model access | Last | After Plus | First |

Exact caps move often, sometimes month to month, and how they reset has itself changed during 2026. Treat any published number, including the ones above, as a shape rather than a spec sheet, and check the plan comparison inside your own account before paying. If you are already hitting a wall, [what to do when ChatGPT says you have reached your limit](https://whizi.io/resources/chatgpt-limit-reached) explains which of the three lookalike limit messages you are actually seeing.

The practical rule: stay on free until you hit a limit mid-task more than once a week. At that point, [ChatGPT Go vs Plus](https://whizi.io/resources/chatgpt-go-vs-plus) covers the cheaper upgrade, and [is ChatGPT Plus worth it](https://whizi.io/resources/is-chatgpt-plus-worth-it) runs the break-even math on the $20 tier. Both are more useful than a review, because the answer depends on your usage rather than on the product.

#### The prompt pattern that actually works

Learning how to write ChatGPT prompts does not require a course. Four elements cover almost everything, and you can hold them in your head: task, context, format, constraint.

**Task.** Lead with a verb and an object. Summarize, rewrite, compare, extract, outline, critique, plan, debug, classify, translate. "Help me with this" is not a task. "Rewrite this paragraph so a non-technical customer understands it" is.

**Context.** Paste the raw material rather than describing it. Add the audience, the goal, and anything the model would otherwise guess: dates, names, prices, constraints, what you already tried. Guessing is what produces the invented details people complain about, and most invented details exist because the prompt left a hole.

**Format.** This is the highest-leverage element and the one most often left out. Ask for a table, a checklist, an email with a subject line, five bullets, a two-paragraph summary, a decision memo, or labeled fields. Naming the format converts a wall of prose into something you can paste straight into your work.

**Constraint.** Length, tone, what to avoid, what must be included. "Under 150 words, no marketing language, keep every number unchanged" removes three rounds of editing before they happen.

A reusable shell: "Your task is to [verb + object]. Context: [paste material, audience, goal]. Output format: [shape]. Constraints: [length, tone, must include, must avoid]. If anything important is missing, ask before writing."

Two habits multiply the results. First, show an example of what good looks like when tone matters, because a sample of your own past writing controls voice far better than any adjective. Second, ask for options rather than an answer when you are still deciding: "give me three versions with different angles, then say which you would send and why."

#### 18 prompts grouped by what you are actually doing

Replace the bracketed parts, paste your own material, and treat the output as a first draft rather than a final one.

**Writing and email**
1. "Turn these notes into a follow-up email with a subject line, three bullets on what we agreed, and one next step. Notes: [paste]."
2. "Rewrite this so it is shorter and clearer without sounding blunt. Keep every fact unchanged. Text: [paste]."
3. "Here are two samples of my writing. Draft [asset] in the same voice: sentence length, directness, vocabulary. Samples: [paste]. Brief: [paste]."
4. "Give me three subject lines and three opening sentences for this email, ranging from formal to casual. Email: [paste]."

**Work and planning**
5. "Turn this goal into a one-week plan. Output a table with day, task, owner, deliverable, and the risk that would derail it. Goal: [goal]. Constraints: [team, deadline]."
6. "Summarize these meeting notes into decisions, open questions, owners, and deadlines. Flag anything ambiguous rather than resolving it yourself. Notes: [paste]."
7. "Compare these options on cost, speed, risk, and reversibility. End with a recommendation and the strongest argument against it. Options: [paste]."
8. "Read this message from a customer and draft a reply that acknowledges the specific complaint, states what we will do, and gives a date. Message: [paste]."

**Learning and explaining**
9. "Explain [concept] to someone smart who has never encountered it. Use one analogy, one worked example, then three mistakes beginners make."
10. "I am about to [do task] and have never done it before. List what I need to decide, what I need to gather, and the two places people usually get stuck."
11. "Quiz me on [topic] with ten questions of increasing difficulty. Ask one at a time, wait for my answer, and tell me what I got wrong before moving on."

**Documents and data**
12. "Read this document and give me a map first: sections, what each covers, and where the important claims sit. Then wait. Document: [paste or upload]."
13. "Extract every [obligation, date, price, name] from this text into a table with the exact quoted wording and where it appears. Do not summarize or paraphrase. Text: [paste]."
14. "Turn this messy list into a clean table. Mark anything uncertain as unknown rather than guessing, and list the rules you applied. Data: [paste]."

**Code**
15. "Explain what this code does, where it could fail, and what tests would raise confidence. Audience is a junior developer. Code: [paste]."
16. "Here is an error, the code, and what I expected. List the most likely causes ranked by probability, then the smallest diagnostic step for each. Do not propose a fix yet. [paste]."

**Personal**
17. "Plan my day from this task list. Group into deep work, admin, quick wins, and defer, and tell me what to drop if I lose two hours. List: [paste]."
18. "I need to have a difficult conversation about [situation]. Help me prepare: what they are likely to say, what I should not say, and three ways to open."

Save the ones that work. A prompt you reuse weekly is worth more than fifty you tried once, and pasting a saved prompt takes ten seconds against the two minutes of rewriting it from memory. [AI workflow templates](https://whizi.io/resources/ai-workflow-templates) has seven of those routines already written out, inputs and quality checks included.

#### Features worth turning on

**File uploads.** Drag a PDF, spreadsheet, image, or document into the chat and ask questions about it. This is the fastest win most people have not tried. Ask for extraction with quoted wording before you ask for interpretation, since a summary of a document you have not checked is a claim, not a fact.

**Web search.** ChatGPT can search the web and return links, which changes it from a model recalling patterns into something that can tell you what happened this week. The citations are worth opening rather than trusting, because a link attached to a sentence is not proof the link supports the sentence.

**Voice mode.** Speak instead of typing. Genuinely useful for thinking out loud, rehearsing a conversation, or working while your hands are busy, and it is much better than dictation for messy half-formed thoughts.

**Custom instructions.** A settings field where you describe who you are and how you want responses written, applied to every new chat. Filling this in once removes the same three corrections you have been typing repeatedly. Keep it short and concrete: what you do, what you are usually working on, and the response style you want.

**Memory.** ChatGPT can retain details across conversations. Convenient, and worth reviewing occasionally: the memory settings show what it has stored and let you delete individual entries. If you share a device or work with sensitive material, know what is in there.

**Projects.** Group related chats with shared files and instructions so you do not re-explain the same background every time you return to an ongoing piece of work. The same folders-plus-pinned-files pattern outside OpenAI is covered in [a ChatGPT Projects alternative](https://whizi.io/resources/chatgpt-projects-alternative).

One rule that overrides all of the above: do not paste customer records, employee data, health or financial information, credentials, or proprietary code into any account your organization has not approved. This has nothing to do with ChatGPT specifically and everything to do with the fact that pasted data leaves your control.

#### What ChatGPT is bad at, and how to check it

Being useful with ChatGPT means knowing where it fails, and the failures are predictable rather than random. It generates text that fits the pattern of a correct answer, which is a different thing from a correct answer, and the two are indistinguishable from tone alone. [What ChatGPT is](https://whizi.io/resources/chatgpt) explains the prediction mechanism underneath that in plain terms.

The reliable weak spots: **specific facts it was not given**, especially dates, prices, statistics, versions, and legal or medical detail that varies by jurisdiction. **Citations**, which can be plausible-looking and wrong, including real authors attached to papers that do not exist. **Arithmetic and counting** across a long document, where it approximates rather than counts. **Recent events**, unless search is actually on. **Anything about you or your organization** that you did not paste in. And **its own confidence**, which stays flat whether it is certain or guessing.

The checks are quick once they become habit. Before you send, publish, or act on anything: verify every number, date, name, and quote against the source. Open any citation that carries weight. Ask "what would make this wrong, and what did you assume?" and read the answer, because it is often more honest than the original. For documents, ask it to quote the passage supporting each important conclusion, then check that passage exists. For code, run it and read the diff.

A second-opinion pass costs almost nothing and catches a surprising amount: paste the output back and ask for the weakest claim in it, or send the same question to a different model and see whether the answers agree. Disagreement between two models is a reliable flag that the question is harder than it looks. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) explains why this is structural rather than a bug awaiting a patch.

#### When a different model does the job better

ChatGPT is the strongest all-round default and a fine place to do most of your work. It is not the best tool for every job, and knowing the handful of cases where something else wins saves more time than any prompt trick.

| If the job is | ChatGPT is | Consider instead |
| --- | --- | --- |
| Everyday questions, drafts, planning | The right default, broadest feature set | Nothing, stay here |
| Writing where voice must survive editing | Good for range and volume | Claude, which tends to need less cleanup on tone |
| Research where you must open the sources | Good synthesis, citations need checking | Perplexity, which answers with links attached |
| Very large documents or whole codebases | Comfortable across normal business documents | Gemini, at the million-token-class end |
| Production image work | Convenient for a quick illustration | Flux or Stable Diffusion for control and consistency |
| A decision that carries real cost | One good answer | Two models, then compare where they disagree |

The habit worth building is not switching allegiance. It is routing: noticing which tool wins for which kind of task and writing that down. Four or five observations from your own work beat any comparison article, including this one. [The best AI to use](https://whizi.io/resources/best-ai-to-use) has the task-by-task verdicts, [ChatGPT vs Claude vs Gemini](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini) compares the three head to head, and [how to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers the draft-with-one, review-with-another loop that catches the most errors. If your day happens inside Word, Excel, and Outlook, [Copilot vs ChatGPT](https://whizi.io/resources/copilot-vs-chatgpt) is the comparison that settles it.

None of that requires paying for several subscriptions. It requires being able to send the same prompt to more than one model when the answer matters, and noticing which one you keep reaching for.

#### Your first week

Week one is about attaching ChatGPT to work you already do, rather than finding new work for it. The people who abandon it after a month are usually the ones who used it for demos instead of tasks.

Days one and two: use it for one recurring task you find tedious. Summarizing notes, drafting a routine reply, cleaning up a list. Same task, same prompt, twice.

Days three and four: fill in custom instructions, then upload a real document and ask it questions. These two things change the experience more than anything else in the settings.

Days five and six: take something with a real deadline and use the full pattern, including the verification pass. Notice how much editing stands between its draft and something you would send, because that number, not a benchmark, is what the tool is worth to you.

Day seven: look back at what you used it for. The tasks that repeat become saved prompts. The tasks where the output disappointed you are the ones to try in another model before assuming AI cannot do them. That is the whole practice, and everything else is detail.

Key points:

- The single change that improves most answers is naming the output format, because it turns vague help into a deliverable you can use
- The free plan is genuinely enough for a large share of people, and the honest upgrade signal is hitting a limit mid-task more than once a week
- ChatGPT is confident whether or not it is right, so every workflow that matters needs a verification step attached to it

Checklist:

- Start with one real task from today, not a test question
- Paste the source material instead of describing it
- Name the output format in every prompt: table, email, bullets, checklist, memo
- Reply to refine rather than starting a new chat and rewriting the prompt
- Fill in custom instructions once so you stop repeating the same corrections
- Stay on the free plan until you hit a limit mid-task more than once a week
- Verify every number, date, name, quote, and citation before you act on the answer
- Keep customer, employee, health, financial, and credential data out of the chat box
- Send the same prompt to a second model when the decision carries real cost
- Save the prompts that work so they become a workflow instead of a one-off

FAQ:

Q: How do I start using ChatGPT?
A: Open chatgpt.com or install the app, sign in, and type a task-shaped prompt: what you want done, the material it needs, and the format you want back. Pick something real from your own day rather than a test question, then refine the answer with a follow-up instead of starting over.

Q: Can I use ChatGPT for free?
A: Yes. The free plan gives you the flagship model for a capped number of messages before switching you to a lighter fallback, plus daily allowances for image generation, file uploads, and web search. That covers most casual use. Paying buys headroom, faster access to new features, and the deeper reasoning and research tools.

Q: How do I write better ChatGPT prompts?
A: Use four elements: task, context, format, and constraint. Lead with a verb, paste the actual material rather than describing it, name the shape you want back, and state length, tone, and what to avoid. Naming the output format is the single change that improves most answers.

Q: What can ChatGPT do?
A: It drafts and rewrites text, summarizes notes and documents, extracts structured data, explains concepts, plans work, compares options, writes and reviews code, translates, reads uploaded files and images, generates images, and searches the web. It is a drafting and reasoning tool, so the output is a starting point rather than a finished answer.

Q: Is ChatGPT accurate, and how do I check it?
A: It is fluent whether or not it is right, and its tone does not change when it is guessing. Verify every number, date, name, quote, and citation against a source. Ask what it assumed and what would make the answer wrong. For anything that carries cost, send the same question to a second model and look at where they disagree.

Q: Should I use ChatGPT or a different AI?
A: ChatGPT is the best single default and covers most work well. Claude usually needs less cleanup on tone-sensitive writing, Perplexity is better when you need sources you can open, and Gemini leads at the very large document end. Rather than picking one forever, note which tool wins for which task and route accordingly.

### How to use Claude: a task-first guide for beginners

URL: https://whizi.io/resources/how-to-use-claude/
Updated: August 2026

Quick answer: Use Claude by opening claude.ai or the Claude app, signing in free, and typing a task with three things attached: the context Claude needs, the constraints it must respect, and the output format you want. Upload documents rather than pasting them, correct the first draft, and start a fresh conversation when the topic changes.

A practical guide to using Claude: signing up, the prompt pattern that works, 18 copy-paste prompts, documents and Projects, how the limits behave, and when another model is the better tool.

#### Getting started, and what it costs

Claude is Anthropic's AI assistant. It runs in a browser at claude.ai and in mobile and desktop apps, and using it is the same in all of them: you type a request, it answers, and the conversation stays there so you can push back on the answer. Creating an account is free and takes a minute. Anthropic gates accounts by country rather than by payment, so if signup fails outright, check the supported regions list before assuming something is broken.

Start on the free plan even if you already intend to pay. Free is a working product rather than a demo. It drafts and edits, reads documents you upload, searches the web, remembers things you told it in earlier conversations, and lets you create up to five Projects. A week on free tells you two things a review cannot: whether you like the writing, and whether you personally run into the limits.

| What you get | Free | Pro (about $20 a month) |
| --- | --- | --- |
| Writing, editing, and everyday questions | Yes | Yes |
| Web search | Yes | Yes |
| Memory across conversations | Yes | Yes |
| Uploading documents and images | Yes | Yes |
| Projects | Yes, capped at five | Uncapped, with deeper retrieval over uploaded material |
| Choice of Claude models | A default model | More models, including the stronger ones |
| Claude Code, the terminal coding agent | No | Yes, at limited usage |
| Usage before you are asked to wait | A session window that resets every five hours | Substantially more, plus a weekly ceiling aimed at very heavy use |
| Priority access when demand is high | No | Yes |

Prices are Anthropic's United States list prices as of August 2026 and differ by country, with heavier Max tiers above Pro. If you are weighing the upgrade specifically, [is Claude free good enough](https://whizi.io/resources/is-claude-free-good-enough) has a one-week test and [is Claude Pro worth it](https://whizi.io/resources/is-claude-pro-worth-it) covers what the paid plan changes and what it still will not do.

#### What Claude is strongest at

Three things, consistently, and knowing them is most of what separates a good first week from a frustrating one.

**Writing that needs less editing.** Give Claude a rough memo, a difficult email, or a set of notes and the draft that comes back usually needs fewer fixes than the alternatives before you would send it. It holds a register, follows nuanced style instructions, and avoids the padding rhythms that make AI text easy to spot. This is the single clearest reason people choose it.

**Reading long and dense material.** Contracts, filings, research papers, board packs, interview transcripts. Claude is careful with qualifications rather than flattening everything into a tidy summary, and it is more willing than most assistants to say a document does not answer your question instead of inventing something that sounds like an answer.

**Reviewing code rather than only writing it.** Ask what could break, where the hidden coupling is, and which tests would actually prove the fix works, and the answers tend to be specific. It also pushes back on an approach instead of cheerfully implementing whatever you asked for, which is the useful behaviour in a reviewer.

Two more habits worth knowing early. Claude takes correction well: tell it the third paragraph is wrong and it usually adjusts properly rather than apologising and repeating itself with new adjectives. And it follows long briefs, so a nine-point list of constraints tends to come back with all nine honoured.

What it does not do: generate images, and voice is well behind ChatGPT and Gemini. Very large context work, whole codebases or a year of transcripts, is Gemini territory. [Claude vs ChatGPT](https://whizi.io/resources/claude-vs-chatgpt) has the head-to-head version of all of this with prompts you can run yourself.

#### The prompt pattern that works

Anthropic's own prompt engineering guidance and the experience of most heavy users land in the same place: be explicit, give context, and say what the output should look like. Four parts, in this order.

**Task.** The verb, stated plainly. Draft, rewrite, summarize, extract, compare, review, critique, plan, explain. "Help me with this" is not a task. "Rewrite this paragraph so it is 40 percent shorter without losing the caveat about delivery dates" is.

**Context.** Who it is for, what they already know, what you are trying to achieve, and the raw material. Claude guesses less when you tell it more, and it is unusually good at using background that other assistants ignore. Include the awkward parts: the deadline that slipped, the person who will be annoyed, the constraint you cannot change.

**Constraints.** What it must avoid as much as what it must include. Length, tone, banned phrases, facts that cannot change, claims you cannot make. "Keep every number exactly as I have it" and "do not add flattery" each fix a specific failure you would otherwise edit out by hand.

**Format.** A table, a memo, an email with a subject line, a checklist, five bullets, a two-paragraph answer. Naming the shape turns a vague reply into something you can use, and it is the cheapest single improvement available.

Copy-paste skeleton: `Your task is to [task]. Context: [audience, goal, background, source material]. Constraints: [tone, length, must include, must avoid, facts to preserve]. Format: [table/email/memo/checklist]. Before you write, list anything you need from me that is missing.`

That last sentence earns its place. Asking Claude what it is missing before it drafts surfaces the assumption it was about to make, and it is faster to answer one question than to rewrite three paragraphs built on a guess. [Prompt engineering for beginners](https://whizi.io/resources/prompt-engineering-beginners) goes deeper on the technique if you want it.

#### 18 prompts to start with

Replace the bracketed parts, paste your own material where it says paste, and keep the ones that work. Saving the prompts that earn their keep is what turns Claude from a novelty into a tool you actually reach for.

**Writing and email**
1. `Here is a rough version of what I want to say. Rewrite it so it sounds professional but still like a person wrote it. Keep every fact exactly as I have it. No flattery, no filler openers. Under 150 words. Draft: [paste].`
2. `Write a message telling [person] that [difficult thing]. Direct but not cold. Do not over-apologise, and do not add hedging that invites an argument about the fact itself. End with one clear next step and a date.`
3. `Rewrite this in the voice of the sample below. Match sentence length, directness, and how much it explains. Do not change any fact or the order of the argument. Sample: [paste]. Draft: [paste].`
4. `Tell me the three weakest things about this piece of writing. Quote each one. Do not rewrite it, and do not tell me what is good about it. Draft: [paste].`

**Documents and reading**
5. `Read the attached document and give me: the five things that matter most, anything that would cost me money or create an obligation, and anything important that is vague or missing. Quote the exact text for each point so I can check it.`
6. `Compare these two documents. Produce a table of every place they disagree, with the wording from each side and which one is more favourable to me. Do not summarize the parts where they agree.`
7. `From this transcript, extract every decision, every owner, and every deadline into a table. Mark anything that was discussed but not decided as open. Do not infer owners that were not named. Transcript: [paste].`
8. `Here is a report. Before summarizing anything, give me a map: what each section covers and how long it is. I will tell you which sections to read closely.`

**Thinking and deciding**
9. `I am deciding between [options]. Score them against these criteria: cost, speed, risk, reversibility, and effect on [group]. End with a recommendation and the strongest argument against it.`
10. `Argue the opposite of the position in this document as convincingly as you can, using only evidence that is actually in it. Then tell me which of your counterarguments I should take most seriously. Document: [paste].`
11. `Here is my plan. List the assumptions it depends on, ranked by how much damage it does if that assumption is wrong. Do not suggest improvements yet. Plan: [paste].`

**Code**
12. `Review this diff for correctness, edge cases, and regression risk. Prioritise things that could actually break in production over style preferences. Tell me which tests would prove the fix works. Diff: [paste].`
13. `Explain what this code does, where it is likely to fail, and what a new developer would misunderstand about it. Keep it readable for someone who has never seen this codebase. Code: [paste].`
14. `Here is an error, the relevant code, and what I expected to happen. Give me the three most likely causes in order, the smallest diagnostic step for each, and the safest fix. Do not refactor anything I did not ask about. [paste].`
15. `I am considering [approach] for [problem]. Argue against it. What breaks at ten times the current scale, and what would I regret in six months?`

**Learning and everyday**
16. `Explain [concept] to someone smart who has never encountered it. One analogy, then a concrete example, then the three mistakes beginners make. No jargon unless you define it in the same sentence.`
17. `I have [time] a day for [number] weeks to learn [skill]. Build a plan with daily practice, a checkpoint at the halfway mark, and one small finished project at the end.`
18. `Here is my task list and my calendar for tomorrow. Group the tasks into deep work, admin, quick wins, and things to drop, and tell me what will not fit. Do not be optimistic about how long things take. [paste].`

#### Documents, Projects, and Artifacts

Three features change how Claude works day to day, and beginners usually find them late.

**Upload documents instead of pasting them.** Attaching a PDF, spreadsheet, or image gives Claude the actual file rather than whatever survived a copy and paste, which preserves structure and stops tables turning into soup. For long files, ask for a map of the document first, then ask for extraction from the sections you care about, then check the quotes against the original. That sequence catches more than asking for a summary in one go.

**Projects hold context so you stop repeating yourself.** A Project is a workspace with its own uploaded material and its own instructions. Put a style guide, a brand voice document, and three examples of good past work into one, and every conversation started inside it already knows them. Free accounts can create up to five; paid plans remove the cap and add deeper retrieval over the material you upload. For anything you produce repeatedly, this is the feature that changes your habits. The equivalent pattern outside Anthropic, folders of chats with pinned files and their own instructions, is covered in [a ChatGPT Projects alternative](https://whizi.io/resources/chatgpt-projects-alternative).

**Artifacts render work beside the conversation.** When Claude produces a document, a piece of code, or a small interactive thing, it appears in a panel next to the chat and updates in place as you ask for changes, rather than being reprinted as another wall of text. Iterating on a draft this way is much easier than scrolling back through a chat log to find the version you liked.

One habit that costs nothing and helps constantly: start a new conversation when the topic changes. Claude re-reads the whole conversation on every turn, so a thread that has drifted through four unrelated subjects is both slower and more expensive than four short ones, and the answers get worse as the irrelevant history piles up.

#### The limits, and how they actually behave

Usage is measured in tokens rather than in messages, which is the fact that explains almost every confusing experience people have. A document is re-processed on every turn of the conversation it sits in, so one question about a 200-page PDF discussed across thirty messages costs vastly more than thirty short separate questions. Two people who sent the same number of messages can hit completely different walls.

Free accounts get a session window that resets every five hours, and how far it stretches also depends on how busy Anthropic's servers are at that moment. Paid plans get substantially more, plus a weekly ceiling aimed at very heavy use. Anthropic does not publish a fixed message count for any tier, and it moves with demand, so treat any article quoting a specific number, including the numbers you will find in forum threads, as a guess.

What drains a session fastest, in order: long documents inside long conversations, extended thinking on hard problems (the reasoning tokens count even though you only see a summary), large code files pasted repeatedly instead of referenced once, and very long threads that never get restarted.

The fixes are behavioural and they work. Start a new conversation when the topic changes. Attach reference material to a Project once instead of pasting it into every chat. Send the twenty relevant pages rather than the whole report. When a session gets long, ask for a written handoff summary and start fresh with it. People who work this way rarely meet the cap; people who keep one enormous conversation open all day meet it constantly.

Two other limits worth knowing. Claude's built-in knowledge stops at its training date, so anything recent needs web search turned on or the source pasted in, and it will answer confidently either way. And a large context window is not a promise of perfect attention across all of it, which is why verifying quotes against the original matters. [What is a context window](https://whizi.io/resources/what-is-a-context-window) explains the mechanism in plain terms.

#### When a different model is the better tool

A limit you hit while pushing Claude at work it is not built for is a wrong-tool problem, and more quota does not fix it. This table is the short version of where each job goes.

| The job | Reach for | Why |
| --- | --- | --- |
| A draft that has to sound like a person wrote it | Claude | Fewest edits between the first draft and something you would send |
| Reading a dense contract, filing, or research packet | Claude | Careful with caveats, and willing to say a document does not answer the question |
| Reviewing code and pressure-testing an approach | Claude | Specific about risk and coupling rather than agreeable |
| Twenty headline variants or five versions of a page | ChatGPT | Range and speed rather than polish |
| A whole repository, twenty files, or a year of transcripts | Gemini | Context window in a different class |
| An answer with links you can open right now | Perplexity or a search-first tool | Checkable in seconds rather than minutes |
| Generating or editing an image | ChatGPT, Gemini, or a specialist model | Claude reads images but does not generate them |
| Talking to an assistant hands free | ChatGPT or Gemini | Voice is not where Anthropic has invested |

The pattern that beats picking a favourite is drafting with one model and reviewing with another. Write in Claude, paste the result into a different model and ask for the weakest claim, then take the criticism back to Claude to fix. That loop is covered properly in [how to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together), and a broader task-by-task ranking is in [the best AI to use](https://whizi.io/resources/best-ai-to-use).

Running that loop across providers usually means several accounts, several histories, and several bills at roughly $20 each, which [the cheapest way to use ChatGPT and Claude together](https://whizi.io/resources/cheapest-way-to-use-claude-and-gpt) prices out route by route. A multi-model workspace like Whizi puts Claude alongside GPT and Gemini behind one login and one conversation history, which suits people who genuinely switch. It is the wrong choice if you want Claude Code, Projects with deep retrieval, or first access to Anthropic's new features, all of which live in Anthropic's own product.

Key points:

- The free plan is enough to learn on, and enough for most short tasks, so there is no reason to pay before you know what you would be paying for
- A four-part prompt pattern, 18 copy-paste prompts, and the upload habit that makes document work far more accurate
- The two things that actually stop people: a session window that resets every five hours, and reaching for Claude on jobs it is not built for

Checklist:

- Create a free account and spend a week on it before paying for anything
- Write the task as a verb, not as a topic
- Give Claude the awkward context: the deadline, the constraint, the person who will object
- Say what the output should look like before it starts writing
- Ask what it is missing before it drafts, rather than rewriting after a bad guess
- Upload documents as files instead of pasting them, and ask for a map before a summary
- Put recurring reference material into a Project once instead of into every chat
- Start a new conversation when the topic changes
- Verify quotes and numbers against the original document before anything leaves your desk
- Send the job somewhere else when it needs images, voice, live sources, or an enormous context window

FAQ:

Q: How do I start using Claude?
A: Go to claude.ai or install the Claude app and create a free account. Type a task with the context it needs and the output format you want, then correct the first draft rather than accepting it. Upload documents as files instead of pasting them, and start a fresh conversation when you move to a different topic.

Q: Do I need to pay to use Claude?
A: No. The free plan writes, edits, reads uploaded documents, searches the web, remembers context across conversations, and allows up to five Projects. What you get with Pro is more usage before the session cap stops you, access to more models, Claude Code, and priority access when demand is high. Spend a week on free first to find out which of those you actually need.

Q: What is Claude best at compared to other AI tools?
A: Three things: drafts that need less editing before you would send them, careful reading of long or dense documents, and code review that names real risks instead of agreeing with you. It does not generate images, its voice features lag behind ChatGPT and Gemini, and Gemini handles much larger context.

Q: How do I use Claude Projects?
A: A Project is a workspace with its own uploaded files and its own instructions, so every conversation started inside it already knows that material. Put a style guide, reference documents, and examples of good past work into one, then work inside it instead of re-pasting context. Free accounts can create up to five; paid plans remove the cap and add deeper retrieval.

Q: Why does Claude tell me to wait?
A: Free accounts get a usage window that resets every five hours, and how far it stretches depends partly on how busy the service is at that moment. Usage is counted in tokens rather than messages, so a long document discussed across many turns drains it far faster than many short questions. Shorter conversations and partial uploads usually fix it without paying.

Q: How do I write better Claude prompts?
A: State the task as a verb, give the context including the awkward parts, list the constraints it must respect, and name the output format. Then add one line asking Claude to tell you what information is missing before it drafts, which surfaces the assumption it was about to make.

### How to use Gemini: a task-first guide for beginners

URL: https://whizi.io/resources/how-to-use-gemini/
Updated: August 2026

Quick answer: Use Gemini by opening gemini.google.com or the Gemini app and signing in with a Google account, then typing a task rather than a keyword. Attach the file, image, or Drive document you want it to work from, name the output format you need, and check any facts before you use the answer.

How to use Google Gemini from the first prompt: signing in, free versus Google AI Pro, the prompt pattern that works, 18 prompts by scenario, and Gemini in Gmail and Docs.

#### Getting started: your first Gemini session

Gemini runs in three places, and the one you pick changes almost nothing about how you prompt it. The web app at gemini.google.com is the fullest version. The Gemini app on Android and iOS is the same assistant with camera, photo, and voice input attached. Gemini features also appear inside Gmail, Docs, Sheets, Drive, and Meet for accounts that have them switched on, which is covered further down.

1. **Open gemini.google.com** in any browser, or install the Gemini app from the Play Store or the App Store.
2. **Sign in with a Google account.** A personal Gmail address works. A work or school account depends on what your administrator has enabled, and some Workspace accounts have Gemini turned off entirely.
3. **Type a task in the box at the bottom.** Not a keyword, a task. "Summarize" beats "summary", and a sentence beats both.
4. **Attach anything the answer depends on** with the upload control: a PDF, an image, a screenshot, a spreadsheet, or a file from Drive.
5. **Read the answer, then push back.** Follow-up messages stay in the same conversation, so "shorter", "more direct", or "add the numbers from page 4" are complete instructions on their own.

Two things surprise people on day one. The first is that Gemini keeps a chat history tied to your Google account, so you can reopen a conversation later and it still has the context. The second is that your first answer is rarely the good one. The productive pattern is a first prompt that sets the job, then two or three corrections. That takes less time than trying to write one perfect prompt.

If you have never used any AI assistant before, [how to use AI](https://whizi.io/resources/how-to-use-ai) covers the general habits that apply to all of them, and this guide assumes none of it. If you keep typing search queries into it out of habit, [AI vs Google search](https://whizi.io/resources/ai-vs-google-search) gives a simple rule for which box a question belongs in.

#### Free Gemini versus Google AI Pro

You can use Gemini for free, permanently, with nothing more than a Google account. The free tier handles ordinary drafting, summarizing, explaining, brainstorming, and file questions, and a lot of people never hit a wall in it. Paying buys higher limits, priority access to the newest models, more generous file handling, and the Gemini features inside Google Workspace apps.

| What you get | Free Gemini | Google AI Pro |
| --- | --- | --- |
| Price | $0 | $19.99 a month in the United States |
| Everyday chat, drafting, and questions | Yes | Yes |
| Newest and slower-thinking models | Available with tighter caps | Higher limits and earlier access |
| File, image, and PDF uploads | Yes, with smaller limits | Larger files and more of them |
| Gemini inside Gmail, Docs, and Drive | Partial and account dependent | Included |
| Deeper research runs and video generation | Capped or unavailable | Included, still metered |
| Storage | The free 15 GB Google account allowance | An expanded Google One allowance |

The row people miss is the last one. Google AI Pro is sold as a Google One membership, so the storage and the AI features are the same subscription and cancelling one cancels both. That matters later rather than now, and the detail is in [how to cancel Google AI Pro](https://whizi.io/resources/how-to-cancel-google-ai-pro). Google also sells a cheaper AI Plus tier at $4.99 a month in the United States, which sits between free and Pro on limits. Prices are United States list prices as of August 2026 and differ by country.

A reasonable rule for deciding: stay free until you hit a limit in the middle of real work more than about once a week. Running out of a model allowance while you are trying to finish something is the signal to upgrade. Curiosity is not.

#### What Gemini is strongest at

Gemini is a general assistant and it will do almost anything the others do. Three areas are where it tends to earn the first attempt rather than the second.

**Large inputs.** Google documents context windows of one million tokens and above on several Gemini models, and the long-context guidance is aimed at exactly the case where you hand over a lot of material at once: a folder of notes, a long transcript, a full specification, many documents together. When the problem is keeping everything visible in one prompt, start here. The caveat is real though: a large stated window is not a promise of even attention across all of it, so ask for extraction with section references before you ask for conclusions.

**Mixed media.** Screenshots, photos, charts, scanned pages, and PDFs are first-class inputs rather than an afterthought. In the mobile app the camera is part of the prompt, which makes "what is this error message" or "what does this label say in English" a two-second job. Ask for a description of visible evidence before interpretation, and you get far fewer confident inventions.

**Anything already inside Google.** If your source material lives in Drive, Docs, and Gmail, Gemini removes the copy-and-paste step that quietly eats the time savings. This is the least glamorous advantage and often the largest one in practice.

For the head-to-head detail behind those claims, [Gemini vs ChatGPT](https://whizi.io/resources/gemini-vs-chatgpt) covers the document and multimodal comparison, and [Claude vs Gemini](https://whizi.io/resources/claude-vs-gemini) covers long context against writing quality.

#### The prompt pattern that works

Most disappointing Gemini answers come from prompts that are missing one of four things. Supply all four and the quality jump is larger than any model upgrade.

**Context first.** Put the source material and the background at the top, then the instruction. Say who the reader is, what they already know, and what decision the answer feeds. Google's own long-context guidance points the same direction: give the model the material up front rather than dribbling it in.

**One clear task verb.** Summarize, compare, extract, rewrite, plan, critique, translate, classify. "Help me with this" is not a task. "Rewrite this in 150 words for a customer who is already annoyed" is.

**Constraints.** Length, tone, what to include, what to leave out, and what to do when information is missing. The most useful constraint of all is a single line: "If something is not in the source, write Not found rather than guessing."

**A named output format.** Table, checklist, email with a subject line, memo with headings, five bullet points, a two-column comparison. Formats turn a wall of prose into something you can actually use.

A reusable shell: "Here is [the source material]. Acting for [audience], [task verb] it into [format]. Include [must include]. Leave out [must avoid]. If a detail is not in the source, mark it Not found rather than inferring it. Finish with the three things I should check myself."

Then iterate. Gemini responds well to short corrections in the same thread, and "cut it in half", "less formal", "add the dates", and "show me where in the document that came from" all work as complete messages. If you want the deeper version of this, [prompt engineering for beginners](https://whizi.io/resources/prompt-engineering-beginners) goes further.

#### 18 Gemini prompts by scenario

Replace the bracketed parts, attach the file where one is mentioned, and expect to run a follow-up or two. These are starting points rather than magic words. The translation prompt repays extra care in particular, and [how to use AI to translate](https://whizi.io/resources/use-ai-to-translate) covers keeping tone and register intact across languages.

**Everyday and personal**

1. "Explain [topic] to me as if I know nothing about it. Use one analogy, one concrete example, and three mistakes beginners make."

2. "I have [ingredients] and [time] minutes. Give me three things I can cook, with steps, and tell me which is easiest."

3. "Translate this into [language] and keep the tone polite and slightly formal. Then list any phrase that does not translate cleanly and explain why. Text: [paste]."

4. "Here is a photo of a letter I received. Tell me what it says, what it is asking me to do, and by when. Flag anything that looks like a deadline or a payment demand."

**Documents and files**

5. "Read the attached PDF and give me a document map first: sections, tables, dates, and the main claims. Do not summarize yet."

6. "From the attached document, extract every number, date, and named commitment into a table with columns for item, value, where it appears, and confidence. Write Not found where a field is missing."

7. "Compare these two attached documents. Return a table of where they agree, where they conflict, and what one covers that the other does not."

8. "Summarize this 60-page report for someone who has eight minutes. Structure: what it says, what it assumes, what is weak, and what I should verify."

**Gmail and Workspace**

9. "Draft a reply to the email below. Goal: [outcome]. Tone: warm and direct, not apologetic. Include a clear next step and a subject line. Email: [paste]."

10. "Summarize this email thread into decisions made, open questions, who owes what, and any deadline mentioned. Thread: [paste]."

11. "Turn these meeting notes into a status update for [audience] with three sections: progress, risks, next steps. Keep it under 250 words. Notes: [paste]."

12. "Take the attached spreadsheet and describe what is actually in it: columns, obvious data quality problems, and three questions the data could answer."

**Research and decisions**

13. "I am deciding between [option A] and [option B] for [purpose]. Build a comparison table using cost, effort, risk, and reversibility, then give a recommendation and the strongest argument against it."

14. "Research [topic] and give me the answer with the sources you used. For each claim, tell me which source supports it, and mark anything you are inferring rather than reading."

15. "Give me the ten questions I should be able to answer about [topic] before I make a decision, plus what evidence would answer each one."

**Images, screenshots, and visual work**

16. "Here is a screenshot of an error. Describe only what is visible first, then list the likely causes in order, then the smallest thing I should try."

17. "Read the chart in this image. Give me the underlying numbers as a table, then tell me what the chart makes look bigger than it is."

18. "Generate an image of [subject] in [style], [aspect ratio]. Then make three variations that change only the [background / color / angle]."

Whichever of these you use, the finishing move is the same: ask "what in this answer would you not stake anything on", then check those parts yourself.

#### Using Gemini in Gmail, Docs, and Drive

The Workspace integration is the part of Gemini that has no real equivalent elsewhere, and it is also the part with the most account-dependent behavior. Availability varies by plan, by country, and by whether an administrator has enabled it, so the honest answer to "why do I not see the button" is usually the account rather than the feature.

Where it shows up when it is on: a Gemini side panel in Gmail, Docs, Sheets, Slides, and Drive; a "Help me write" control in Gmail compose and in Docs; summary generation on long threads; and Gemini in Meet for notes. In the side panel you can reference your own files directly, which is the actual point of the integration.

The move that makes it worth having: reference a file by name in the prompt instead of pasting its contents. In the Docs or Gmail side panel, ask something like "Using the Q3 planning doc in my Drive, draft a reply to this email confirming the dates and flagging the two risks it lists." That single prompt would otherwise be three copy-and-paste steps and a lost afternoon.

Two cautions worth holding on to. First, drafted email is still your email once you press send, and Gemini will confidently agree to a commitment you did not intend to make, so read the draft rather than skimming it. Second, work accounts have data handling rules that personal accounts do not. Check your organization's policy before pasting customer, employee, health, or financial data into any assistant, Gemini included.

#### When another model does the job better

Gemini is not the right first attempt for everything, and knowing where it is second-best is more useful than any feature list. The leaders have separated by task, and they have stayed separated.

| The job | Try Gemini first if | Try something else first if |
| --- | --- | --- |
| Everyday drafting, planning, questions | You already work inside Google apps all day | You want the broadest feature set and tooling, where ChatGPT is the safest single default |
| Polished writing and editing | The draft starts from documents already in Drive | Voice and restraint matter most, where Claude usually leaves less to clean up |
| Long documents and PDFs | The source material is genuinely large and needs to be seen at once | The document is short and what you need back is polished prose, not extraction |
| Coding | The relevant context is spread across a large repo or a long specification | You want a bounded fix, tests, or a careful review of a patch |
| Research with sources | You want synthesis across many sources and will open the links yourself | Every claim needs a citation attached, where a search-first tool is built for it |
| Images | You want generation and editing inside the same conversation | The image ships, where specialist models lead on control and consistency |

The habit that beats picking a favorite: send the same real task to two models and note which one you had to fix less. Three or four of those notes is a routing rule, and a routing rule is worth more than any comparison article, including this one. [The best AI to use](https://whizi.io/resources/best-ai-to-use) has the task-by-task verdicts, and [how to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers the draft-with-one, review-with-another loop in detail.

One thing does not change with the model. Gemini can state something wrong in the same tone it uses for something right, because confident wrong answers are structural rather than a bug awaiting a patch. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) explains the mechanism, and the practical defense is short: verify numbers, dates, quotes, and anything you are about to send.

Key points:

- Gemini is free to use with any Google account, and the free tier covers most everyday drafting, summarizing, and question answering
- The one habit that changes output quality most is attaching the real source material and naming the format you want back
- Gemini is at its strongest on large files, mixed media, and work that already lives in Gmail, Docs, and Drive

Checklist:

- Sign in at gemini.google.com or in the Gemini app with the Google account you want the history on
- Write a task with a verb in it rather than a search-style keyword
- Attach the real file, screenshot, or Drive document instead of describing it
- Name the output format you want: table, checklist, email, memo, or bullets
- Add the line "if it is not in the source, write Not found" to any extraction prompt
- Ask for a document map and extraction before you ask for conclusions on a long file
- Correct the answer in the same thread rather than starting over
- Check numbers, dates, and quotes yourself before sending or publishing anything
- Stay on the free tier until you hit limits during real work more than once a week
- Send the same task to a second model when the decision actually matters

Steps:

1. Open Gemini and sign in: Go to gemini.google.com in a browser, or install the Gemini app on Android or iOS, and sign in with the Google account you want the chats saved to.
2. Type a task, not a keyword: Describe the job in a sentence: who it is for, what you want back, and how long it should be. A search-style phrase gets a search-style answer.
3. Attach the source material: Use the upload control to add a PDF, image, spreadsheet, or Drive file so Gemini works from your material instead of guessing at it.
4. Name the output format: Ask for a table, a checklist, an email with a subject line, or a memo with headings. Format instructions are the fastest quality improvement available.
5. Check the answer before you use it: Verify numbers, dates, quotes, and any claim you plan to send or publish. Ask Gemini to point at the part of the source that supports each conclusion.

FAQ:

Q: How do I use Gemini for free?
A: Open gemini.google.com or the Gemini app and sign in with any Google account. The free tier covers everyday chat, drafting, summarizing, and file uploads, with lower limits on the newest models and smaller file allowances. There is no trial clock, so you can stay on it indefinitely.

Q: Do I need a Google account to use Gemini?
A: Yes. Gemini signs you in with a Google account, which is also where your chat history is saved. A personal Gmail address works. A work or school account depends on whether your administrator has enabled Gemini, which is the usual reason the features do not appear.

Q: How do I use Gemini in Gmail?
A: Where it is enabled, Gemini appears as a side panel in Gmail and as a "Help me write" control in the compose window. You can ask it to draft a reply, summarize a long thread, or pull details from a file in your Drive. Availability depends on your plan, country, and administrator settings.

Q: Can Gemini read PDFs, images, and spreadsheets?
A: Yes. Upload the file directly or reference one from Drive. For long or dense files, ask for a document map and a structured extraction with section references before asking for a summary, then verify the numbers and quotes against the original yourself.

Q: What is Gemini best at compared to ChatGPT and Claude?
A: Gemini tends to lead on very large inputs, mixed media like screenshots and charts, and anything already living in Gmail, Docs, and Drive. ChatGPT is the broader all-round default, and Claude is the usual first pick for polished writing and careful review.

Q: How do I get better answers from Gemini?
A: Give it the source material up front, use one clear task verb, set constraints including what to do when information is missing, and name the output format. Then correct the first answer in the same thread instead of rewriting the prompt from scratch.

### How to use multiple AI models together (four patterns that work)

URL: https://whizi.io/resources/how-to-use-multiple-ai-models-together/
Updated: August 2026

Quick answer: To use all the major AI models you have two routes: open a separate account with each provider, or use one multi-model workspace that holds them in a single conversation. Either way, split the work across them. Draft with the strongest writer, critique with a different model, and verify wherever two models disagree.

Four practical workflows for using more than one AI model on the same task, including the critique pass and the disagreement check that catches confident mistakes.

#### Why one model is not enough

The usual argument for using several AI models is that each is better at something, so you should pick the best one per task. True, but it is the least interesting reason.

The better reason is that **a model cannot see its own blind spots.** Ask a model to check its own work and it will mostly confirm it, because the same weights that produced the error also produced the confidence. Ask a different model, trained differently on different data with different failure modes, and it will spot things the first one could not. This is not a trick, it is the same reason you ask a colleague to read your email before you send it to the client.

The second reason is practical: model leadership keeps changing. Whoever is best at coding this quarter may not be next quarter. A workflow built on one vendor has to be rebuilt each time the ranking shifts. A workflow built on the idea that you switch models has no such problem.

What follows is four patterns, in order of how much value they add relative to effort. You do not need all four. Most people get almost everything from the first two.

#### Pattern 1: draft, then critique with a different model

The highest value pattern, and the one almost nobody uses. Write with one model, review with another.

It works because generating and evaluating are different jobs. A model producing text is optimising for a coherent continuation. A model evaluating text it did not write has no investment in the argument, so it will say the third paragraph does not follow rather than smoothing over it.

The prompt matters enormously here. Asking "what do you think of this" gets you compliments. Ask for this instead:

`You did not write the text below and you have no stake in it. Review it as a skeptical editor. List, in order: 1) any factual claim that is wrong, unsupported, or too strong, 2) any point where the argument does not actually follow, 3) the single weakest paragraph and exactly why. Do not rewrite anything. Do not tell me what is good.`

The three instructions at the end do the real work. "Do not rewrite" stops it producing a version in its own voice, which is not what you asked for. "Do not tell me what is good" removes the reflexive praise that pads most AI feedback. "You did not write this" is a surprisingly effective framing, because it stops the model defending choices it thinks are its own.

Then take the critique back to the first model with the parts you agree with. You are the editor deciding which notes to accept, which is exactly the right division of labour.

Where this is worth the extra two minutes: anything going to a client, anything public, anything where being wrong is expensive. Where it is not: a slack message, a first-pass outline, a shopping list.

#### Pattern 2: the disagreement check

This is the closest thing to a reliable hallucination detector that does not involve manually verifying everything.

Ask two different models the same factual question, independently, with the same prompt. Then compare. Where they agree, you are probably fine. **Where they disagree, you have found the exact spot that needs checking.**

That is the whole technique, and it is powerful because hallucinations are usually not shared. When a model invents a statistic, a case name, a function signature, or a date, a different model trained differently rarely invents the same one. Agreement is weak evidence of correctness. Disagreement is strong evidence that something is wrong, and it points straight at it.

For anything with several claims, make the comparison do the work for you:

`Below are two answers to the same question, marked A and B. Ignore style and length. List every substantive point where they disagree, including differences in numbers, dates, names, and how strongly something is claimed. For each disagreement, say which is more likely correct and what single source would settle it.`

You end up with a short list of things to check instead of a long document to fact-check, and the list is usually right about where the problems are.

Two warnings. First, models can share a mistake if it is common on the internet, so this catches invention rather than widespread misinformation. Second, do not skip verification because two models agreed. This narrows what you check, it does not remove the need to check.

Use it on: statistics, legal and medical claims, historical facts, API details, anything you will be quoted on. Skip it on: opinion, creative work, and anything where being approximately right is fine.

#### Pattern 3: route by task

The obvious pattern, worth doing properly. Send each kind of work to the model that is best at it, rather than sending everything to whichever one you have open.

| Task | Reach for | Because |
| --- | --- | --- |
| First draft of anything written | The strongest writing model | You are buying down editing time |
| Review of that draft | A different model | Fresh eyes, different failure modes |
| Anything about this week | A model with live search | Training data is always behind |
| A very long document or codebase | A large-context model | It has to fit before it can be understood |
| Hard logic, math, or a tricky bug | A reasoning model with extended thinking | Slower and much more accurate on multi-step problems |
| High-volume repetitive work | A cheap fast model | Frontier quality is wasted on classification and tagging |
| Anything sensitive | Whichever meets your data rules | Capability does not override your obligations |

The one habit worth building: **stop defaulting.** Most people use whichever model their subscription gave them for everything, including tasks it is poor at, and conclude that AI is bad at those tasks. Ten seconds of thought about which model fits changes the output more than an hour of prompt tuning.

For the fuller version of this decision, see [how to choose an AI model](https://whizi.io/resources/how-to-choose-ai-model), and [what is a context window](https://whizi.io/resources/what-is-a-context-window) for why the long-document row exists.

#### Pattern 4: relay with a written handoff

For work that spans several stages, the failure point is the gap between stages. You explain the project to one model, get somewhere useful, then move to another model and either paste a wall of history or explain everything again badly.

Fix it by asking for a handoff explicitly before you switch:

`Write a handoff brief for another assistant that has not seen this conversation. Include: what we are trying to produce, decisions already made and why, constraints and things to avoid, what has been tried and rejected, and exactly what I need next. Be specific enough that it can continue without asking me any questions.`

This is useful even when you are not switching models. It is the cleanest way to escape a long conversation that has become slow and vague, because you keep the substance and drop the accumulated noise. Paste the brief into a fresh chat and the quality usually jumps immediately.

A realistic relay looks like this. Give a 200 page report to a large-context model and ask for a structured summary with the ten passages that matter, quoted. Take those into a strong writing model and produce the client note. Run the critique pass from pattern one with a third model. Total time, about fifteen minutes, and every stage used a tool suited to it.

#### When not to bother

Multi-model workflows have overhead, and pretending otherwise is how people end up with an elaborate process for writing a two line email.

Use one model when: the task is small, the cost of being wrong is low, you are exploring rather than producing, you are iterating quickly and the friction would break your flow, or the work is genuinely creative and a second opinion just muddies your voice.

Use two or more when: the output goes to someone else, factual accuracy matters, the task has distinct stages that need different strengths, you are stuck and need a genuinely different angle, or you are about to make a decision based on what the model told you.

A reasonable rule: **if you would ask a colleague to look at it, ask a second model.** If you would not, do not.

#### Making it practical

All of this is straightforward in principle and annoying in practice if each model lives behind a different subscription. The friction is real: separate tabs, separate histories, copying context between them, losing formatting, and the small resistance that means you skip the critique pass on the day you most need it.

That friction is exactly why the patterns above are underused. They are not difficult, they are just tedious enough to skip when you are busy.

This is what a multi-model workspace is for. In [Whizi](https://whizi.io/pricing), the models sit in one conversation, so a critique pass is switching the model and asking, not opening another product and pasting your draft. One history, one search across everything you have ever asked, one bill. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation) and [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side) for how those two steps actually work.

If you would rather assemble it yourself from separate subscriptions, these patterns still work and you should still use them. The only thing that changes is how often you can be bothered. If you are already paying for two or three plans to make this possible, run the [savings calculator](https://whizi.io/tools/calculator) first, because that is usually the more expensive way to get the same result.

Start with pattern one. Take the next thing you write that someone else will read, run the critique prompt through a different model, and see what it finds. That single habit is worth more than the rest of this article.

Key points:

- Two models disagreeing is the cheapest hallucination detector available
- A critique pass from a second model catches what the writing model cannot see
- Most tasks need one model, and knowing which ones do not is the whole skill

Checklist:

- Draft with one model and critique with a different one before anything goes out
- Use the critique prompt that forbids rewriting and forbids praise
- Ask two models the same factual question and check every point where they differ
- Route long documents, hard logic, and live questions to models suited to them
- Ask for a written handoff brief before switching models or starting a fresh chat
- Skip all of this for small, low-stakes, or exploratory work

Steps:

1. Draft with the model that writes best: Produce the first version with whichever model needs the least editing for your kind of work, usually a strong writing model.
2. Critique with a different model: Paste the draft into a second model and ask it to review as a skeptical editor without rewriting, listing errors, weak arguments, and the worst paragraph.
3. Check disagreements on anything factual: Ask the same factual question of two models and treat every point where they differ as a claim that needs verifying.
4. Route specialist work to the specialist: Send long documents to a large-context model, hard logic to a reasoning model, and current questions to a model with live search.
5. Hand off with a written brief: When moving between models or starting fresh, ask for a handoff brief covering context, decisions, constraints, and the next step.

FAQ:

Q: Why use more than one AI model?
A: Because a model cannot see its own blind spots. A second model, trained differently, catches errors the first one was confident about. It also means your workflow does not break every time a different lab takes the lead on a particular task.

Q: How do I catch AI hallucinations?
A: Ask two different models the same question independently and compare. Invented facts are rarely invented the same way twice, so any point where the answers disagree is where you should verify. Agreement narrows what you check but does not remove the need to check.

Q: Which model should write and which should review?
A: Draft with whichever model needs the least editing for your kind of work, usually a strong writing model, and review with a different one. The reviewer matters less than the fact that it is a different model with different failure modes.

Q: Is using multiple models worth the extra time?
A: For anything that goes to another person or that you will be held to, yes, and the critique pass takes about two minutes. For quick, low-stakes, or exploratory work it is overhead you do not need. If you would ask a colleague to check it, ask a second model.

Q: Do I need several subscriptions to do this?
A: No. A multi-model workspace gives you the major models on one plan and one history, which removes the friction that makes people skip these steps. Separate subscriptions work too, they just make the good habits harder to keep.

### How to Create AI Images for Beginners (A Simple Guide)

URL: https://whizi.io/resources/image-generation/
Updated: August 2026

Quick answer: To use AI image generation, you type a description and a picture appears. Getting a good one means naming six things: subject, setting, lighting, framing, style, and mood or colour. Anything you leave out is filled with a generic default, and what you do not want goes in a separate negative list.

Learn how to easily create stunning AI images from text. Discover simple prompt tips to generate beautiful art, graphics, and photos in seconds.

#### How it works, in one paragraph

You type a description and a picture appears. Underneath, the tool starts with random visual noise and repeatedly nudges it toward something matching your words, having learned what words look like from an enormous number of captioned images.

One consequence explains most beginner frustration: the model is matching your description against patterns, not following instructions like a person would. If you say "no dog", the word dog is still in there and a dog may well appear. If you leave something out, it does not ask; it fills the gap with the most average version of that thing it has seen.

That is the whole skill. Say what you want specifically, and say what you do not want in the place designed for it.

#### The six-part formula

Nearly every good image prompt contains the same six things. Missing ones get filled with defaults, and the defaults are exactly what makes an image look generic.

| Part | What to say | If you leave it out |
| --- | --- | --- |
| Subject | The specific thing, with the details that matter | The most generic version of that noun |
| Setting | Where it is, and how much you see | A blank or cluttered background |
| Lighting | Time of day, direction, quality | Flat, evenly lit, lifeless |
| Framing | Close up or wide, from what angle | A centred medium shot, every time |
| Style | Photograph, oil painting, 3D render, illustration | Glossy digital art |
| Mood or colour | The feeling and palette | Oversaturated and high contrast |

A weak prompt: `a cozy living room`.

A strong prompt: `A small living room with a lit fireplace and a tabby cat asleep on a worn armchair, late afternoon light through a west-facing window, wide shot from the doorway, realistic photograph, 35mm, warm browns and soft amber, calm and lived in.`

Both take about the same effort to think of. The second one is a picture; the first is a category.

Then add the part beginners skip entirely: a negative list. `Negative: text, watermark, extra fingers, cluttered background, stock-photo smiling.` Most of what makes an image look AI generated comes from a small set of recurring artefacts, and naming them removes them.

#### The trick that makes this easy

You do not have to write these prompts yourself. Ask a text AI to write them for you.

`Write an image generation prompt for: [your rough idea]. It is for [purpose]. Structure it as subject, setting, lighting, framing, style, colour and mood, then a negative list. Give me three versions that differ in framing rather than in adjectives.`

Language models are considerably better at writing image prompts than most people are, because the format is a known one and they have seen enormous numbers of examples. This is why generating images inside a chat is genuinely more useful than a standalone image box: the tool that writes the prompt is sitting right next to the one that draws it.

#### Why your image came out wrong

A short diagnostic list covering nearly every beginner problem.

**Boring and generic.** You did not specify lighting or framing. Those two do more work than any other pair.

**The right things, arranged wrongly.** Say the framing explicitly: `wide shot from below`, `close up, subject in the left third`. The model has no idea where you want things unless you say.

**Something you asked not to see is in the picture.** Naming it in the main prompt makes it more likely, not less. Move it to the negative list.

**Hands, faces, or small repeated details are wrong.** Still the weakest area everywhere. Try a closer crop, a different angle, or hands not being visible. Sometimes just regenerate.

**Text is gibberish.** Expected. See the next section.

**It looks like stock photography.** Add `candid`, describe an action rather than a pose, and put `stock-photo smiling at camera` in the negatives.

**Wrong shape for where you need it.** Ask for the aspect ratio you actually need. Cropping a square into a banner throws away the composition you asked for.

#### Iterating without going round in circles

Beginners regenerate the same prompt hoping for a better roll. Some randomness is unavoidable, but most wasted attempts come from changing several things at once.

- **Change one thing per attempt.** Lighting, or framing, or style. Not all three, or you will not know what helped.
- **Fix composition before detail.** Getting the framing and light right first is far more efficient than perfecting a subject that is in the wrong part of the frame.
- **Describe what was right.** When an image is close, say so in words in the next prompt rather than assuming the tool remembers.
- **Save prompts that worked.** Six or eight reliable prompt shapes are worth more than any single image, and they are what make a set of pictures look like they belong together.

#### What it still cannot do

**Text.** Words in images remain unreliable everywhere. A short word sometimes works; a headline or a logo usually does not. Generate the picture clean and add text in any design tool, which also gets you the right font.

**Exact arrangements.** "Exactly three objects, the red one on the left" is not reliably followed. If the arrangement matters, generate the pieces and compose them yourself.

**The same character twice.** Getting an identical person or product across a series is difficult. Detailed repeated descriptions help and identical results are not guaranteed.

**Editing a real photo precisely.** It can adjust background, lighting, and style well. It cannot reliably move an object two inches to the left.

And on using the results: check the rules before anything generated goes somewhere commercial. Take particular care with images resembling a real person, a recognisable place, or the distinctive style of a living artist. Terms differ between tools and the responsibility sits with whoever publishes the image.

Key points:

- A six-part prompt formula, with a weak and a strong version of the same request
- Why images come out wrong, and the specific fix for each cause
- What AI image tools still cannot do, so you stop trying

Checklist:

- Include all six parts: subject, setting, lighting, framing, style, mood
- Put what you do not want in a negative list, never in the main description
- Ask a text AI to write the image prompt from your rough idea
- Generate at the aspect ratio you actually need
- Change one variable per attempt so you learn what helped
- Add text in a design tool rather than in the image model
- Save the prompts that worked so your images look consistent

FAQ:

Q: Do I need artistic skills to generate AI images?
A: No, but you do need descriptive ones, and that is the part beginners underestimate. The gap between a disappointing image and a good one is almost entirely how specifically you described the lighting, the framing, and the style. If describing a scene in detail does not come naturally, ask a text AI to expand your rough idea into a structured prompt, which works well.

Q: Can I use AI to help me come up with image ideas?
A: Yes, and it is more useful than most people realise. Ask a chatbot for five descriptive concepts, then ask it to turn your favourite into a full structured image prompt with a negative list. Language models write image prompts better than most beginners do, which is why generating images inside a chat beats a standalone image box.

Q: Why does the text in my image look like nonsense?
A: Because rendered text is genuinely the weakest area across every image model, and no prompt reliably fixes it. Short single words sometimes come out; headlines, logos, and labels usually do not. Generate the image without text and add typography in any design tool, which also lets you use the correct font and place it properly.

Q: Which image generator is best?
A: It depends on the subject more than people expect. Flux is strongest for photorealism and anything that should look like it came from a camera. Stable Diffusion gives more stylistic control and variety for illustration. The fastest way to decide is to run the same prompt through two of them and look at the results side by side, since the winner on portraits is often not the winner on interiors.

Q: Do I need a separate subscription for image generation?
A: Not necessarily. Image generation is included in Whizi on Pro and above, alongside the text models, which covers ordinary creative and content work without a second bill. A dedicated image tool still makes sense for people doing this professionally every day and wanting very fine stylistic control.

### Is AI safe to use? An honest answer, not a reassurance

URL: https://whizi.io/resources/is-ai-safe-to-use/
Updated: August 2026

Quick answer: AI is safe enough for everyday tasks with the same caution you give any cloud service, since your messages travel to the provider's servers rather than staying on your device. Check the training setting under data controls once, keep ID numbers, card details, and passwords out of the chat, and verify any answer you would act on.

Where your messages go when you press send, when chats get used for training, the short list of things you should never paste, and the risk that hurts people more often than a data leak.

#### Where your message actually goes

You type a question, press send, and an answer appears in about two seconds. That speed makes the whole thing feel like it happened on your phone. It did not. Your words left your device, traveled over the internet to a data center full of specialized chips owned by the company that makes the model, got processed there, and came back. A chatbot is nothing like a calculator app humming away offline. It is closer to sending an email, or uploading a photo to a cloud service: the useful work happens on someone else's computer. If the technology itself is still fuzzy, [what AI is](https://whizi.io/resources/what-is-ai) covers the ground floor in plain language.

Once your text sits on a company's servers, what happens next is decided by that company's policy and its access controls, not by anything physical on your end. That also puts the file question in proportion. A chatbot cannot go rummaging through your phone, your email, or your work drive on its own. It sees what you hand it, plus whatever permissions you granted the app when you installed it. Two minutes well spent today: open your phone settings, find the app, and switch off camera, microphone, or photo library access if you never use those features. Here is what actually reaches the provider.

- **The text you type or paste.** All of it, including the paragraph you copied from a document and did not reread.
- **Files you attach.** A PDF, a photo, a spreadsheet. Only the ones you pick, only at the moment you pick them.
- **Account and connection details.** Your email address, an approximate location from your IP address, your device type and app version.
- **Nothing else on your device.** Not your photo library, not your contacts, not your other apps, not documents you did not attach.
- **Sometimes, your past chats.** Many apps now carry memory across conversations, which is convenient and means old messages can resurface in new ones.

#### Training, tiers, and the setting nobody opens

The fear behind "does AI steal your data" is usually narrower than theft. What people actually want to know is whether their words end up inside the next version of the model, and whether a stranger could ever see them. The industry pattern is consistent enough to describe, even though every provider words it differently and rewrites that wording every few months. Broadly: the more the account looks like a business account, the less likely your content is used for training, and the more that promise is written into a contract rather than a settings toggle.

| What you are on | The usual pattern | What to check |
| --- | --- | --- |
| Free consumer account | Conversations may be used to improve models unless you opt out | The opt-out switch, and whether it is on |
| Paid consumer plan (roughly 20 dollars a month) | Often the same default as the free tier, with the same switch | Do not assume paying changed it, because it often does not |
| Business, team, or enterprise plan | Training on your content is usually off by default and contractual | Your admin's console, since the choice may not be yours |
| Developer or API access | Content is typically excluded from training by default | The provider's API data usage page |

The switch exists in every major consumer app. It sits under a heading like data controls, data and privacy, or activity, and providers move it often enough that printing a menu path here would be wrong within a month. Go and find it now, before you have anything sensitive in there. It is a five-minute job you do once per app.

Two things surprise people when they do. First, turning training off almost always applies going forward, not backward, so conversations you have already had may stay in whatever pool they were already in. Second, off means "not used to train", not "never stored". Providers generally keep chats for a period for abuse monitoring, safety review, and support, commonly around 30 days, before deletion runs. That is also true of chats you delete yourself. If you want the specifics for your provider, read their policy rather than an article summarizing it, including this one.

#### The never-paste list, and the fix that costs nothing

Some things should never go into a chat box, on any provider, under any setting. Not because a breach is likely, but because the damage is permanent if you are wrong once. You cannot un-send a national ID number.

- **Government ID numbers.** Social security, national insurance, passport, driver's license, tax ID.
- **Card and banking details.** Full card numbers, security codes, account and routing numbers, and crypto seed phrases above all.
- **Passwords and access keys.** Including the temporary one you swear you are about to change.
- **Other people's private business.** A parent's diagnosis, a friend's salary, a client's home address. That consent is not yours to give on their behalf.
- **Anything covered by a work confidentiality agreement.** Unreleased financials, customer lists, code your employer owns, documents under NDA. Check whether your workplace has an approved tool before you decide this one for yourself.
- **Unredacted contracts and legal documents.** A signed agreement usually carries names, addresses, and signatures the model does not need in order to answer your question.

The fix is unglamorous and it works: replace the sensitive parts with placeholders before you paste. Models read structure and language rather than identity, so a clause with [LANDLORD] in it gets the same analysis as one with a real name. Suppose you want a lease clause explained. Instead of pasting the lease, you send this:

> Review this clause from a residential lease. Tenant is [TENANT], landlord is [LANDLORD], monthly rent is [AMOUNT], property is at [ADDRESS]. Clause: "The tenant shall be responsible for all repairs and maintenance to the interior of the premises, without limit."

Then you read the answer and put the real values back yourself. On the work beginners actually bring to AI, summarizing, drafting, explaining, and checking, the quality loss from placeholders is close to zero. The same habit applies to uploads: redact the document before you attach it, which the guide to [summarizing a PDF with AI](https://whizi.io/resources/summarize-pdf-with-ai) walks through step by step.

#### The risk more likely to bite you

Privacy is the risk people ask about. Accuracy is the risk that actually costs them something. An AI model produces text at one steady pitch of confidence whether it is right or inventing, and there is no tremor in the voice when it fabricates a case citation, a dosage, a refund deadline, or a statistic with a decimal point in it. Beginners read fluency as knowledge, because in people it usually is. That instinct is the thing to unlearn first, and the fastest cure is to check three or four answers you were certain about and see what turns up.

The second failure is timing. A model learns from text gathered up to a cutoff date, so prices, tax thresholds, laws, product features, and anything from the last few months can be quietly out of date while sounding perfectly current. Some apps search the web to patch this and some do not, and few of them announce which one just happened. When an answer depends on a live number, ask outright whether it searched, then check the number at its source anyway.

The third one does the real damage: taking a fluent answer as advice. AI is genuinely good at preparing you for a medical, legal, or financial conversation. It explains what a term means, what your options are called, and which questions are worth asking. It is not a doctor, a lawyer, or a licensed advisor. It does not know your history, it cannot examine you, and it will not tell you when it is out of its depth. Use it before the appointment, not instead of it. [Why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) explains the mechanism, which is what makes a wrong answer easier to catch before you act on it.

#### Five habits you will actually keep

Safety advice fails when it demands effort every single time. These five cost you about ten minutes once, then a couple of seconds each thereafter, which is why they survive contact with a busy week.

1. **Check the data setting once, today.** Open settings, find data controls or privacy, decide whether your chats can be used for training, and move on. Repeat only when you switch apps.
2. **Use placeholders by default.** [NAME], [COMPANY], [AMOUNT]. Make it a reflex rather than a judgment call, because judgment calls are exactly what fail when you are rushing.
3. **Verify anything you would act on.** If an answer would make you send money, sign a document, take a medication, or quote a number to your boss, confirm it somewhere that is not the chatbot.
4. **Keep work and personal accounts separate.** Use your employer's approved tool for work material and your own account for everything else. This protects you at least as much as it protects them.
5. **Treat every answer as a first draft.** Rewrite it in your own words before it goes anywhere. You catch the errors while you edit, and the result sounds like you instead of like software.

None of this drops the risk to zero, and anyone who tells you otherwise is selling something. What it does is move the leftover risk into a category you already live in: your email sits on someone's server, your photos are in someone's cloud, and you decided long ago that the trade was worth it. AI is that same trade with a newer set of companies on the other end, so it deserves the same ordinary caution rather than either panic or blind trust.

One practical note on accounts. Every provider you sign up with is another privacy policy, another retention window, and another data control to hunt down, so four accounts means running that checkup four times and forgetting at least one of them. Consolidating cuts the number of policies you are responsible for: Whizi puts GPT, Claude, Gemini, and image models behind one subscription, which leaves one settings screen instead of four, while adding one more company to the chain. That is a trade, not a safety guarantee, and the never-paste list is identical either way. For the useful half of all this, [how to use AI](https://whizi.io/resources/how-to-use-ai) lays out a first week of real tasks, and [how to talk to AI](https://whizi.io/resources/how-to-talk-to-ai) covers the phrasing that gets better answers out of any model.

Key points:

- Your message does not run on your phone, it travels to the provider's servers, and that single fact is the root of every privacy question about AI
- Whether your chats are used for training depends on the provider and on one setting, usually filed under data controls, that you can find and change today
- The risk most likely to cost you something is a confident answer that happens to be wrong, delivered in the same steady tone as a correct one

Checklist:

- Open your AI app once and find the data control that governs training.
- Assume everything you send leaves your device and sits on a provider server.
- Replace names, numbers, and addresses with placeholders before you paste.
- Never send government ID numbers, card details, passwords, or seed phrases.
- Keep other people's medical, financial, and legal details out of the chat.
- Verify any answer you would act on against a source that is not the chatbot.
- Use separate accounts for work material and personal material.

FAQ:

Q: Is ChatGPT safe to use?
A: For everyday tasks like drafting, summarizing, and explaining, yes, with the same caution you would apply to any cloud service. Your messages are processed on OpenAI servers rather than your device, so the practical safety question is what you send rather than whether the app is trustworthy. Check the data controls in settings once, keep ID numbers and passwords out of the chat, and verify facts you plan to act on.

Q: Does AI steal your data?
A: Stealing is the wrong frame. Major providers state what they collect and how they use it in published policies, and the real question is whether your conversations may be used to improve future models, which is often a setting you control on consumer plans. The safer assumption is that anything you type could be stored for a period and reviewed by a human in rare cases, so decide what you send on that basis.

Q: Is it safe to put personal information into an AI chat?
A: Ordinary personal context, such as your job, your city, or your situation, is generally fine and usually makes answers better. Identifying numbers are a different category: no government ID, card, bank, or password details, ever, and no private information belonging to other people. Use placeholders like [NAME] and [AMOUNT] instead, which barely affects answer quality.

Q: Can AI see the files on my phone or computer?
A: No. A chatbot cannot browse your device, your photo library, or your work drive on its own. It only sees a file when you deliberately attach or upload it, and only that file. If an app has camera or photo permissions you never use, you can revoke them in your phone settings without losing anything.

Q: Is it safe to use AI for work documents?
A: That depends on your employer, not on the AI. Many companies have an approved tool with a business plan where content is excluded from training by contract, and using a personal account for confidential material can breach your agreement even if nothing leaks. Ask what is approved, use that account for work, and redact client names and financial figures out of habit.

### Is ChatGPT Plus worth it? An honest breakdown

URL: https://whizi.io/resources/is-chatgpt-plus-worth-it/
Updated: August 2026

Quick answer: ChatGPT Plus is worth it if you hit the free plan's limits often enough to notice and ChatGPT is where most of your AI work happens. It is not worth it for occasional users, for sourced research that needs links, or if you already pay for two other AI subscriptions.

A straight answer on whether ChatGPT Plus is worth paying for, based on what the free plan already does, the limits you actually hit, and the break-even math.

#### The short version

ChatGPT Plus is worth it if you hit the free plan's limits often enough to notice, and if ChatGPT is where most of your AI work happens. It is not worth it if you use AI a few times a week, if you mostly need sourced answers with links, or if you are already paying for two other AI subscriptions that cover the same ground.

That sounds obvious, but most reviews answer a different question. They ask whether AI is useful, decide that it is, and recommend the subscription. The useful question is narrower: **what does Plus give you that free does not, and do you personally run into those walls?** For a lot of people the honest answer is no, and they pay $20 a month for a plan whose main benefit is that they stop seeing a limit message they were only seeing twice a month anyway.

ChatGPT Plus has been $20 a month for a long time, with a $200 Pro tier above it and Team seats in between. Prices and plan names move, so check the current numbers on OpenAI's pricing page before you commit. What matters more than the exact figure is the shape of the decision, and that has stayed the same: free covers casual use well, Plus buys headroom and access, Pro buys volume that almost nobody needs.

#### What you actually give up on the free plan

The free tier is much stronger than it was two years ago. You get the flagship model for a limited number of messages, then a quieter fallback model for the rest of the day. You get file uploads, image understanding, image generation, web browsing, and voice, all with tighter caps. For a student writing a few essays a month or someone drafting the occasional email, that is genuinely enough, and [the student workspace](https://whizi.io/use-cases/students) sets out what that kind of coursework actually needs.

Here is where the plans separate. The specific caps change often, sometimes month to month, so treat the numbers in any article (including this one) as a shape rather than a spec sheet.

| What you get | Free | Plus |
| --- | --- | --- |
| Flagship model | Limited messages, then a smaller fallback | Much higher cap, rarely hit in normal use |
| Reasoning models for hard problems | Very limited or unavailable | Included, with usage caps |
| Deep research style multi-step reports | Small monthly allowance | Larger allowance |
| Image generation | A few per day | Higher daily limit |
| Advanced voice | Short daily window | Extended |
| Projects, memory, custom instructions | Available, smaller scale | Full |
| Peak-time access | Degrades first when demand spikes | Priority |
| New feature access | Weeks or months later | Usually first |

Read that table and one thing stands out: almost every line is about **volume and access, not intelligence**. With one exception. The reasoning models, the ones that think longer before answering, are a real capability difference, not just a bigger bucket. If your work involves multi-step logic, tricky math, dense code review, or analysis where a plausible-sounding wrong answer costs you something, that is the single strongest argument for paying.

Everything else on the list is headroom. Headroom is worth money if you are running into the ceiling. It is worth nothing if you are not.

#### The break-even math nobody actually does

The standard pitch is that $20 a month is 67 cents a day, so if AI saves you fifteen minutes a week it has already paid for itself. That math is true and also useless, because it is measuring the wrong thing. The free plan is also saving you that fifteen minutes. You are not deciding between AI and no AI. You are deciding between free AI and paid AI.

So run the honest version. Look back at the last two weeks and count two things: how many times you saw a usage limit message, and how many times you gave up on a task or moved it elsewhere because the model you had access to was not good enough.

| Times you hit a wall in two weeks | What it means | What to do |
| --- | --- | --- |
| 0 to 1 | You are a light user and free is doing the job | Stay free, revisit in three months |
| 2 to 5 | You are on the edge | Pay for one month, then count again |
| 6 to 12 | You are working around the plan regularly | Plus will pay for itself in saved friction |
| More than 12 | AI is core to your day | Pay, and start asking whether one provider is enough |

If you bill for your time, add the second calculation. Take your effective hourly rate, divide by 60, and multiply by the minutes per month you lose to limit messages, waiting, retrying prompts on a weaker model, and redoing work the fallback model got wrong. At $50 an hour, Plus needs to save you 24 minutes a month to break even. At $150 an hour, it needs eight minutes. Most people who use ChatGPT daily clear that easily. Most people who use it twice a week do not, no matter how much they enjoy it.

One more number worth knowing before you subscribe: your total AI spend. Plus alone is easy to justify. Plus plus a research tool plus a coding assistant plus an image tool is a different conversation, and it is the one most people avoid having. The [AI subscription savings calculator](https://whizi.io/tools/calculator) totals it in about two minutes.

#### Who should pay, and who should not

These profiles come from the pattern we see most often when people compare their stacks. Find the one closest to you.

**Pay for Plus if you are:**

- A daily user who opens ChatGPT before opening a blank document, and who has seen the limit message more than once this week.

- Someone whose work depends on careful reasoning: financial modelling, legal analysis, debugging logic errors, anything where a confident wrong answer is expensive. The reasoning models justify the price on their own.

- A solo operator or small team without other AI tools, where ChatGPT is doing the job of a researcher, an editor, and a junior analyst at once.

- Someone who works with long documents and images together, and who needs the higher upload and analysis limits to get through a real workload.

**Do not pay for Plus if you are:**

- An occasional user. Two or three sessions a week almost never touches the free ceiling, and paying is a subscription to a feeling of readiness.

- Primarily doing sourced research where every claim needs a link. ChatGPT browses, but tools built around citation do that job better. Compare [Perplexity as an alternative](https://whizi.io/alternatives/perplexity) before you settle.

- A developer who lives inside an editor. In-IDE assistants that see your whole repository will beat a chat window for day to day coding, though a chat model is still better for design discussions and code review. Our [AI for coding guide](https://whizi.io/resources/ai-for-coding) covers the split.

- Already paying for Claude and Gemini. Three general assistants at $20 each is the most common form of AI overspending, and the overlap between them is larger than the difference. Read [ChatGPT Plus vs Claude Pro vs Gemini Advanced](https://whizi.io/resources/chatgpt-plus-vs-claude-pro-vs-gemini-advanced) before adding a third.

#### Where ChatGPT Plus is genuinely weak

No subscription is good at everything, and it is worth knowing the edges before you pay rather than after.

**You are locked to one lab's judgment.** When OpenAI changes which model is the default, deprecates one you liked, or adjusts limits, your workflow changes with it. That is fine when the flagship is ahead and annoying when it is not. Model leadership has changed hands several times, and it will again.

**Sourced research is a compromise.** Browsing works, but the output is a summary with some links attached rather than an answer built from sources. If you need to check every claim, you will spend the time you saved verifying it.

**Long-document work has better homes.** Models with very large context windows handle a 400-page PDF more gracefully. If most of your work is synthesis across long material, test that specific task elsewhere before committing. Our [context window explainer](https://whizi.io/resources/what-is-a-context-window) covers why this matters more than the marketing numbers suggest.

**Writing voice is subjective, and many people prefer Claude.** This is the least measurable difference and the one people feel most strongly about. Run the same brief through both before deciding whose default voice needs less editing. [Claude vs ChatGPT](https://whizi.io/resources/claude-vs-chatgpt) has the head to head.

None of these make Plus a bad product. They just mean the honest recommendation depends on what you do all day, which is exactly what most reviews skip.

#### Run a seven-day test before you decide

Reviews cannot tell you whether a subscription fits your work. A week of your own data can. This test takes about ten minutes of overhead a day.

1. **Pick five real tasks** from your last two weeks of work. Not demos. Actual things: the client update you rewrote four times, the spreadsheet you needed explained, the function you could not get right, the research summary you built by hand.

2. **Run each one on the free plan first.** Save the output. Note how long it took, how many retries you needed, and how much editing the result needed before you could use it.

3. **Start the paid month and run the same five tasks.** Same prompts, same order. Use the reasoning model where the task calls for it, since that is the capability you are actually buying.

4. **Score each pair on three things:** minutes to a usable result, edits required, and whether you trusted the answer enough to send it without checking.

5. **Count your limit hits during the week.** If the number is near zero even under a heavier test load, that is your answer.

6. **Decide on the totals, not on the best moment.** Everyone can remember one impressive response. The question is the average.

If Plus wins clearly, keep it and stop reading comparison articles. If it wins narrowly, the more interesting question is whether a different model would have won bigger on those same five tasks, which is something you can only know by running them side by side.

#### The stacking trap, and the alternative

Here is the pattern we see constantly. Someone pays for ChatGPT Plus. A few months later they read that Claude writes better, so they add Claude Pro. Then a project needs long-document work, so they add Gemini. Now they are at $60 a month, they have three tabs open, their prompt history is split across three products, and they still are not sure which one to use for a given task.

Each individual decision was reasonable. The stack is not. And the fix is not brand loyalty, because the reason people stack is real: different models genuinely are better at different jobs, and the gap moves with every release.

The alternative is to keep the multi-model access and drop the multi-subscription cost. That is what [Whizi](https://whizi.io/pricing) does: one plan, the major models from OpenAI, Anthropic, Google, and others in one place, the same conversation continued across models, and one prompt history instead of three. If you want the direct comparison, [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) lays out where each one wins.

To be clear about the tradeoff: if ChatGPT is the only AI tool you use and you never wonder what a different model would say, Plus is a perfectly good purchase and a consolidation tool adds a layer you do not need. The case for consolidating starts the moment you open a second subscription, or the moment you find yourself pasting the same prompt into two products to compare the answers.

Whichever way you go, decide it with numbers. Count your limit hits, run the seven-day test, total your current spend with the [calculator](https://whizi.io/tools/calculator), and then pick a plan on the [pricing page](https://whizi.io/pricing) that matches how you actually work.

Key points:

- The real question is not whether AI is worth $20, it is whether the gap between free and Plus is worth $20
- A two-week limit count tells you the answer faster than any review
- Plus is a poor fit for sourced research, in-editor coding, and anyone already paying for two other AI plans

Checklist:

- Count how many times you hit a usage limit in the last two weeks
- Separate volume upgrades from real capability upgrades, especially reasoning models
- Calculate your break-even in minutes saved per month, not in vague productivity
- Test five real tasks on free before you pay, and the same five after
- Check whether a citation tool or an in-editor assistant fits your work better
- Total every AI subscription you already pay for before adding another

FAQ:

Q: Is ChatGPT Plus worth it in 2026?
A: It is worth it for daily users who regularly hit the free plan's limits and for anyone who needs the reasoning models for multi-step problems. It is not worth it for occasional users, since most of what Plus adds is headroom rather than a smarter model.

Q: What is the real difference between free ChatGPT and Plus?
A: Mostly volume and access: higher message caps on the flagship model, larger allowances for research and image generation, priority at peak times, and earlier access to new features. The one genuine capability difference is fuller access to the reasoning models that think longer before answering.

Q: Is ChatGPT Pro worth it instead of Plus?
A: For almost everyone, no. Pro costs around ten times as much and mainly buys near-unlimited usage plus the heaviest reasoning modes. It makes sense for people running AI workloads all day, such as researchers and heavy engineering users, not for anyone deciding whether $20 is justified.

Q: Does ChatGPT Plus make the answers better?
A: Not for a simple question. The same flagship model answers both free and paid users. Answers improve when a task benefits from a reasoning model, which Plus gives you far more access to, and when you stop being pushed onto the smaller fallback model after hitting a cap.

Q: Can I cancel ChatGPT Plus if it is not worth it?
A: Yes, and you keep access until the end of the billing period you already paid for. Our [step by step cancellation guide](https://whizi.io/resources/how-to-cancel-chatgpt-plus) covers the web, iOS, and Android paths, which are different and catch people out.

### Is Claude Pro worth it? What you get and what you will hit

URL: https://whizi.io/resources/is-claude-pro-worth-it/
Updated: August 2026

Quick answer: Claude Pro is worth it if written output quality is what you are buying and you work in sessions rather than all-day marathons. It produces drafts needing less cleanup than anything else at this price. Skip it if you use AI heavily and continuously, or need image generation and voice.

An honest look at Claude Pro: the writing quality that justifies it, the usage limits that frustrate people, who should pay, and who should pick something else.

#### The short answer

Claude Pro is worth it if the quality of written output is what you are buying, and if your working style is sessions rather than marathons. For drafting, editing, careful analysis, and code review, it produces work that needs less cleanup than anything else at this price. That is a real, repeatable saving and it shows up in every session, not just the impressive ones.

It is not worth it if you work in long unbroken stretches of heavy use, because you will meet the usage limits, and meeting them is genuinely disruptive in a way that other plans are not. It is also the wrong pick if you need image generation, voice, or a wide set of workplace integrations, none of which are Anthropic's focus.

One line summary: **Claude Pro gives you the best writing for the money and the tightest leash.** Whether that trade works depends almost entirely on how heavily you use it, which is why this article spends more time on limits than on features.

#### What you actually get for the money

Around $20 a month, cheaper if you pay annually, with much heavier Max tiers above it. Verify current pricing before subscribing, since the tiering here has changed more than once.

**Substantially more usage than free.** The free tier is a taster: enough to judge the writing quality, not enough to work with. Pro multiplies your capacity by roughly five times and gives you the more capable models for real work.

**Access to the stronger models.** The lighter models are fine for quick questions and noticeably weaker at the things people choose Claude for. Pro gives you the ones worth having.

**Projects.** A persistent workspace with its own uploaded knowledge and its own instructions. Put a style guide, a brand voice document, and three past examples into a project, and every conversation inside it starts already knowing them. For anyone producing recurring work, this is the feature that changes your habits, and it is under-marketed.

**Artifacts.** Documents, code, and small interactive pieces render beside the conversation and update in place instead of being regenerated as walls of text. Better for iterating on a document than scrolling a chat log.

**Claude Code.** A terminal-based coding agent, included with Pro at limited usage. This is the most underrated line in the package. For a solo developer, having a capable agent that reads your repository, makes multi-file changes, and runs tests is worth the subscription on its own. Heavy use will push you toward Max, but light and moderate use fits inside Pro.

**Connectors and extensions.** Anthropic's integration story has improved fast, including linking to tools you already use and working directly in the browser. Still narrower than ChatGPT's, but no longer a gap you can dismiss.

#### The limits, explained properly

This is where most Claude Pro reviews are useless, because they quote a number that changes and then move on. What you need is the shape of how limits behave.

Claude limits usage in two overlapping ways: a **rolling session cap** that resets every few hours, and a **weekly ceiling** for very heavy use. The session cap is the one you meet in normal work. The weekly one is aimed at people running agents all day.

Crucially, **it is not counted in messages, it is counted in tokens.** That means what you send matters as much as how often. Ten short questions cost far less than one question about a 200 page PDF, because the whole document is re-read on every turn in that conversation. Two people can send the same number of messages and have completely different experiences.

Here is what drains a session fastest, in order:

1. **Long documents in a long conversation.** The document is re-processed every turn. A 100 page PDF discussed over thirty messages is by far the most expensive thing you can do.
2. **Extended thinking on hard problems.** Reasoning tokens count even though you only see a summary of them.
3. **Large code files pasted repeatedly** rather than referenced once.
4. **Very long conversations** that never get restarted, since the full history is re-read each turn.

Which means the fix is mostly behavioural, and it works:

- Start a new conversation when the topic changes rather than continuing a giant thread.
- Attach a document to a Project once instead of pasting it into every chat.
- Send the relevant twenty pages instead of the whole report.
- Ask for a written handoff summary and start fresh when a session gets long.

People who work this way rarely hit the cap. People who keep one enormous conversation running all day hit it constantly, then write a review saying the limits are unusable. Both experiences are real, and the difference is habit. Our [context window explainer](https://whizi.io/resources/what-is-a-context-window) covers why this happens under the hood.

#### Where Claude genuinely wins

**Prose that needs less editing.** This is the whole argument, and it is a strong one. Claude holds a register, avoids the padding rhythms that make AI text recognisable, follows nuanced style instructions, and produces first drafts that read like a competent human wrote them in a hurry rather than like a template got filled in. If you write professionally, run one real brief through Claude and one through anything else, and count the edits. That number is your answer.

**Reading long and dense material.** Contracts, research papers, board packs, code reviews. Claude is careful, flags ambiguity, and is more willing than most to say a document does not answer your question rather than inventing something that sounds like an answer. For work where a confident wrong answer costs you, that restraint is worth paying for.

**Following instructions precisely.** Give it a nine-point brief with constraints and exceptions and it tends to honour all nine. Other models drop point six.

**Code review over code generation.** It reads unfamiliar code well, explains architectural risk in plain language, and pushes back on approaches rather than cheerfully implementing whatever you asked for. Our [AI for coding guide](https://whizi.io/resources/ai-for-coding) covers where that fits alongside in-editor tools.

**Being told it is wrong.** Correct Claude and it usually adjusts properly rather than apologising and repeating itself with different adjectives. That sounds minor and it is not, because most real work is iterative.

#### Where it will let you down

**Image generation.** Not part of the product. If you need images, you need something else, and our [best AI image generators guide](https://whizi.io/resources/best-ai-image-generators) covers the options.

**Voice.** Far behind ChatGPT and Gemini. If you talk to your assistant while walking or driving, this is decisive.

**Very large context work.** Claude's window is generous but Gemini's is in a different class. For whole-codebase reasoning or a year of transcripts, that gap is real.

**Ecosystem breadth.** Fewer integrations, fewer third-party tutorials, a smaller community. When you hit an unusual problem, there is less written about it.

**Heavy daily use.** Say it plainly: if AI is running constantly through your working day, Pro will not be enough and you will be looking at Max, which costs five to ten times more. That is a legitimate product for people who need it, and it is a very different budget conversation from $20.

**Real-time information.** Web access exists but is not the strength here. For live questions, see [Grok vs ChatGPT](https://whizi.io/resources/grok-vs-chatgpt) and [Perplexity vs ChatGPT](https://whizi.io/resources/perplexity-vs-chatgpt).

#### Should you pay? Match yourself to a profile

| You are | Verdict | Why |
| --- | --- | --- |
| A writer, editor, or marketer | Yes, first choice | Editing time saved pays for it in the first week |
| A consultant or analyst | Yes | Careful synthesis of long documents is the core strength |
| A solo developer | Yes, for Claude Code and review quality | Watch usage if you run the agent all day |
| A student | Free tier first | Free is enough for essays and reading, upgrade at deadline season |
| A heavy all-day user | Pro will frustrate you | Budget for Max or use a multi-model workspace instead |
| Someone who needs images or voice | No | Not what this product does |
| Someone in Google Workspace all day | Probably not | Gemini's integration is worth more to you day to day |
| Already paying for ChatGPT Plus | Read the next section first | Two general assistants is the most common overspend |

If you want the three-way version of that decision, [ChatGPT Plus vs Claude Pro vs Gemini Advanced](https://whizi.io/resources/chatgpt-plus-vs-claude-pro-vs-gemini-advanced) puts all three side by side at the same price point.

#### The one-week test, and the second-subscription question

Before you subscribe, spend a week on the free tier deliberately. Take the three pieces of writing you produce most often and run them through Claude with a proper brief. Count the edits you had to make before the text was usable, and compare that honestly with whatever you use now. Then note how often you hit the free limit, because that tells you whether Pro's five times capacity will actually be enough or whether you are a Max person in disguise.

Then there is the question most people arrive at eventually. You already pay for ChatGPT Plus, Claude writes better, and you are considering paying for both. That is a real and reasonable impulse, because the models genuinely are different. It is also $40 a month, two histories, two sets of saved prompts, and a small decision every time you start a task. Before you commit to that second plan, [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) sets out what a single subscription covers instead.

If what you want is Claude's specific product features, Projects, Artifacts, and Claude Code, buy Claude Pro. Those are not available anywhere else and they are good.

If what you want is access to Claude's writing alongside other models, that is what a multi-model workspace is for. [Whizi](https://whizi.io/pricing) includes Claude alongside GPT, Gemini, Grok, and DeepSeek on one subscription, so you can draft in Claude, stress-test the argument with a different model, and keep it all in one conversation and one history. [Whizi vs Claude Pro](https://whizi.io/vs/whizi-vs-claude-pro) is the direct comparison, including what you give up.

Whichever way you go, total your existing AI spend first with the [savings calculator](https://whizi.io/tools/calculator). A surprising number of people discover they are already paying for three plans and using one.

Key points:

- Claude Pro is the strongest $20 plan for writing and code review, and the most likely to make you feel rationed
- Usage limits are the real decision, not features, so estimate your session length before subscribing
- Claude Code included in the plan is the most underrated part of the package for solo developers

Checklist:

- Test Claude's writing on your three most common tasks before paying
- Count edits required, since that is the number the subscription is buying down
- Estimate your usage in tokens, not messages, because documents dominate the cost
- Attach recurring reference material to a Project instead of pasting it every time
- Start fresh conversations rather than running one long thread all day
- Check whether you need Max before assuming Pro will be enough

FAQ:

Q: Is Claude Pro worth it compared to ChatGPT Plus?
A: For writing, editing, long-document analysis, and code review, Claude Pro is the stronger choice at the same price. ChatGPT Plus is better for breadth, integrations, image generation, voice, and data analysis with code execution. Pick based on which half of that describes your week.

Q: What are Claude Pro's usage limits really like?
A: Usage is measured in tokens across a rolling session window, with a weekly ceiling for very heavy users. Long documents held in long conversations are what drain it fastest, since the whole thing is re-read each turn. Starting fresh chats and using Projects for reference material keeps most people well clear of the cap.

Q: Is Claude Code included with Claude Pro?
A: Yes, with limited usage. That is enough for light and moderate development work and it is one of the strongest reasons for a solo developer to pick this plan. Running the agent continuously will push you toward a Max tier.

Q: Is the free version of Claude good enough?
A: It is good enough to judge the writing quality and to handle occasional tasks, but the limits are tight enough that it is a sample rather than a workspace. Use it for a week to decide, then upgrade if you keep running out mid-task.

Q: Should I get Claude Max instead?
A: Only if you have already hit Pro's limits repeatedly in normal use, typically because you run coding agents or process large documents all day. Max costs five to ten times more, so start with Pro and let your own usage decide rather than buying headroom you may not need.

### What is Meta Llama 3 AI (Explained Simply for Beginners)

URL: https://whizi.io/resources/llama-3/
Updated: August 2026

Quick answer: Llama 3 is one generation of the AI models Meta builds. You type a question and it answers, like ChatGPT or Claude. The difference is distribution: Meta releases the weights, the numbers that constitute the trained model, so anyone can download Llama and run it, even with the internet off.

Learn what Meta Llama 3 is, how it works behind the scenes, and how beginners can use this lightning-fast AI for quick, everyday answers.

#### What Llama 3 is

Llama is a family of AI models built by Meta, the company behind Facebook and Instagram. Like ChatGPT and Claude, you type a question and it answers. The interesting part is not what it does but how it is distributed.

Most AI models are locked inside their maker's product. You use ChatGPT through OpenAI, and the model itself never leaves their servers. Meta releases Llama's weights, which are the numbers that constitute the trained model, for anyone to download and run.

A useful analogy: most AI is a restaurant, where you order and they cook. Llama is closer to publishing the recipe. You can still eat at the restaurant, and you can also cook it yourself, adapt it, or open your own kitchen.

One clarification you will see argued about online. Llama is usually described as open source, and strictly speaking it is "open weights" with a licence that carries some conditions, including restrictions on very large-scale commercial use. For an individual it makes no practical difference. For a company building a product on it, the licence is worth reading.

#### Why that matters even if you never download anything

Three consequences reach ordinary users.

**It is cheap, so it is everywhere.** Because anyone can host it, competition drives the cost down. Llama models power a large amount of the AI you encounter in other products without being labelled, and they are usually what a service means by a fast or economical tier.

**It can run without the internet.** A model on your own machine works on a plane, in a secure environment, or anywhere data must not leave the building. No hosted service can offer that.

**It keeps the market honest.** A capable free option puts a ceiling on what closed models can charge, which is good for you regardless of what you use.

#### What it is actually good at

Being honest about this is more useful than enthusiasm. Llama comes in several sizes, and the trade-off is consistent: smaller models are extremely fast and cheap, larger ones are more capable but no longer cheap to run.

| Task | Llama | Better elsewhere |
| --- | --- | --- |
| Quick factual questions | Fast and good enough | No |
| Short lists, brainstorming, simple rewrites | Very good, and instant | No |
| Repetitive work across many items | Ideal, because cost per item matters | No |
| Anything needing to run offline or privately | The only real option | No |
| Writing a person will read end to end | Usable | Claude, noticeably |
| Complex multi-step reasoning | Weaker than the frontier | GPT or Claude |
| Very long documents | Limited | Gemini |
| Anything from the last few weeks | It does not know | A model with web access |

The pattern: excellent for volume and speed, competitive for everyday questions, and behind the frontier models on hard reasoning and polished writing. That is a fair trade for something free, and it is why "which is better" is the wrong question. Use it where speed and cost matter, and switch when the task is hard.

#### Three ways to use it

**1. Through a workspace, with no setup.** The simplest option. Llama is included in Whizi alongside GPT, Claude, and Gemini, so you can send the quick questions to it and switch to a stronger model in the same conversation when the task gets harder. Nothing to install.

**2. On your own computer.** Tools such as Ollama and LM Studio download a model and run it locally. Realistic expectations: you need a reasonably modern machine with plenty of memory, the smaller models are the ones that run comfortably, and answers will be slower than a hosted service unless you have a good GPU. In exchange, nothing you type leaves your machine, and it works with the internet off. Genuinely worth it if privacy or offline use is the requirement, and unnecessary otherwise.

**3. Through an API.** If you are building software, providers host Llama at very low per-token cost, which is often the reason to choose it over a frontier model for high-volume tasks.

#### Getting good answers out of it

Smaller models reward a different prompting style than the frontier ones. Three habits make a large difference.

**Be direct and specific.** `List 10 name ideas for a coffee shop, one line each, no explanations` works better than a conversational request. Smaller models follow simple explicit instructions well and drift on elaborate ones.

**One thing at a time.** Frontier models handle six constraints in one prompt. Smaller models do better with a sequence of single requests.

**Say the format.** `Answer in one sentence` or `return only a numbered list` prevents the padding that smaller models produce when uncertain.

And know when to switch. If an answer is vague, contradicts itself, or misses something obvious, that is the signal to move the same question to a stronger model rather than to keep rephrasing. In a workspace that holds several, this is one click and the conversation carries over.

#### What to be careful about

**It does not know what happened recently.** Its knowledge stops at its training date, and unless it is connected to search, it will answer a question about last month from general knowledge, confidently.

**It invents things, like every model.** Smaller models do it somewhat more. Check any specific fact, date, or number.

**Free does not mean private, if it is hosted.** Running Llama on your own machine is private. Using it through someone's service means their data policy applies, exactly as with any other model.

**Version confusion.** Llama 3 was one generation, and newer versions have shipped since. Ask which version you are using if it matters, because capability differences between generations are large.

Key points:

- What open weights actually mean, and why it changes what you can do rather than just who owns it
- The two things Llama is genuinely best at, and the tasks where you should use something else
- Three ways to actually use it, from zero setup to running it on your own machine

Checklist:

- Use it for quick questions, short lists, and repetitive tasks where speed matters
- Switch to a stronger model when the task needs careful reasoning or polished writing
- Keep prompts direct, one request at a time, and state the output format
- Run it locally only if you actually need offline or private use
- Check any specific fact, date, or number, as with any model
- Ask which version you are using, since generations differ a lot

FAQ:

Q: Do I need to download software to use Llama 3?
A: No. Downloading and running it locally is one option and it is the right one if you need offline or fully private use, but it requires a reasonably powerful machine and gives slower answers than a hosted service. Most people use it through a platform in the browser, where it sits alongside other models and takes no setup at all.

Q: Is Llama 3 faster than ChatGPT?
A: The smaller Llama models are noticeably faster, which makes them excellent for short questions, quick lists, and anything repetitive. The trade-off is capability: on complex reasoning and polished writing, the frontier models are clearly better. The sensible approach is to use Llama for speed and switch when a question turns out to be harder than it looked.

Q: Is Llama really free?
A: The model weights are free to download and the licence permits most uses, with some conditions that matter mainly to very large-scale commercial deployments. Running it is not free in practice, since it costs either your own hardware and electricity or a hosted provider’s per-token fee, which is simply much lower than the frontier models charge.

Q: Is it private if I run it on my own computer?
A: Yes, and that is the strongest reason to run it locally. Nothing you type leaves your machine, it works with the internet disconnected, and no provider policy applies because no provider is involved. The costs are setup effort, a machine with enough memory, slower responses without a good GPU, and being limited to the smaller models that run comfortably.

Q: Should I use Llama instead of ChatGPT or Claude?
A: Instead of, rarely. Alongside, often. It is the right choice for fast simple questions, high-volume repetitive work, and anything that must run privately or offline. For writing someone will read carefully, for hard multi-step reasoning, and for very long documents, the frontier models are meaningfully better. Having both available means the choice is per task rather than a commitment.

### Multimodal AI: how to work with text, images, and documents

URL: https://whizi.io/resources/multimodal-ai/
Updated: August 2026

Quick answer: Multimodal AI means a model can work with more than one kind of input or output, usually text plus images, screenshots, PDFs, charts, and documents. The workflow that makes it reliable has four moves: observe what is visible, extract structured fields, interpret, then verify against the source. Weak prompts skip straight to interpretation.

Learn practical multimodal AI workflows for text, images, screenshots, PDFs, long documents, and structured extraction tasks.

#### What multimodal means

Multimodal AI means an AI system can work with more than one kind of input or output. In daily work, that usually means text plus images, screenshots, PDFs, charts, tables, documents, or slide decks. Instead of only typing a question, you can give the model visual or document context and ask it to extract, explain, compare, summarize, or transform what it sees.

The useful question is not "what is multimodal AI?" It is "which parts of my work need evidence from text, images, and documents at the same time?" That is where it stops being a demo and starts saving you an afternoon.

A product manager uploads a screenshot and asks for UX issues. A founder uploads three competitor pricing pages and asks for a comparison table. A researcher uploads a PDF and asks for claims, caveats, and source-backed takeaways. A support team pastes a customer complaint alongside a screenshot and asks for a diagnosis.

A strong multimodal workflow has four moves: observe, extract, interpret, and verify. Observation asks the model to describe what is visible or present. Extraction turns the input into structured fields. Interpretation explains what the evidence might mean. Verification checks whether the answer stayed grounded in the source. Most weak multimodal prompts skip straight to interpretation, which is where confident mistakes sneak in.

The practical rule: make the model prove that it noticed the source before asking it to reason about the source. For images, ask for visible evidence. For PDFs, ask for a document map. For long document packets, ask for sections, claims, tables, and unknowns before the final summary. This turns multimodal AI from a novelty into a repeatable work system.

#### Image -> text workflows

Image input AI models are useful for screenshots, charts, whiteboards, product photos, invoices, handwritten notes, social ads, dashboards, diagrams, and visual QA. The key is to treat image analysis as evidence collection before opinion. Ask the model what it can see, what it cannot read, and what it is inferring.

Use this image workflow when accuracy matters: upload the image, ask for a visible-evidence inventory, extract structured details, ask for interpretation separately, then run a verification pass. If the image includes tiny text, cropped edges, blurry regions, or visual ambiguity, tell the model to flag those limits instead of filling gaps.

Screenshot analysis prompt: "Analyze this screenshot in four sections. First, list only visible elements: headings, buttons, errors, labels, numbers, and layout issues. Second, extract any readable text into a table with location and confidence. Third, explain likely user problems based only on visible evidence. Fourth, list what is too small, cropped, or unclear to verify."

Chart extraction prompt: "Review this chart. Extract the title, axes, labels, legend, time period, highest and lowest values, visible trends, and any caveats. Separate directly visible data from interpretation. If exact values are not readable, say approximate and explain why."

UX review prompt: "Look at this product screen as a usability reviewer. Start with visible observations. Then identify friction points, accessibility concerns, confusing labels, missing states, and likely next actions. Return a prioritized table with issue, evidence, impact, suggested fix, and confidence."

Multimodal prompting improves when the output format is explicit. A vague prompt like "What do you think of this?" invites a vague answer. A better prompt asks for a table, a checklist, a set of risks, or a before/after rewrite. When you need work product, specify the work product.

#### Doc/PDF workflows

Documents add a different challenge. PDFs and long files can contain text, page layout, tables, footnotes, charts, appendices, scanned pages, and contradictions. An AI that can analyze documents with AI does not automatically make every summary trustworthy. You still need a workflow that preserves source grounding.

Start with a document map. Ask the model to identify the title, date, author or organization if visible, sections, page ranges, tables, figures, appendices, and hard-to-read areas. Do not ask for the final answer yet. A document map tells you whether the model noticed the parts that matter.

Document map prompt: "Before summarizing, create a document map. Include title, date, visible author or organization, major sections, page ranges if available, tables, figures, appendices, repeated terms, and any areas that appear unreadable. Do not infer missing details. Use Not visible when needed."

PDF extraction prompt: "Extract structured information from this document into a table with columns for claim, number or metric, date or period, entity, source page or section, confidence, and verification note. Only include facts supported by the document. If a field is missing, write Not found."

Long-context synthesis prompt: "Use this document packet to answer this question: [question]. First list the relevant sources or sections. Then summarize the evidence. Then identify contradictions, caveats, and open questions. Finish with a decision memo in bullet form. Do not use outside knowledge unless I ask for it."

Long context AI for documents is powerful when the model can keep a large body of relevant material available at once. Gemini documentation emphasizes long-context use cases, while Anthropic documents PDF support for text and visual content with practical limits. The takeaway is not that one model should always win. The takeaway is that document tasks should be tested with the source material you actually use.

#### Prompt patterns

Good multimodal prompts are structured like a small operating procedure. They define the input, the role, the evidence rules, the output format, and the verification step. This is especially important when the prompt combines image and text, because the model may blend what it sees with what it assumes.

Use this universal multimodal prompt template: "You are helping with [task]. Use only the attached image/document and the context below. First list observable evidence. Then extract the requested fields. Then provide analysis. Mark uncertainty clearly. Return the final answer as [table/checklist/memo/JSON-style fields]. Context: [context]. Fields or questions: [fields]."

Structured extraction template: "Extract these fields: [field list]. For each field, include value, source evidence, confidence, and verification note. If the source does not contain the answer, write Not found. Do not guess." This works for invoices, competitor pages, charts, PDFs, onboarding forms, support screenshots, and research notes.

Image plus document template: "Compare this screenshot with the attached document. Identify where the screenshot matches the documented process, where it differs, and what a user should do next. Use a table with columns for observed item, document reference, match status, risk, and recommended action."

QA template: "Review your previous answer against the source. Find unsupported claims, missing caveats, unreadable areas, incorrect numbers, and assumptions. Return a corrected version and a short list of changes." This prompt is boring in the best way. It catches errors before they become embarrassing.

For recurring work, save prompts as workflows rather than one-off questions. A reusable prompt pack might include screenshot triage, chart extraction, PDF map, evidence table, decision memo, and verification pass. The goal is not to become a prompt artist. The goal is to make reliable work easier to repeat.

#### Which model to pick for multimodal tasks

The best model for multimodal AI depends on the input and the output. Use Gemini as a strong candidate when the task involves long context, large document packets, or multimodal source review. Use Claude as a strong candidate for careful reading, visual analysis, PDF workflows, and structured reasoning over source material. Use OpenAI models as strong candidates for broad productivity workflows, image-and-text tasks, drafting, coding, structured outputs, and turning analysis into polished deliverables.

| Task | Start with | What to check |
| --- | --- | --- |
| Large PDF packet | Gemini or Claude | Coverage, page/section grounding, missed caveats |
| Screenshot or UI review | Claude or OpenAI | Visible evidence, accessibility issues, actionable fixes |
| Chart or dashboard analysis | Gemini, Claude, or OpenAI | Number accuracy, labels, uncertainty around unreadable values |
| Image plus written instructions | OpenAI, Claude, or Gemini | Whether the model follows both visual and text constraints |
| Final memo or client-ready summary | OpenAI or Claude | Structure, tone, traceability, and cleanup required |

If you are comparing [Gemini vs ChatGPT](https://whizi.io/resources/gemini-vs-chatgpt), start with the same image or PDF prompt in both and score each output. If your task is mainly PDF review, use the deeper [summarize PDFs with AI](https://whizi.io/resources/summarize-pdf-with-ai) workflow. The winner should be the model that produces the most accurate, usable, and verifiable output for your source material.

Whizi makes this easier because you can test models in one place instead of maintaining a separate workflow for every assistant. Pick one real task from your week: a screenshot, PDF, customer document, product image, or competitor page. Run the same prompt across models, compare the evidence, and save the model-plus-prompt combination that performs best.

When you are ready to build a repeatable workflow, create your account at [Whizi registration](https://whizi.io/register). If you are deciding whether a consolidated workspace makes sense for your team, compare options at [Whizi pricing](https://whizi.io/pricing).

Key points:

- Multimodal AI is most useful when you separate observation, extraction, interpretation, and verification
- Image and PDF workflows work best when the prompt asks for visible evidence before conclusions
- The right model depends on the input: screenshots, charts, PDFs, long documents, or polished writing outputs

Checklist:

- Ask for observable evidence before interpretation
- Use structured fields when extracting from images, charts, PDFs, or screenshots
- Tell the model to mark unreadable, cropped, blurry, or missing information
- Separate extraction prompts from analysis prompts for higher accuracy
- Run a verification pass before relying on numbers, claims, dates, or recommendations
- Compare the same multimodal prompt across models before saving a workflow
- Use Whizi to keep model selection flexible instead of picking one model forever

FAQ:

Q: What is an example of multimodal AI?
A: A common example is uploading a screenshot or PDF and asking the AI to extract visible details, summarize the content, identify issues, and return a structured table or memo.

Q: Can multimodal AI read images accurately?
A: It can be useful, but accuracy depends on image quality, readable text, crop, resolution, and task complexity. Ask the model to separate visible evidence from interpretation and flag anything unclear.

Q: Which AI model is best for multimodal work?
A: There is no permanent winner. Gemini, Claude, and OpenAI models can all be useful depending on whether the task involves long documents, images, PDFs, structured extraction, or polished writing. Test the same prompt across models.

### One subscription for all AI models: what you actually get

URL: https://whizi.io/resources/one-subscription-all-ai-models/
Updated: August 2026

Quick answer: Yes, several products sell one subscription covering GPT, Claude, and Gemini together, including Poe, ChatHub, TypingMind, and Whizi, which publishes this page. Plans usually run $10 to $30 a month against roughly $60 for the three direct subscriptions. What you give up is each provider's deepest integrations and first access to new features.

A plain explanation of the multi-model AI subscription category: what it is, what you get and give up compared to paying each provider directly, and how to evaluate one before you buy.

#### The short answer

Yes. Several products sell a single subscription that covers GPT, Claude, and Gemini together, so the category is real rather than a marketing claim. Whizi, which publishes this page, is one of them. Poe, ChatHub, and TypingMind are others, and they work differently enough from each other that the right pick depends on how you actually work.

One plan in this category usually costs between $10 and $30 a month. Buying ChatGPT Plus, Claude Pro, and Google AI Pro directly runs close to $60 a month in the United States as of August 2026. That gap is the entire reason the category exists.

What that price does not buy is the deep product work each provider builds around its own model, like Claude Code in a terminal or Gemini inside Google Docs. That trade is covered plainly further down, because for some people it is the reason to skip this category and pay a provider directly instead.

#### What this category actually is

A multi-model AI subscription is a single paid plan that gives you access to chat models from more than one company, usually GPT from OpenAI, Claude from Anthropic, and Gemini from Google, inside one app with one login and one monthly charge. You are not buying a new AI model. You are buying access to models that already exist, bundled and resold as a subscription, with an interface built around comparing and switching between them.

The category exists for a plain structural reason: every major AI provider prices by seat, not by task. A ChatGPT Plus account costs money. A separate Claude Pro account costs money. A separate Google AI Pro account costs money. None of them lets you touch the other two. If your work genuinely benefits from more than one model, the direct route is to open three accounts, enter three cards, and pay three bills, and the total is the sum of the three, not some blended discount. As of August 2026, ChatGPT Plus is $20 a month, Claude Pro is $20 a month, and Google AI Pro (the plan that replaced the Gemini Advanced name) is $19.99 a month, so three direct subscriptions run close to $60 a month in the United States, and local pricing differs by country. Price the plan you would actually buy against the specific plans you would cancel, not a category average, since the arithmetic changes person to person. A multi-model subscription is the answer to that specific arithmetic problem: one price that includes several models instead of three prices that each include one.

This is different from a model itself, and different from a router API that developers use to call several providers from code. It is a consumer product: a chat app, usually with history, documents, and sometimes image generation, that happens to let you pick which underlying model answers your message. If you have compared it against [ai subscription costs](https://whizi.io/resources/ai-subscription-costs) and found your own stack adds up to more than one plan, this is the category built for that exact situation.

#### Why nobody but an outsider can build this

OpenAI is not going to sell you Claude access inside ChatGPT, and Anthropic is not going to sell you Gemini access inside Claude. Each company has an obvious commercial reason to keep you inside its own app, so if the three are ever going to sit behind one bill, the bundling has to be done by someone with no model of its own to protect. That is the whole business model of this category: it does not build models, it buys access to several providers' models and resells it under one subscription, and it makes money on the margin between what serving you costs and what you pay for not managing three accounts. Everything good and everything limited about these products follows from that one fact.

#### What you genuinely get

Five things the category delivers, done properly, that paying providers separately does not.

- **One bill instead of several.** A single monthly charge replaces multiple recurring charges on your card, which matters more than it sounds like once you have three or four AI line items to track.
- **One history instead of scattered ones.** Your conversations, documents, and generated images live in one place rather than split across ChatGPT, Claude, and Gemini's own separate histories.
- **Model switching inside a single workflow.** You can draft with one model and hand the draft to another for a critique, in the same conversation, instead of copying text between four open tabs. [How to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers the actual workflows this enables, like drafting with one model and having a second one edit it.
- **A cheaper way to try models you would not otherwise pay for.** If you are curious whether Claude writes better than GPT for your kind of work, or whether Gemini's huge context window helps with a big document, a shared subscription lets you find out for a fraction of what three separate trials would cost.
- **Possibly lower total cost, depending on what you would otherwise buy.** Most of these plans sit somewhere between $10 and $30 a month, so one of them beats two or three direct subscriptions at around $20 each. It does not automatically beat a single provider's own standard-tier account, so do the subtraction with the specific plans you would cancel, not with the category averages.

None of this requires taking any vendor's word for it. Take a real task from last week, the reply you had to write to an unhappy customer or the spreadsheet formula that kept returning an error, and send the exact same prompt to GPT, Claude, and Gemini in one afternoon. Read the three answers side by side. Either two of them are clearly better than the third for your kind of work, in which case the switching is worth paying for, or all three are close enough that you should pick the cheapest single account and stop reading comparison articles.

#### What you genuinely give up, stated plainly

Five things you give up, which the pages selling this category tend to leave out.

1. **First access to new features.** When a provider ships something new, a live-search mode, a new voice capability, a new agentic tool, it ships inside that provider's own app first. A multi-model subscription adds support later, sometimes weeks later, sometimes not at all if the feature depends on infrastructure the provider does not expose through an API.
2. **The deepest product integrations.** Gemini inside Google Docs and Gmail, Claude Code running in a terminal against your actual repository, ChatGPT's custom GPTs and connectors to your other tools. These are built by each provider around its own ecosystem, and no aggregator, however good, replicates them. If your daily work depends on one of these specifically, the aggregator gives you the underlying model but not the tool built around it.
3. **The provider's own mobile app and its platform hooks.** Native Siri or Google Assistant integration, an app that's tuned by the company that trained the model, voice modes built by the same team that built the model. A multi-model app builds its own mobile experience, which can be very good, but it is not the same artifact as the provider's own app.
4. **Sometimes, the newest model on day one.** Providers occasionally gate a brand new model to their own direct subscribers for a period before it reaches third-party access. Nobody publishes how long that window runs and it has varied by release, so treat any promise of same-day parity as a claim to check rather than a feature.
5. **Direct account history with that one provider.** If you eventually want a native ChatGPT or Claude account of your own, for a feature only available there, your usage inside a third-party app does not carry over. It is a separate account with a separate history.

None of this is a defect in a specific product. It is a structural fact about being a reseller of access rather than the company that trains the model.

#### Who this category suits, and who it does not

| You are | Direct provider subscription | Multi-model subscription |
| --- | --- | --- |
| A heavy daily user of one model for one deep workflow, like Claude Code in a terminal, Gemini inside Docs, or an agency built entirely on one vendor's tools | Better fit, you need the product depth | Weaker fit, you lose the integration you actually use |
| Someone who wants the newest feature the day it ships, no matter the provider | Better fit for that one provider | Weaker fit, features arrive on a lag |
| Someone who uses two or more models regularly for general writing, research, or coding help | Weaker fit, you are paying for and managing several accounts | Better fit, one bill and one place to compare |
| Someone still figuring out which model actually suits their work | Weaker fit, you are guessing before you have tried alternatives | Better fit, cheap way to compare before committing |
| A team that wants everyone using a consistent set of tools without managing several vendor billing relationships | Weaker fit administratively | Better fit, one invoice and one admin panel |

The honest version, stated once more: if you are a heavy single-model user who lives inside one provider's deep product features every day, going direct to that provider is very likely the better and sometimes the cheaper option. A multi-model subscription earns its price when your actual work already spans more than one model, or when you have not yet found out which model suits you and want to find out cheaply. It does not earn its price for someone who has already settled on one tool and uses its native integrations daily.

#### How to evaluate a multi-model subscription before you pay

The products here are easy enough to name. Poe from Quora is the oldest and the widest catalogue, and it charges in compute points rather than flat messages. TypingMind wants you to bring your own provider API keys, which is the cheapest route if you already have those keys and the worst one if you do not. OpenRouter is not a consumer subscription at all: it is a pay-as-you-go router that you buy credits for and call from code, so it only belongs on your list if you are comfortable metering your own usage. ChatHub is a browser extension that puts several chat windows side by side. Whizi, which publishes this article, is a flat monthly plan. [Poe vs TypingMind vs Whizi](https://whizi.io/resources/poe-vs-typingmind-vs-whizi) compares them properly, including where the others win, and [the best all-in-one AI platforms](https://whizi.io/resources/best-all-in-one-ai-platforms) maps the wider category, including the bring-your-own-key clients and the self-hosted tools. What no comparison can do for you is stay current, because prices and limits in this category move every quarter, so check these six things yourself against whatever you are considering, whoever is selling it.

1. **Which specific models are included, by name, not by category.** "Access to top AI models" is marketing language. A real answer names GPT, Claude, and Gemini specifically, and tells you whether that means the current versions or an older, cheaper one substituted in behind the scenes. If a page will not name the actual models on the plan you would pay for, that is itself an answer.
2. **What the usage limit actually is, in the unit that matters to you.** "Unlimited" rarely means unlimited; it almost always means a soft cap, a rate limit, or a drop to a smaller model once you are heavy on a given day. Find the specific number or the specific mechanism before you commit, not after your first busy week.
3. **What happens above the limit.** Do you get blocked until the next billing cycle, dropped to a weaker model, or charged extra? All three exist across different products in this category, and they are not equally acceptable, so find out which one applies before you pay.
4. **Whether history and documents live in one place or per model.** Part of the value of this category is a single continuous history. Some products keep separate threads per model under the hood, which quietly leaves you with several disconnected tools behind a single interface.
5. **How the free trial or entry offer actually converts.** A one dollar or heavily discounted first period is common in this category. Confirm the price it rolls into, the billing date, and how many clicks cancellation takes, before you enter a card.
6. **Whether cancellation is actually simple.** Test this before you need it if you can, or at minimum read the cancellation policy rather than assuming. A subscription that is easy to start and hard to stop is a pattern worth watching for in any recurring charge, not just this category.

Run this checklist against whatever you are considering, including Whizi. If a product will not give you a straight, specific answer to all six, that is worth weighing on its own, separate from price.

#### A word on the parts of this category to avoid entirely

Search results for this topic include listings for shared logins, resold accounts, and services that route around a provider's country availability or payment restrictions. These are not a cheaper version of the same category. Consumer plans from OpenAI, Anthropic, and Google are sold per person, their terms are written around one account holder rather than a group, and shared logins get suspended without warning, taking the history and the paid time with them. Read each provider's terms if you want the exact wording. Reselling a single account across many buyers is the same problem at scale.

Country availability is the other thing to be straight about. Every one of these providers publishes its own supported countries list, OpenAI, Anthropic, and Google alike, and several of them do not operate everywhere. If a signup or a checkout fails for that reason, it is a location limit rather than a problem with your card, and no amount of retrying with a different payment method fixes it. Check the provider's own availability page before you spend an evening on it, and if the answer is no, the honest response is to use a service that does operate where you live. If you have hit a hard wall with one provider specifically, [ChatGPT alternatives](https://whizi.io/resources/chatgpt-alternatives) covers legitimate options rather than ways around the restriction.

A legitimate multi-model subscription is a company with its own commercial agreements to provider APIs, its own infrastructure, and its own terms with you as a paying customer. That is a completely different thing from a shared password in a spreadsheet, and it is worth being able to tell the two apart before you type a card number into either.

Key points:

- The category exists because every provider prices per seat, so using GPT, Claude, and Gemini directly means paying three separate monthly fees for three separate accounts
- You give up the provider-native apps, first access to new features, and the deepest integrations. You keep model access, one bill, and one place to compare answers
- A person who lives inside one provider all day, using its coding tool or its Docs integration, is usually better off paying that provider directly than joining this category

Checklist:

- Add up what you would pay for direct subscriptions to every provider you actually use
- Name the specific models included in any multi-model plan you are considering, not the category name
- Find the real usage limit and what happens when you hit it, in writing
- Check whether history and documents stay in one place across models or split per model underneath
- Confirm the price an entry offer rolls into and how cancellation actually works
- Decide honestly whether you live inside one provider's deep product features daily, and if so, price that provider directly
- Never use a shared or resold account to get around a price or a country restriction

FAQ:

Q: I am paying for three AI subscriptions. How do I consolidate them?
A: Spend one normal week noting which model you actually opened for which task, rather than guessing from memory. Then check that a single multi-model plan covers those specific models by name, move your recurring work into it for a week while the old plans are still running, and only cancel once you have confirmed nothing you depend on is missing. Cancel in the place you subscribed, since a plan bought through the iOS App Store is cancelled in Apple subscription settings rather than on the provider website, and let each plan run to the end of the period you already paid for.

Q: What is a multi-model AI workspace, and who makes one?
A: A multi-model AI workspace is a chat application that puts models from several different companies behind one login and one bill, so you can send the same question to GPT, Claude, and Gemini and compare the answers without opening three accounts. The companies that make one are independents rather than the model providers themselves, because OpenAI has no reason to sell Claude access and Anthropic has no reason to sell Gemini access. Whizi publishes this page and is one. Poe from Quora is the oldest and widest. ChatHub, TypingMind, and several newer entrants take different approaches to pricing and to whether you bring your own API keys.

Q: Is there really one subscription that includes ChatGPT, Claude, and Gemini?
A: Yes, several products package access to models from OpenAI, Anthropic, and Google under one subscription, including Whizi. What you get is access to the underlying models through a shared app, not a login to each provider's own native app. The provider-built apps, their newest features, and their deepest integrations, like Claude Code or Gemini in Google Docs, stay separate from what an aggregator can offer.

Q: Is a multi-model subscription cheaper than paying for ChatGPT Plus, Claude Pro, and Gemini separately?
A: It depends on which plans you are replacing. As of August 2026, ChatGPT Plus and Claude Pro are $20 a month each and Google AI Pro is $19.99 a month in the United States, so all three together run close to $60 a month. Multi-model plans generally sit well below that, which is the whole reason the category sells. Compare the specific plan you would buy against the specific plans you would cancel, and check local pricing, since these figures are United States ones and Google in particular varies by country.

Q: What do I lose by using an aggregator instead of the providers directly?
A: You lose first access to brand new features, the provider's own mobile apps and platform integrations (like Siri or Google Assistant hooks), and the deepest product features built around one vendor's infrastructure, such as Claude Code or Gemini inside Google Docs. You keep access to the general chat models themselves, usually at a lower combined price if you use more than one.

Q: Who should not use a multi-model AI subscription?
A: Anyone whose daily work depends on one provider's deep product integration, such as an engineer living in Claude Code, or someone whose entire workflow runs through Gemini inside Google Workspace. For those cases, a direct subscription to that one provider is very likely the better fit, and often the better value too, since you are paying for the integration you actually use rather than for models you don't.

Q: Are cheap resold or shared AI accounts a safe way to get several models for less?
A: No. Consumer plans from OpenAI, Anthropic, and Google are sold per person and their terms are not written for a login passed around a group, so a shared or resold account can be suspended without warning, taking the history and any paid time with it. A legitimate multi-model subscription has its own commercial agreement with each provider and its own billing relationship with you, which is a fundamentally different thing from a shared password.

### How to pay for ChatGPT without an international card, and why cards get declined

URL: https://whizi.io/resources/paying-for-ai-when-your-card-is-declined/
Updated: August 2026

Quick answer: Buy the subscription through the Apple App Store or Google Play, where the charge runs on the store's payment rails and can come from a locally bought gift card or carrier billing. Or check OpenAI's own checkout for a local method: UPI in India, Pix in Brazil, GoPay in Indonesia, and local Korean options.

Card declined for ChatGPT, Claude or Gemini? The six real causes, how to tell which one you hit from the error, the legitimate fixes, and the cases no fix reaches.

#### The short answer

If your card has no way out of your country at all, two routes skip the card problem instead of solving it. Buy the subscription inside the Apple App Store or Google Play, where the charge runs on the store's payment rails and can come from a gift card balance bought locally or from carrier billing. That one only exists if the app is actually listed in your country's storefront, so check that before you buy a gift card. Or check the provider's own checkout page for a local method: OpenAI's checkout has shown UPI in India, Pix in Brazil, GoPay in Indonesia and local Korean options. Both routes are set out further down, and for a lot of people one of them is the whole answer.

If your card should work and was refused anyway, the rest of this article is for you, and it starts in a place most people skip. In almost every case the AI company did not reject you. Your bank did. Your card details travel to your issuer, which decides yes or no in about a second, and the merchant gets back a short code and nothing else. OpenAI says this in its own billing help: it typically does not receive detailed decline reasons from banks, and it points you at the bank as the fastest route to an answer.

So the support chat at the AI company cannot tell you why, while your bank can see it on the first screen the agent opens. The most useful sentence in this article is the one you say to them: "Please enable international transactions and online, card not present transactions on this card, and release any fraud block on a charge to a foreign online merchant." That resolves a large share of declines. The rest of this article is for the cases where it does not, plus one honest exception at the end: in a few countries no payment method works at all, because the service is not legally available there.

#### The six things that actually cause it

Declines feel random because the error text is deliberately vague. They are not. Nearly all are one of these six, and each has a different owner.

| Cause | What is happening | Who fixes it |
| --- | --- | --- |
| International use not enabled | Many banks ship debit cards with cross-border use switched off | You, by calling the bank or toggling it in the app |
| Online (card not present) use not enabled | A separate switch. The card works in local shops and fails on every website | You, same call |
| A domestic-only card network | Local schemes such as RuPay, Elo, Verve and Troy are built for payments inside their own country, so a domestic variant has no route to a foreign merchant. Russia's Mir is a separate case, being sanctioned rather than merely domestic | Nobody. You need a different card or route |
| A first foreign charge fraud block | The issuer sees an unfamiliar merchant, a foreign currency and a recurring setup, and blocks it | You, by approving the prompt you missed |
| 3D Secure did not complete | The bank demanded step-up authentication and the popup was blocked or the redirect died | You, by allowing popups and dropping the VPN |
| Card, currency or address mismatch | Your billing address does not match bank records, or the card cannot take the currency shown | You on the address, the bank on the currency |

Prepaid and reloadable cards are the most common near miss, and they fail structurally rather than on balance: many carry no verifiable billing address for the address check, and many are not enabled for recurring charges, which a subscription needs at renewal even when the first charge clears.

#### Reading the error to work out which one you hit

Much of the industry processes cards through Stripe, which publishes its decline codes with what each one means. You rarely see the raw code, but the wording on your screen derives from it, so the phrasing is a real clue.

| What you see | Code | What it means |
| --- | --- | --- |
| "Does not support this type of purchase" | card_not_supported | Ask the issuer to allow this kind of transaction |
| "Your card was declined", no detail | generic_decline, do_not_honor, transaction_not_allowed | The issuer refused and gave no reason, so in most cases only the bank can tell you. One exception worth knowing: Stripe documents generic_decline as also covering payments its own fraud screening stopped. The most common case by far |
| An authentication screen that failed | authentication_required | 3D Secure was demanded and did not complete. Retry with popups allowed |
| "Does not support this currency" | currency_not_supported | The card cannot take the currency shown. Ask the issuer, or look for local currency pricing |
| "Incorrect postal code" or address | incorrect_zip, incorrect_address | Type the address exactly as your statement has it |
| A funds or limit message | insufficient_funds, card_velocity_exceeded | Balance, credit limit or a cap. Some banks cap foreign spending separately |
| "Contact your card issuer" | call_issuer, restricted_card | The bank has flagged the card. A phone call, not a retry |

Two notes. A code saying the issuer refused for an unknown reason is genuinely unknown to the merchant, so a support agent who tells you exactly why is guessing. And stop after two failures: repeated attempts look like card testing, and a soft block can harden into a locked card.

#### The fixes, in the order worth trying them

Work down this list. It is ordered by how often each one turns out to be the answer.

1. **Check the obvious three.** Number, expiry, security code, then the billing address exactly as your bank statement prints it, postal code included. Address mismatches are extremely common where the local format differs from the US one the form expects.
2. **Look at your phone.** Many blocks arrive as an SMS or in-app prompt asking you to confirm the transaction. Approve it, retry within a few minutes, and it often clears with no phone call.
3. **Call the bank and ask for both switches by name.** International transactions and online card not present transactions are separate settings at most issuers, often exposed in the app under card controls or travel settings. Ask them to whitelist the merchant too.
4. **Retry cleanly.** VPN off, popups and redirects allowed, ad blocker disabled, cookies cleared. OpenAI lists all of these itself, because 3D Secure is a redirect and anything that eats redirects eats the payment.
5. **Try a credit card rather than a debit card.** Credit cards are far more often enabled for cross-border and recurring use out of the box. And if you are charged in USD on a card that cannot take USD, the issuer refuses regardless of balance.
6. **Use the app store route.** A subscription bought through the Apple App Store or Google Play runs on the store's payment rails, not the merchant's. Apple accepts Apple Account balance, cards, PayPal and carrier billing depending on country. Google Play accepts cards, PayPal, Play balance and carrier billing with some carriers. Buying a gift card in your own country for your own country's store is entirely legitimate, and for many people it is the whole solution.
7. **Look for a local payment method.** OpenAI's checkout has shown UPI in India, Pix in Brazil, GoPay in Indonesia and local Korean options, and the list keeps changing. Read the payment page before assuming a card is the only choice.

Two cautions. On the store route, prices there are sometimes higher than the web price because the store takes a cut, and cancelling runs through the store rather than the provider, which [How to cancel ChatGPT Plus](https://whizi.io/resources/how-to-cancel-chatgpt-plus) covers and is worth knowing before you subscribe rather than after. On local methods, which plans a local method covers can differ from which plans a card covers, so check the plan you actually want rather than the cheapest one on the page.

#### The services that advertise a way around this, and when the answer is genuinely no

Search this problem and the first results are virtual card sellers, crypto top up services and account resellers, at least one advertising that it runs no identity checks. They sell relief, and they are a bad deal for reasons that have nothing to do with morality.

- **The account is what you are risking.** Providers suspend accounts over payment patterns that look like circumvention, and losing the account means losing the history, projects and files inside it.
- **No identity checks is a fact about the operator, not a feature for you.** Regulated issuers run those checks because they must, so a service advertising their absence is telling you which rules it sits outside. You are the one holding a balance with it.
- **Prefunded balances are unsecured deposits.** When these operators disappear, and they do, the money is gone. No chargeback, no regulator.
- **A resold account has someone else's hand on it.** Whoever sold it can usually still reset the password, and everything you type passes through their access.
- **It often does not even work.** Many of these cards fail exactly the checks above, so people pay a fee, top up a balance and still get declined.

And here is the part every article selling something skips. Some countries cannot be billed at all, and no payment trick changes that. OpenAI publishes a supported countries and territories list, and states that accessing or offering access to its services outside those territories may result in your account being blocked or suspended. Comprehensive US sanctions programs are law rather than company preference, and as of August 2026 they covered Cuba, Iran and North Korea along with the Crimea, Donetsk and Luhansk regions of Ukraine. That list moves, in both directions: the comprehensive Syria program was terminated in 2025 and many articles still list it, and the Ukraine order lets the US Treasury designate further regions. Check OFAC's own sanctions programs page for the current position rather than any article, including this one. Most other absences are commercial or regulatory decisions, and those do change. The two are easy to confuse and worth telling apart before you spend an afternoon on payment methods, which [ChatGPT is not available in your country](https://whizi.io/resources/chatgpt-not-available-in-your-country) sets out in full.

If that is you, the honest position is that this product is not available, and the advertised workarounds risk your account, your money and in some places your legal position. What does remain is real: open weight models such as the Llama, Qwen, Mistral and DeepSeek families run on your own hardware with no billing relationship at all, and several assistants have free tiers that need no card. [Free ChatGPT alternatives](https://whizi.io/resources/free-chatgpt-alternatives) covers the ones that are usable rather than just technically free, and [is ChatGPT Plus worth it](https://whizi.io/resources/is-chatgpt-plus-worth-it) is worth settling before you chase a payment method at all.

Key points:

- Almost every decline comes from your own bank, and the merchant is rarely told the real reason
- The boring fix usually works: ask your bank to enable international and online transactions on the card
- Some countries cannot be billed at all, and the virtual card sellers advertising a way around it do not change that

Checklist:

- Assume your bank refused, not the AI company, and start the conversation there.
- Ask the bank for two separate things: international transactions and online card not present transactions.
- Check your phone for a fraud confirmation prompt before anything else.
- Type the billing address exactly as it appears on your bank statement.
- Retry with the VPN off, popups allowed and the ad blocker disabled so 3D Secure can complete.
- Stop after two failed attempts so you do not trigger a hard block.
- Check the checkout page for a local payment method, then check the app store route.
- Avoid prepaid cards for subscriptions: they often clear once and fail at renewal.
- Never buy accounts, resold subscriptions or no identity check cards.

FAQ:

Q: How can I pay for ChatGPT without an international card?
A: Check the checkout page for a local payment method first. OpenAI has added local options in several markets, including UPI in India, Pix in Brazil, GoPay in Indonesia and local Korean methods, and the list changes. The other legitimate route is subscribing through the Apple App Store or Google Play, which runs on the store's own payment methods, including your country's gift card balance and carrier billing where supported.

Q: Why does my debit card work everywhere else but not for an AI subscription?
A: Two settings most people never see. Many banks ship debit cards with international transactions off, and with online card not present transactions as a separate switch. A card with both off works in local shops and fails on every foreign website. Ask your bank to enable both by name. If the card is on a domestic-only network such as RuPay, Elo, Verve or Troy, there is no route to a foreign merchant and no setting fixes it. Check the card itself before concluding that, though: several of those networks also issue internationally co-badged versions, and a card carrying a second network mark is not domestic-only.

Q: Can I use a prepaid or virtual card?
A: Sometimes for the first charge, often not for the renewal. Prepaid cards frequently carry no verifiable billing address for the address check, and many are not enabled for recurring billing, which a subscription needs every month. Separately, do not buy cards from services advertising no identity checks: those balances are unsecured and the operators disappear.

Q: Will a different browser or a VPN fix a declined card?
A: A clean browser sometimes helps, and a VPN usually hurts. 3D Secure runs through a popup or a redirect, so ad blockers, popup blockers and stale cookies can break a payment that would otherwise succeed. OpenAI lists clearing cache, allowing popups and disabling blockers in its own decline guidance. A VPN adds a location mismatch on top.

Q: What if my country is not supported at all?
A: Then the service is genuinely unavailable and no payment method changes it. OpenAI states that accessing its services outside the listed territories may result in your account being blocked or suspended. Check first whether your country is merely absent from a commercial list, which does change, or covered by a comprehensive US sanctions program, which is law. Either way, what remains is real: open weight models such as Llama, Qwen, Mistral and DeepSeek run on your own hardware with no billing relationship at all, and several assistants have free tiers that need no card.

### Perplexity vs ChatGPT: an answer engine and an assistant

URL: https://whizi.io/resources/perplexity-vs-chatgpt/
Updated: August 2026

Quick answer: Perplexity is an answer engine and ChatGPT is an assistant, so the choice depends on whether your work is more finding out or making something. Perplexity searches first and cites every claim, which suits research you have to defend. ChatGPT is stronger for writing, coding, file work and anything you produce.

Perplexity and ChatGPT solve different problems. Here is what each one is built for, where citations genuinely help, and why the answer for most people is not one of them.

#### They are not the same kind of product

Perplexity answers questions. ChatGPT helps you do work. That distinction explains almost every difference between them, and it is why side-by-side feature tables tend to mislead.

Ask both "what changed in the EU AI Act enforcement timeline this quarter" and Perplexity will give you a tight answer with numbered sources you can click, drawn from pages published this week. ChatGPT will give you a longer, more conversational answer that may be excellent and may quietly be six months stale.

Now ask both to take that answer and turn it into a client briefing note with three recommendations and a slide outline. ChatGPT does that comfortably. Perplexity will produce something, but you are using a research tool as a writing tool and it shows.

So the real question is not which is better. It is which half of your work is bigger: finding out, or making something.

#### What Perplexity does that ChatGPT does not

**Sources attached to every claim.** This is the product. Each statement carries a numbered citation you can click, so checking whether an answer is trustworthy takes seconds instead of a fresh search. For anyone whose work gets checked by someone else, this changes the economics of using AI for research at all.

**Genuinely current information by default.** Perplexity searches first and answers second, every time. ChatGPT browses when it decides browsing is needed, which means you sometimes get a confident answer built from training data without realising it. For anything time-sensitive, "always searches" beats "usually searches".

**Research that follows a thread.** Suggested follow-ups are good, and the flow of question, answer, sharper question is where the product feels designed rather than assembled. Twenty minutes in Perplexity often covers ground that would take an hour of tabs.

**Focused search scopes.** Restricting to academic papers, or to discussion forums, or to the general web, is a small feature with a large effect. Academic mode alone makes it a serious tool for literature scanning.

**Spaces for ongoing projects.** Collections with their own uploaded files and instructions, so a long-running research topic keeps its context instead of starting over.

**Speed.** For a factual question with a findable answer, it is simply the fastest route from question to sourced answer that exists right now.

#### What ChatGPT does that Perplexity does not

**Sustained work on one thing.** Drafting, revising, arguing about structure, rewriting the third paragraph again. ChatGPT holds a working session; Perplexity is built around discrete questions.

**Making things rather than finding things.** Code, documents, spreadsheets, images, slide outlines, data analysis with actual execution. This is most of what people use AI for and it is not what an answer engine is for.

**Deep file work.** Upload a stack of documents and interrogate them across a long session. Perplexity handles files, but this is ChatGPT's home ground.

**Creative and voice-driven work.** Fiction, brand voice, tone matching, comedy, brainstorming that goes somewhere odd and useful.

**Building on top of it.** Custom assistants, connectors, automations, a mature API. If you want the tool to become part of a workflow rather than a tab you visit, that is a different category of product.

**Reasoning without searching.** Some problems need thinking, not sources. "Why is this function returning undefined" is not a search query, and a tool that reflexively searches the web is the wrong shape for it.

#### The citation trap

This is the most important paragraph in the article, so it gets its own section.

**A citation is a pointer, not a proof.** Perplexity attaches sources to claims, and most of the time the source supports the claim. Not always. The failure mode is subtle: the answer says something slightly stronger, slightly more specific, or slightly differently scoped than the cited page actually says. It is not fabrication, it is drift, and it is much harder to catch than an obvious hallucination because the citation makes it look checked.

It also inherits whatever it finds. If the top results are a content-marketing blog, an SEO listicle, and a forum post, you get a well-cited summary of those. The confident presentation is identical whether the underlying sources are peer-reviewed or paid placement.

What to actually do about it:

- **Click through on anything that matters.** Any number, date, or claim you will repeat to someone else. This takes ten seconds and it is the entire point of having citations.
- **Look at what got cited, not just how many.** Eight sources that are all aggregators of one press release is one source.
- **Be suspicious when every source agrees perfectly.** Genuine questions usually have some disagreement in the literature, and its absence often means you are reading one narrative repeated.
- **Use academic scope for anything technical or medical**, and treat general web results on those topics as a starting point only.

None of this makes Perplexity untrustworthy. It makes it a research assistant rather than a research replacement, which is a perfectly good thing to be. The tool is at its best in the hands of someone who was going to check anyway and now checks faster.

#### Task by task

| Task | Better tool | Why |
| --- | --- | --- |
| Fact-checking a specific claim | Perplexity | Citations make verification fast |
| Market or competitor research | Perplexity first, then ChatGPT | Gather with sources, then synthesise into a deliverable |
| Academic literature scanning | Perplexity | Scoped search over papers is a genuine advantage |
| Writing a report or article | ChatGPT | Sustained drafting and revision |
| Coding and debugging | ChatGPT | Not a search problem |
| Analysing your own files | ChatGPT | Deeper file handling and code execution |
| Current events and prices | Perplexity | Always searches, always dated |
| Brainstorming | ChatGPT | Perplexity keeps trying to find the answer rather than invent one |
| Learning an unfamiliar subject | Both, in order | Perplexity to map the territory, ChatGPT to have it explained |
| Anything with a deadline and an editor | Perplexity to source, ChatGPT to write | The split most professionals land on |

That last row is the honest conclusion, and it is inconvenient. Most people who use both properly end up using both, which is exactly how you arrive at two subscriptions.

#### Pricing, and the thing to know about Perplexity

Both have usable free tiers and both charge around $20 a month for their standard paid plan, with much more expensive top tiers above. Perplexity's free tier gives you a limited number of advanced searches per day, which is enough for casual use. Check current pricing on both sites, since the tiers move.

Here is the part worth understanding before you subscribe: **Perplexity is largely an interface over other labs' models.** Its paid plan lets you choose which model answers, including models from OpenAI, Anthropic, and Google, alongside its own search-tuned models. The value it adds is the search pipeline, the citation layer, and the interface, not a proprietary frontier model.

That is not a criticism. Orchestration is real engineering and the product is good. But it changes what you are buying. You are not choosing Perplexity's model over ChatGPT's model, you are choosing Perplexity's research workflow, which happens to run on the same models everyone else uses.

It also means the comparison people should be making is often between Perplexity and other multi-model tools rather than between Perplexity and ChatGPT. Our [Perplexity alternatives page](https://whizi.io/alternatives/perplexity) and [Whizi vs Perplexity Pro](https://whizi.io/vs/whizi-vs-perplexity-pro) cover that comparison directly.

#### How to decide

Look at your last two weeks of AI use and sort it into two buckets: questions you needed answered, and things you needed made. Whichever bucket is bigger points at your primary tool. If they are close, you are in the group that ends up wanting both.

If you can only pay for one:

- **Choose Perplexity** if you are a journalist, an analyst, a researcher, a student writing sourced work, or anyone who has to defend claims to someone else.
- **Choose ChatGPT** if you write, code, build, or produce things, which covers most people.

If you want both without paying twice, the practical option is a workspace that gives you the underlying models plus web access in one place. [Whizi](https://whizi.io/pricing) includes GPT, Claude, Gemini, Grok, and DeepSeek on one subscription, so the research and the writing happen in the same conversation instead of across two products and two histories. The tradeoff to be clear about: a general workspace does not replicate Perplexity's dedicated citation pipeline, so if sourced answers are the core of your job, that specialist tool earns its place.

For the wider decision about which model to reach for on which task, read [how to choose an AI model](https://whizi.io/resources/how-to-choose-ai-model), or [how to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) for the workflow.

Key points:

- Perplexity is an answer engine and ChatGPT is an assistant, which is why comparing them feature by feature misleads
- Citations change how fast you can verify a claim, not whether you need to
- Perplexity runs on other labs' models, so choosing it is choosing an interface rather than a model

Checklist:

- Sort your recent AI use into questions answered and things made
- Click through citations on any claim you plan to repeat
- Check whether the cited sources are independent or one press release repeated
- Use academic scope for technical, medical, or scientific questions
- Remember that Perplexity runs on other labs' models, so you are buying the workflow
- Decide whether you need a specialist research tool or broader model access

FAQ:

Q: Is Perplexity better than ChatGPT?
A: For sourced research and current information, yes. Perplexity searches first and cites every claim, which makes verification fast. For writing, coding, file work, and anything you are producing rather than looking up, ChatGPT is the stronger tool. They are different categories of product.

Q: Does Perplexity use ChatGPT?
A: Partly. Perplexity runs on models from several labs, including OpenAI, Anthropic, and Google, alongside its own search-tuned models, and paid users can pick which one answers. What Perplexity adds is the search pipeline and citation layer rather than a proprietary frontier model.

Q: Are Perplexity's citations reliable?
A: They are reliable pointers, not proof. The common failure is drift, where the answer states something slightly stronger or differently scoped than the cited page supports. Click through on anything you will repeat, and check whether the sources are genuinely independent.

Q: Is Perplexity Pro worth $20 a month?
A: It is if your work involves defending claims to other people: research, journalism, analysis, or academic writing. If most of your AI use is drafting, coding, or producing deliverables, the same $20 buys more elsewhere.

Q: Can one tool do both research and writing?
A: Mostly. A multi-model workspace with web access covers research and writing in one place and one history, which suits people whose research is a step in producing something. A dedicated answer engine still wins when sourced accuracy is the deliverable itself.

### Poe vs TypingMind vs Whizi: the multi-model apps compared

URL: https://whizi.io/resources/poe-vs-typingmind-vs-whizi/
Updated: August 2026

Quick answer: Poe suits people who want the largest model catalogue and can manage a compute-point allowance. TypingMind is cheapest for heavy technical users willing to set up their own API keys. Whizi is the flat-subscription option for predictable billing and no setup. What separates them is how each one charges you, not the model list.

An honest comparison of the apps that give you many AI models in one place, including Poe, TypingMind, OpenRouter, ChatHub, and Whizi, and how each one charges you.

#### A note on who wrote this

Whizi is one of the tools in this comparison. You should read the rest of it knowing that, and you should be suspicious of any section where a competitor has no strengths.

The reason to write it anyway is that this category has no good honest overview. Every existing comparison is either a listicle that got the pricing models wrong or a vendor page that pretends the other options do not exist. People end up choosing on which name they saw first, then discovering three months later that they picked the wrong billing model for how they work.

So the rule for this piece: **every tool gets a section on what it is genuinely best at, and the recommendation section names cases where you should not pick Whizi.** If you want the marketing version, the [pricing page](https://whizi.io/pricing) is right there and it is doing its job.

#### The thing that actually separates these tools

Every product in this category lists the same models. GPT, Claude, Gemini, Llama, DeepSeek, Grok, Mistral, plus a long tail. Comparing model lists is close to useless, because they all pull from the same providers and the lists change monthly.

What genuinely differs is **how you pay**, and that decides everything else: how predictable your bill is, how much setup you face, whether heavy use is cheap or ruinous, and how much you think about cost while working.

| Billing model | How it works | Feels like | Breaks down when |
| --- | --- | --- | --- |
| Flat subscription | Fixed monthly price, usage included | A normal app | You use it very lightly |
| Credit or point system | Monthly allowance, each model costs different amounts | A prepaid card | You do not know what anything costs until you spend it |
| Bring your own key | You buy a licence, then pay providers directly per token | A power tool | You are not technical, or usage is unpredictable |
| Pure pay as you go | Top up credits, pay per token | An API bill | You want to stop thinking about cost |

Almost every regret in this category traces back to picking the wrong row of that table. A light user on a bring-your-own-key setup pays for complexity they do not need. A heavy user on a credit system spends the last week of every month rationing. Work out which row suits you before you compare anything else.

#### Poe

Built by Quora, and the oldest well-known product in this category. Around $20 a month, with an annual discount and a higher tier above.

**What it is best at:** breadth. Poe has the largest catalogue by a distance, including a huge library of user-created bots with their own prompts and personalities. If you want to try an obscure model the day it is released, or want a ready-made bot for a niche task, Poe usually has it first. The creator ecosystem is genuine, and creators can earn from bots people use, which keeps it stocked.

**What it is worst at:** cost transparency. Usage runs on compute points, and different models consume points at wildly different rates. A conversation with a cheap model costs almost nothing; a few exchanges with a top reasoning model can eat a meaningful chunk of a monthly allowance. You cannot easily predict what a working session will cost until you have done it a few times, and the mental overhead of "can I afford to ask the good model" is a real tax on how you work.

**Pick Poe if:** you want maximum model variety, you enjoy trying new releases immediately, and you are comfortable managing an allowance.

#### TypingMind

A different model entirely: you buy a licence, usually a one-time payment with tiered feature levels, then connect your own API keys from OpenAI, Anthropic, Google, and others. You pay those providers directly for what you use.

**What it is best at:** control and, for the right person, cost. There is no per-seat subscription tax, you pay exactly the provider rate for exactly what you consume, and you get a genuinely well-built interface with prompt libraries, plugins, custom characters, and local chat storage. For a technical heavy user this is frequently the cheapest option available, and it is not close.

It is also the best answer for anyone who wants their conversation history stored locally rather than on someone else's server, and for teams who already have provider accounts with negotiated rates.

**What it is worst at:** everything before the first message. You need accounts with each provider, API keys generated and pasted, billing set up in several places, and enough understanding to know why one model rejected your request and another did not. Costs are also unbounded: a runaway long-context session can produce a surprising bill, because nothing stops you. And there is no free tier to fall back on, since every token is billed by a provider.

**Pick TypingMind if:** you are technical, you use AI heavily, you want the lowest cost per token, and setting up four API accounts sounds like a Tuesday rather than a barrier.

#### OpenRouter

Primarily a routing layer for developers rather than a chat app, though it has a usable chat interface. You top up credits and pay per token across hundreds of models through one API.

**What it is best at:** being infrastructure. One API key, one bill, automatic fallback between providers, transparent per-model pricing, and access to models you cannot easily get elsewhere. If you are building software, this is often the correct choice and the rest of this article does not apply to you.

**What it is worst at:** being a daily workspace. The chat interface is a convenience, not a product. No serious prompt management, no polished document handling, and the whole thing assumes you are comfortable reading a token-price table.

**Pick OpenRouter if:** you are a developer routing model calls, or you want a single API bill across many providers. See our [OpenRouter alternatives page](https://whizi.io/alternatives/openrouter) if you started there and want something more finished.

#### ChatHub, Merlin, and the extension category

A separate shape worth knowing about. These are browser extensions rather than standalone apps, and they sit alongside the sites you already use.

**ChatHub** shows several chatbots side by side in one window, sending your prompt to all of them at once. Its trick is that it can use your existing logged-in web sessions, so you can compare using accounts you already pay for rather than API keys. As a pure comparison tool it is elegant and cheap.

**Merlin** is a broader assistant extension: summarise this page, reply to this email, rewrite this selection, plus chat access to several models on a credit allowance. It is convenient and it lives where you already work.

**Where the category struggles:** these are companions rather than workspaces. Conversation history, document handling, and long project work are weaker, and extensions that depend on your existing web logins are fragile when providers change their sites. They also add a browser extension with broad page access, which is a legitimate thing to think about before installing.

**Pick an extension if:** your main need is comparing answers or acting on pages you are already reading, and your serious work happens elsewhere.

#### Whizi

A flat subscription, with the major models included and no API keys to manage. Sign in and start working.

**What it is built for:** predictability and continuity. One price, one history, and the ability to switch models inside a conversation without losing the thread, so you can draft with one model and have another critique it in the same context. Documents, images, and projects live in the same place as the chat rather than in a separate tool.

**Where it is genuinely the right pick:** you want several frontier models, you do not want to think about tokens, and you would rather pay one predictable price than manage keys or ration points. That describes most people who are not developers, and quite a few who are.

**Where it is not the right pick, plainly:** if you are a very heavy technical user with your own provider accounts, bring-your-own-key will cost you less per token. If you want every obscure model the day it ships, Poe's catalogue is bigger. If you are building software, use an API router rather than a chat product. And if you use one vendor's deep product features every day, such as Claude Code in a terminal or Gemini inside Google Docs, no aggregator replicates those, and you should buy that vendor's plan.

If you want the direct comparisons rather than the category view: [Whizi vs Poe](https://whizi.io/vs/whizi-vs-poe), [Whizi vs OpenRouter](https://whizi.io/vs/whizi-vs-openrouter), [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus).

#### Side by side

Verify all pricing directly with each vendor before deciding, since every product in this category has repriced at least once in the past year.

| | Poe | TypingMind | OpenRouter | Extensions | Whizi |
| --- | --- | --- | --- | --- | --- |
| You pay | Monthly subscription | Licence plus your own API usage | Per token credits | Small subscription or credits | Monthly subscription |
| Setup time | Minutes | An hour or more | Developer setup | Minutes | Minutes |
| Cost predictability | Medium, points vary by model | Low, depends on usage | Low | Medium | High |
| Cheapest for heavy technical use | No | Yes | Yes | No | No |
| Model catalogue | Largest | Whatever you have keys for | Largest for developers | Limited | Major models |
| Switch models mid-conversation | Limited | Yes | Manual | Compare only | Yes |
| Best for | Trying everything | Power users | Building software | Working inside the browser | Daily work without admin |

#### Pick by how you work

Find yourself in this list and stop reading.

- **You use AI a few hours a week for writing, research, and admin.** A flat subscription. The cheapest option on paper is irrelevant if it costs you an evening of setup and a monthly cost calculation.
- **You are a developer with provider accounts and heavy usage.** Bring your own key, through TypingMind for a polished interface or OpenRouter if you are building rather than chatting.
- **You want to try every new model immediately.** Poe. Nothing else keeps up with the catalogue.
- **You need to compare two answers occasionally and already pay for a plan.** A comparison extension is a cheap add-on to what you have.
- **You live inside one vendor's product features.** Buy that vendor's plan. Aggregators give you model access, not product depth, and pretending otherwise would be dishonest.
- **You are paying for two or three separate AI subscriptions right now.** Any tool in this category will save you money. Start by totalling what you spend with the [savings calculator](https://whizi.io/tools/calculator), because the number is usually higher than people expect.

The one thing worth avoiding is what most people do by default, which is to accumulate single-vendor subscriptions one at a time until the total is embarrassing and the workflow is split across four histories. Whatever you choose here, choosing deliberately beats that.

If you want the workflow rather than the shopping decision, [how to use multiple AI models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers what to actually do with more than one model once you have them.

Key points:

- The real difference between these tools is how they charge you, not which models they list
- Bring-your-own-key tools are cheapest for heavy technical users and worst for everyone else
- Whizi built this comparison and tells you where the other options win, because you will find out anyway

Checklist:

- Decide which billing model suits you before comparing model lists
- Estimate your monthly usage honestly, since it decides which option is cheapest
- Count the setup cost, including provider accounts and API keys, as part of the price
- Check whether you depend on a single vendor's product features that no aggregator replicates
- Test switching models mid-conversation, since implementations differ a lot
- Total your current AI subscriptions before adding or replacing anything

FAQ:

Q: What is the best app for using multiple AI models?
A: It depends on how you want to pay. Poe has the largest catalogue on a subscription, TypingMind is cheapest for heavy technical users who bring their own API keys, OpenRouter is the right choice for developers routing calls, and a flat-subscription workspace like Whizi suits people who want predictable billing and no setup.

Q: Is Poe or TypingMind cheaper?
A: For light or moderate use, Poe's flat subscription is usually cheaper and far simpler. For heavy use, TypingMind plus your own API keys is normally cheaper per token, because you pay provider rates with no subscription margin. The crossover point depends on which models you use, since reasoning models cost several times more.

Q: Do these apps give you the same models as the official apps?
A: The same underlying models, yes, through provider APIs. What you do not get are vendor product features built around the model, such as Claude Code, Gemini inside Google Docs, or ChatGPT's custom GPTs and connectors. If those features are central to your work, buy the vendor plan.

Q: Why do Poe's compute points run out so fast?
A: Because models cost very different amounts to run. Top reasoning models consume points many times faster than lightweight ones, and extended thinking generates tokens you never see. Watch which model you have selected, and use cheaper ones for routine questions.

Q: Is it safe to put my API keys into a third-party app?
A: It depends on the app. Prefer tools that store keys locally in your browser rather than on their servers, set spending limits on every provider account, use separate keys per tool so you can revoke one without breaking the others, and rotate them periodically.

### Prompt engineering for beginners: a framework that works

URL: https://whizi.io/resources/prompt-engineering-beginners/
Updated: August 2026

Quick answer: Prompt engineering for beginners is a four-part brief: instruction, context, constraints, and format. Name the action and deliverable, supply the audience, goal, and source material, set rules on length, tone, and evidence, then request a specific output shape. Treat the first answer as a draft and revise one thing at a time.

Learn prompt engineering for beginners with a reusable framework, prompt template, task examples, iteration loop, and checklist for better AI outputs.

#### Why prompts fail

Prompt engineering for beginners is not about memorizing clever phrases. It is about giving an AI model the job, evidence, boundaries, and output shape it needs to produce something useful. Most bad prompts fail because they ask for help without defining what "good" means. The model fills the gaps with likely patterns, and those patterns can sound polished while missing your real goal.

Almost every disappointing answer traces back to one of four things:
1. **The task is unclear.** "Make this better" can mean shorter, warmer, more persuasive, more accurate, less technical, or easier to scan. The model picks one and you get the wrong one.
2. **The context is missing.** Without the audience, the source material, the decision criteria, or the goal, the model optimises for generic quality, which is nobody's quality.
3. **The constraints are weak.** No length, no tone, no sources, no privacy rules, no list of things to avoid, so the answer drifts toward the average of everything it has seen.
4. **There is no review loop.** The first answer is a draft. Treating it as the finished product is the most common mistake of all.

A strong beginner workflow treats prompting as a short brief plus a quality check. You define the task, include the context, set constraints, request a format, then ask the model to expose assumptions or uncertainty. That simple habit improves prompts for writing, prompts for research, planning prompts, and analysis prompts because it turns the model from a vague assistant into a structured collaborator.

#### Framework: instruction, context, constraints, and format

Use this framework for almost every serious prompt: Instruction, Context, Constraints, Format. It is memorable, it works across every model, and it is specific enough to change the output rather than just feel organised.

| Part | What it answers | Weak version | Strong version |
| --- | --- | --- | --- |
| Instruction | What should it do? | "Help with market research" | "Create a competitor comparison table for five products in this category" |
| Context | What does it need to know? | Nothing supplied | Audience, goal, source text, decision criteria, examples |
| Constraints | What are the rules? | None | Under 400 words, no jargon, mark anything uncertain, do not invent numbers |
| Format | What shape should it come back in? | Unspecified | A table with columns for claim, evidence, and confidence |

Instruction is the action. Start with a verb and a deliverable: summarize, compare, rewrite, extract, critique, classify, outline, debug, or plan. Instead of "Help with market research," write "Create a competitor comparison table for five products in this category." The clearer the instruction, the easier it is to judge the output.

Context is the material the model needs to avoid guessing. Include audience, goal, background, source text, examples, decision criteria, and definitions. For long inputs, label the source clearly and put your instructions before or after the content in a consistent structure. Long-context guidance from model providers often emphasizes clear organization because large inputs still need signposts.

Constraints are the rules. Add length, tone, scope, excluded claims, required evidence, privacy boundaries, reading level, and what to do when information is missing. Constraints are especially useful when you need prompts that improve accuracy: tell the model to mark unknowns instead of inventing details, separate facts from assumptions, and cite the source section when possible.

Format is the output container. Ask for a table, checklist, memo, outline, JSON-style fields, email, scoring rubric, or step-by-step plan. Structured output guidance from OpenAI and Gemini is useful because it reinforces the same practical lesson: when the desired shape is explicit, outputs are easier to parse, compare, and reuse.

Copy-paste prompt template: "Instruction: [specific action and deliverable]. Context: [audience, goal, source material, examples, definitions, decision criteria]. Constraints: [length, tone, must include, must avoid, evidence rules, privacy rules, uncertainty rules]. Format: [table/checklist/memo/email/outline/JSON-style fields]. Before finalizing, list assumptions, missing information, and one follow-up question that would improve the result."

#### Examples by task

The best prompts for work are reusable patterns, not one-off magic lines. Use the examples below as starting points, then save the ones that consistently produce useful output.

Writing prompt: "Instruction: Rewrite this draft for clarity and specificity. Context: The audience is [audience], the goal is [goal], and the draft is [paste draft]. Constraints: Keep all facts unchanged, avoid hype, use plain language, and make the next step obvious. Format: Return a revised draft, then a table with original issue, edit made, and reason." This works because it gives the model a job, a reader, a standard, and a reviewable edit trail.

Research prompt: "Instruction: Turn these sources into a research brief. Context: I am evaluating [question] for [decision]. Sources: [paste source excerpts or links/notes]. Constraints: Use only the provided material, separate evidence from inference, mark weak evidence, and do not invent statistics. Format: Return key findings, evidence table, open questions, and recommended next sources." This is better than asking "Research this topic" because it protects traceability.

Planning prompt: "Instruction: Create a two-week execution plan. Context: Goal is [goal], team is [team], deadline is [date], constraints are [constraints]. Constraints: Prioritize reversible steps, identify blockers early, and keep each task owner-friendly. Format: Table with day, task, owner, output, risk, and success check." The output becomes something a team can actually use.

Analysis prompt: "Instruction: Compare these options and recommend one. Context: Options are [options]. Decision criteria are [criteria]. Constraints: Include tradeoffs, strongest counterargument, and what would change the recommendation. Format: Scorecard table plus a short decision memo." This encourages judgment without hiding uncertainty.

Review prompt: "Instruction: Critique this output before I use it. Context: Intended use is [use case]. Output to review: [paste]. Constraints: Check accuracy, unsupported claims, missing context, tone, risk, and places needing human review. Format: Return critical issues first, then suggested edits, then a final confidence rating." This turns AI into a quality-control step rather than only a draft generator.

#### Iteration loop

Good prompt engineering is iterative. You should expect the first response to reveal what the prompt forgot. Instead of starting over, use a loop: evaluate, diagnose, revise, compare, save.

Evaluate the answer against five questions. Did it complete the exact task? Did it use the provided context? Did it obey constraints? Is the format easy to use? Are claims, assumptions, and uncertainties visible? If the answer fails, diagnose the prompt before blaming the model. A vague output often means the task was vague. A generic output often means the context was thin. A risky output often means constraints were missing.

Then revise with one precise instruction at a time. Say "make the recommendation more evidence-based and mark unknowns" instead of "try again." Say "convert this into a table with columns for claim, evidence, confidence, and next check" instead of "be more structured."
When the output matters, run the revised prompt across two models and compare them on four things: accuracy, completeness, tone control, and how much editing each one still needs.

Use this iteration checklist: 1. Name the failure in one sentence. 2. Add the missing context or constraint. 3. Tighten the output format. 4. Ask for assumptions and uncertainty. 5. Run a second model when the task is important. 6. Save the improved prompt only after it works on a real task, not a toy example.

#### Build your reusable prompt library

A reusable prompt library is where beginner prompting becomes a daily workflow. Instead of keeping random prompts in chat history, organize templates by job: writing, research, planning, coding, customer analysis, document review, and decision support. Each template should include the same four fields: instruction, context, constraints, and format.

Add a short usage note to every saved prompt: when to use it, what inputs it needs, which model performed best, and how to verify the output. For example, a research prompt might require source excerpts and a claim-evidence table. A writing prompt might require audience, voice sample, and must-keep facts. A coding review prompt might require the diff, expected behavior, and test context.

Whizi fits this workflow because you can save templates and run them across models instead of committing to one model forever. Start with three templates this week: a writing rewrite, a research brief, and a decision scorecard. Run each template on a real task, compare model outputs, revise the template, and keep the version that produces the most useful result with the least cleanup.

For a broader beginner guide, read [How to use AI](https://whizi.io/resources/how-to-use-ai). To turn your best prompts into standing routines, the [AI workflow templates](https://whizi.io/resources/ai-workflow-templates) guide gives you seven ready-made ones. For research-heavy work, try the [Founder Research Stack](https://whizi.io/templates/founder-research-stack). When you are ready to turn prompting into a repeatable workspace, create your account at [register](https://whizi.io/register).

Key points:

- A reusable instruction, context, constraints, and format framework for everyday work prompts
- Copy-paste prompt templates for writing, research, planning, analysis, and review tasks
- An iteration checklist that helps you improve accuracy, structure, and usefulness across models

Checklist:

- Start every serious prompt with a specific instruction and deliverable.
- Add audience, goal, source material, examples, and decision criteria as context.
- Use constraints to control length, tone, evidence, privacy, and uncertainty.
- Request a concrete format: table, checklist, memo, outline, or JSON-style fields.
- Ask the model to list assumptions, missing information, and confidence limits.
- Improve prompts through one focused revision at a time.
- Compare important prompts across models before saving the final template.
- Store reusable prompts by workflow, not by model name.

FAQ:

Q: What is prompt engineering for beginners?
A: Prompt engineering for beginners is the practice of giving AI models clear instructions, relevant context, useful constraints, and a specific output format so the response is easier to trust, edit, and reuse.

Q: How do I write better prompts?
A: Write better prompts by naming the task, giving the model background information, setting boundaries, requesting a clear format, and asking it to state assumptions or missing information before finalizing.

Q: Do prompt templates work across different AI models?
A: Yes, reusable prompt templates usually work across models, but outputs can differ. For important work, run the same prompt across models and compare accuracy, structure, tone, and edit time.

### How to summarize PDFs with AI (and verify accuracy)

URL: https://whizi.io/resources/summarize-pdf-with-ai/
Updated: August 2026

Quick answer: Summarize a PDF with AI in stages rather than one request. Ask for a document map first, covering sections, page ranges, tables, and figures. Then extract claims, numbers, and dates into a table with a source location for every row. Only then ask for a summary aimed at a specific reader, and verify it.

Learn how to summarize PDFs with AI using structured prompts, document maps, extraction steps, and an accuracy checklist.

#### Common failure modes

The fastest way to summarize PDF with AI is to upload a document and ask, "Summarize this." The fastest way to get a risky answer is also to upload a document and ask, "Summarize this." PDF work needs more structure because PDFs are not always clean text. They may include scanned pages, charts, footnotes, tables, appendices, rotated pages, tiny text, legal language, or research methods that matter more than the conclusion.

A good ai pdf summarizer workflow starts by separating reading from reasoning. First, ask the model to map the document. Then ask it to extract specific evidence. Only then ask it to summarize. This prevents the model from rushing into a polished answer before it has understood what is actually in the file.

Four things go wrong, and they go wrong in roughly this order of frequency:

| Failure | What it looks like | What to do about it |
| --- | --- | --- |
| Coverage loss | The summary covers the introduction and conclusion and skips tables, figures, caveats, and appendix data | Ask for a document map first, and check it covers the whole file |
| Invented specificity | A number, page reference, quote, or recommendation that sounds right but is not in the PDF | Require a source location for every precise claim, and verify the ones that matter |
| Visual blindness | Charts, diagrams, layout, and handwritten notes are silently missed | Ask the model directly what it can and cannot see before trusting anything |
| Context overflow | A long document produces a vague summary that gets weaker toward the middle | Split by section, and see [what is a context window](https://whizi.io/resources/what-is-a-context-window) |

Two of those deserve extra attention. **Invented specificity** is the dangerous one, because a fabricated page reference looks exactly like a real one. Treat every precise claim as unverified until you have looked, particularly when the document affects money, law, health, hiring, security, or a commitment to a customer.

**Visual blindness** is the one people do not know to check for. Model support for PDFs varies: some read text and visual content together, some extract text only, and all of them have limits on request size, page count, and legibility. Asking "what parts of this document can you actually see?" takes five seconds and occasionally changes everything that follows.

#### Workflow

Use this five-step workflow whenever you need to summarize long document AI outputs accurately. It works for research papers, board decks, product specs, vendor contracts, policy PDFs, market reports, and customer interview packets.

**Step 1. Prepare the PDF.** Check that it is searchable, upright, complete, and not password protected. Scanned or visually dense files are less reliable, so expect to verify more. Very long documents should be split along natural seams: chapters, exhibits, appendices, page ranges. Keep the original open so you can check page references as you go.

**Step 2. Get a document map.** Before asking for anything else, ask the model to list the major sections, page ranges, tables, figures, appendices, and repeated terms. You are not after insight yet, you are after orientation. If the map misses a big chunk of the file, stop and fix that before going further.

**Step 3. Extract the evidence.** Ask for claims, numbers, definitions, dates, risks, recommendations, and named entities in a table, with a source location for every row. Anything the model cannot locate gets marked "needs verification". This turns the PDF from a blob into something you can inspect.

**Step 4. Summarise for a specific reader.** A CEO summary, an analyst summary, a legal review, and an engineering handoff are four different deliverables. Say who it is for, what decision it supports, what to include, and what to leave out.

**Step 5. Verify and revise.** Run a second pass looking for missing caveats, unsupported claims, and contradictions, ideally with a different model. Then check the highest-value claims yourself, in the actual document.

The shape to remember: map, then evidence table, then audience-specific summary, then gaps, then your own verification. Skipping straight to the summary saves two minutes and regularly loses the facts that mattered.

#### Prompts for summary, outline, Q&A, and extraction

Use these ai document summarization prompts as reusable building blocks. Paste the prompt after uploading or attaching the PDF. If your tool supports multiple models, run the same prompt in two models and compare source accuracy before choosing the final answer.

Document map prompt: "Review this PDF and create a document map before summarizing. Include title, author or organization if visible, publication date if visible, major sections, page ranges, tables, figures, appendices, and any sections that appear hard to read. Do not summarize yet. If something is not visible, write Not visible."

Executive summary prompt: "Summarize this PDF for [audience] who needs to decide [decision]. Use this structure: 1. one-sentence thesis, 2. five key takeaways, 3. important numbers or evidence with page or section references, 4. risks and caveats, 5. recommended next questions. Do not include claims that are not supported by the document."

Research paper prompt: "Summarize this research paper with separate sections for research question, method, sample or dataset, main findings, limitations, practical implications, and what a skeptical reader should verify. Include page or section references where possible. Keep the language clear for a smart non-specialist."

AI PDF to outline prompt: "Turn this PDF into a detailed outline. Preserve the document hierarchy where possible. For each section, include the main point, supporting evidence, tables or figures mentioned, and unresolved questions. Mark anything that appears to be an inference rather than explicit text."

Chat with PDF AI prompt: "Answer my question using only this PDF: [question]. First quote or paraphrase the relevant evidence with page or section references. Then answer directly. Then list anything the PDF does not answer. If the answer requires outside knowledge, say so instead of guessing."

Extraction prompt: "Extract the following fields into a table: claim, exact number if any, unit, date or time period, entity, source page or section, confidence, and verification note. Use Not found when a field is missing. Do not calculate or infer values unless I explicitly ask you to."

Contradiction check prompt: "Review your previous summary against the PDF. Identify any unsupported claims, missing caveats, contradictions, overgeneralizations, and places where a table or figure changes the interpretation. Return a corrected summary and a list of edits made."

Long PDF chunk prompt: "I am sending this PDF in sections. For this section only, extract key claims, numbers, definitions, risks, and open questions. Do not create a final summary yet. Save a running glossary of terms that should stay consistent across sections."

#### Verification checklist

The best ai for pdf summaries is not just the model that writes the cleanest paragraph. It is the workflow that makes the answer checkable. Use this verification checklist before relying on a PDF summary.

Check coverage. Did the model mention the introduction, main body, tables, charts, appendix, limitations, and conclusion? If the PDF has exhibits or figures, ask specifically whether they changed the summary.

Check source grounding. Every important claim should trace back to a page, section, table, figure, or quoted passage. If the model gives page references, spot-check them. If the tool cannot provide reliable page references, ask for nearby headings or exact phrases you can search in the PDF.

Check numbers. Verify percentages, dollar amounts, dates, sample sizes, confidence intervals, deadlines, pricing, and totals manually. A model can copy a number correctly, transpose it, round it incorrectly, or attach it to the wrong entity.

Check scope. A study about one market, population, geography, time period, or product category should not become a universal claim. Ask: "What does this PDF not prove?" Good summaries include boundaries.

Check visual content. If the PDF includes charts, diagrams, or scanned pages, ask the model to describe what it can visibly read. Anthropic notes that PDF processing can combine text extraction with page images, but dense pages and visual limitations still matter. Blurry inputs produce brittle outputs.

Check privacy and permissions. Do not upload sensitive PDFs unless your organization allows that tool and data handling path. Contracts, financial records, medical documents, customer exports, unpublished research, and employee records deserve extra caution.

Check the final format. A summary is only useful if it matches the job. For a meeting, you may need action items. For a research paper, you need method and limitations. For a contract, you need obligations and risk. For a market report, you need assumptions and evidence.

#### Try the workflow in Whizi

Whizi is useful for PDF work because you can test the same document prompt across models instead of guessing which assistant will handle the file best. One model may produce a cleaner summary. Another may catch more caveats. Another may be better at turning the PDF into an outline or extraction table. The right answer is often visible only after you compare outputs.

Start with one real PDF, not a demo file. Upload a report, research paper, product spec, or customer document you actually need to understand. If the file is a lecture handout or a set reading, [the student workspace](https://whizi.io/use-cases/students) covers the study side of the same workflow. Run the document map prompt first. If the map looks complete, run the extraction prompt. Then run the executive summary prompt. Finally, run the contradiction check prompt and compare which model found the most useful corrections.

For a fast test, score each output from 1 to 5 on coverage, source grounding, number accuracy, caveats, and usefulness. Keep the prompt and model pairing that wins. That becomes your repeatable PDF workflow.

When the summary is ready, turn it into the next artifact: a briefing memo, Q&A document, slide outline, action list, or research table. That is the real value of document chat in Whizi. You are not just shortening a PDF. You are turning dense source material into work you can verify and use.

Create your account at [Whizi registration](https://whizi.io/register) to test document chat on your own PDFs. When you are ready to make PDF review part of your regular workflow, compare plan options at [Whizi pricing](https://whizi.io/pricing).

Key points:

- A reliable PDF workflow that moves from document map to extraction to summary
- Copy-paste prompts for summaries, outlines, Q&A, and structured data extraction
- A verification checklist for catching hallucinated claims, missed tables, and weak citations

Checklist:

- Prepare the PDF before uploading: searchable, upright, complete, and split into sections if very long
- Ask for a document map before asking for a summary
- Extract claims, numbers, dates, tables, and risks into a structured table
- Require page, section, table, figure, or nearby-heading references for important claims
- Use a separate prompt for Q&A instead of assuming the summary answers every question
- Ask the model to list what the PDF does not prove
- Manually verify numbers, dates, quotes, legal terms, financial claims, and medical or safety details
- Compare outputs across models when the document matters
- Save the best prompt sequence as a reusable PDF workflow

FAQ:

Q: How should I handle a scanned PDF differently?
A: Expect to verify more, and find out what the model can actually see before you trust anything. Scanned or visually dense files are less reliable, so ask the model directly to describe the charts, diagrams, and scanned pages it can visibly read. Blurry inputs produce brittle outputs, so check any figure you take from a scanned page against the original.

Q: Why does the summary skip the tables and the appendix?
A: That is coverage loss, and it happens when you ask for a summary before the model has mapped the file. Ask it to list the major sections, page ranges, tables, figures, and appendices first. If the map misses a big chunk of the document, fix that before going further. Then ask specifically whether the exhibits or figures changed the summary.

Q: Can I trust the numbers and page references in an AI summary?
A: Not without checking them. Invented specificity is the common failure: a number, quote, or page reference that sounds right but is not in the PDF. Require a source location for every precise claim, then verify percentages, dollar amounts, dates, sample sizes, and deadlines yourself. A model can copy a number correctly, transpose it, round it incorrectly, or attach it to the wrong entity.

Q: What do I do when the PDF is too long to summarize in one go?
A: Split it along natural seams: chapters, exhibits, appendices, or page ranges. Models have limits on request size, page count, and legibility, and a long document tends to produce a vague summary that gets weaker toward the middle. Send one section at a time, pull out the claims and open questions from each, and keep a running glossary so terms stay consistent.

Q: Which PDFs should I avoid uploading?
A: Anything sensitive that your organization has not cleared for that tool and that data handling path. Contracts, financial records, medical documents, customer exports, unpublished research, and employee records deserve extra caution. Settle the permission question before you upload rather than after, and apply the same care to documents that affect money, law, health, hiring, or security.

### How to use AI to translate: better than word-by-word tools

URL: https://whizi.io/resources/use-ai-to-translate/
Updated: August 2026

Quick answer: Translate with an AI chat model by telling it who will read the text, how formal it should be, and which names or terms must stay unchanged, then asking for the translation plus notes on anything ambiguous. Check the result by back-translating it in a fresh conversation with a different model.

Learn how to translate with AI chat models: the prompt pattern for tone and context, document translation, accuracy checks, and when AI beats Google Translate.

#### Why chat models changed translation

Classic translation tools convert words and sentences. They are fast and often accurate, but they take no instructions. You cannot tell them the reader is your boss, the tone should stay playful, the legal terms must remain in English, or the greeting should match how people actually open emails in Japan. An AI chat model takes exactly those instructions.

That is the practical difference. With ChatGPT, Claude, or Gemini, translation becomes a conversation: you explain the situation, the model translates for that situation, and you can ask why it chose a phrase or request a more formal version. For anything where tone and context matter, such as messages to people, job applications, and customer emails, that flexibility beats word-by-word conversion.

Dedicated tools still have their place. For instantly translating a street sign, a menu, or a webpage where you just need the gist, a classic translator app is quicker. The comparison section below gives an honest breakdown of when each option wins.

#### The prompt pattern that gets natural translations

The difference between a stiff translation and a natural one is almost always the instructions. Use this pattern and adapt it:

"Translate the following text into [language]. Audience: [who will read it and your relationship]. Register: [formal, neutral, or informal, and any local convention to follow]. Keep unchanged: [names, product terms, technical words]. After the translation, list any phrases that were ambiguous or that you adapted culturally, and explain your choice in one line each. Text: [paste]."

Three refinements come up constantly. For languages with formality levels, name the form you want, such as usted or tú in Spanish, Sie or du in German, vous or tu in French, and polite or casual style in Japanese and Korean. For idioms, ask the model to adapt rather than translate literally, or the joke will arrive as a corpse. For anything customer-facing, add "make it read like it was written by a native speaker, not translated."

You can also reverse the direction: paste a message you received in another language and ask what it says, what tone it carries, and how a polite reply would normally look in that culture. Beginners underuse this. Understanding incoming messages accurately is half the value of AI translation.

#### Translating documents and PDFs

Modern AI chat apps accept file uploads, which turns document translation from a copy-paste marathon into a short workflow.

1. Upload the document and state what it is: a rental contract, a diploma, a product manual. Context changes vocabulary.
2. Ask for the translation in the same structure as the original: keep headings, numbering, and tables so you can compare side by side.
3. For long documents, work in sections and ask for a glossary first: "List the recurring key terms and how you will translate each, then wait for my approval." Consistent terms are what separate professional from amateur translations.
4. Finish with: "List anything in this document a non-native reader might misunderstand, and any place where you were unsure."

Two limits to respect. First, a chat translation of an official document is not a certified translation; visa offices and courts usually require a sworn translator, so use AI to understand documents and prepare drafts, not to submit them. Second, for very long files, models can lose consistency in the middle sections, which is exactly what the glossary step and a section-by-section pass prevent. Our [PDF summarization guide](https://whizi.io/resources/summarize-pdf-with-ai) covers the upload workflow in more depth.

#### AI chat vs Google Translate and DeepL: an honest comparison

Each option earns its place. The table reflects what beginners actually experience, not vendor claims.

| Situation | Best option | Why |
| --- | --- | --- |
| Street signs, menus, quick gist | Classic translator app | Instant, camera input, works offline |
| Personal and business messages | AI chat model | Follows tone, audience, and formality instructions |
| Documents with fixed terminology | AI chat model | Glossary control and consistent structure |
| Long formal texts between major European languages | DeepL or AI chat | Both strong; compare on your text |
| Rare language pairs | Test both | Quality varies; run the back-translation check |
| Certified official translations | Human sworn translator | Legal requirement in most countries |

On accuracy: for major languages, modern chat models translate at a level most readers cannot distinguish from a competent human for everyday text. Quality drops for rare language pairs, heavy slang, and specialized jargon, which is precisely where the verification workflow in the next section matters. And because models differ by language pair, the model that is excellent for Spanish may be mediocre for Thai. Five minutes of side-by-side comparison on your actual language pair, which is exactly what [Whizi's side-by-side view](https://whizi.io/docs/compare-models-side-by-side/) is built for, answers the question for good.

#### How to check a translation you cannot read

The uncomfortable part of translating into a language you do not speak is that you cannot judge the result. You can still verify it with three checks that take a few minutes.

Back-translation is the workhorse. Open a new conversation, ideally with a different model, paste the translation, and ask for a translation back into your language plus a description of the tone. If the meaning survived the round trip and the tone matches what you intended, serious errors are unlikely. If something important shifted, fix the source sentence and rerun; ambiguous source text is the usual culprit.

Ask for a self-review: "Review this translation against the original. Flag mistranslations, awkward phrasing, and tone mismatches, and rate your confidence per paragraph." Models are decent critics of translation work, especially a second model reviewing the first one's output.

For high-stakes text, add a human check on the target side. Even one native-speaking acquaintance reading the final message catches the cultural misfire no automated check will. Use AI to get 95 percent of the way there and to make the human review fast.

#### Which model should you translate with?

There is no single best model for translation, and anyone who names one without asking your language pair is guessing. Performance genuinely differs across pairs and text types: one model produces the most natural casual German, another is more careful with Japanese honorifics, another holds terminology steady across a long contract.

The practical answer is a ten-minute test on your own languages. Take three texts you actually deal with: a casual message, a formal email, and a dense paragraph from a document. Run the same translation prompt on two or three models, back-translate each result with a different model, and note which one needed the fewest corrections and sounded most natural to any native speaker you can ask.

Running that test across ChatGPT, Claude, and Gemini normally requires three subscriptions, which nobody buys for an experiment. Whizi gives you all of them in one workspace under [one plan](https://whizi.io/pricing/), with side-by-side comparison built in, so the test costs ten minutes instead of sixty dollars. After the test, keep translating with the winner and spot-check important texts with the runner-up.

Key points:

- AI chat models translate meaning, tone, and context, not just words, and you can give them instructions a classic translator tool cannot take
- The quality gap between models differs by language pair, so comparing two models on your language is worth five minutes
- A simple back-translation check catches most serious errors even when you cannot read the target language

Checklist:

- State the audience, relationship, and formality level in every translation prompt.
- List names and terms that must stay unchanged before the model translates.
- Ask for notes on ambiguous or culturally adapted phrases.
- Keep document structure identical so you can compare side by side.
- Build a glossary first for long or technical documents.
- Back-translate important text with a second model in a fresh conversation.
- Use sworn human translators for documents with legal force.
- Test your specific language pair across models before picking a default.

Steps:

1. Tell the model the audience and the relationship: Say who will read the translation and how formal it should be, for example a client, a friend, or an official office.
2. Paste the text with any terms that must stay fixed: Provide the source text and list names, product terms, or phrases that should not be translated or must be translated a specific way.
3. Ask for the translation plus notes: Request the translation and a short list of any phrases that were ambiguous or culturally adapted, so you know where to look closely.
4. Run a back-translation as a check: In a new conversation, ask a model to translate the result back into your language and compare it against your original meaning.
5. Compare a second model for important text: For contracts, applications, or anything public, run the same prompt on a second model and reconcile any differences before using it.

FAQ:

Q: Can ChatGPT translate as well as Google Translate?
A: For everyday text in major languages, chat models like ChatGPT, Claude, and Gemini usually translate as well or better, because they follow instructions about tone, audience, and terminology. Classic tools remain faster for instant gist translations of signs, menus, and webpages.

Q: Is AI translation accurate?
A: For common language pairs and everyday content, accuracy is high. Quality drops for rare pairs, heavy slang, and specialist jargon. The back-translation check catches most serious errors: translate the result back to your language with a second model and compare meaning and tone.

Q: Can AI translate a whole PDF or document?
A: Yes. Upload the file, state what kind of document it is, and ask for the translation with the original structure kept. For long files, request a glossary of key terms first and translate section by section for consistency.

Q: Can I use an AI translation for official paperwork?
A: Usually not. Courts, visa offices, and universities typically require certified translations by sworn translators. Use AI to understand official documents and prepare drafts, then hire a certified translator for the submitted version.

Q: Which AI model is best for translation?
A: It depends on the language pair and text type. The reliable way to choose is a short test: run the same three texts through two or three models, back-translate the results, and keep the model that needed the fewest corrections. Whizi lets you run that test in one app.

### What is a context window? The AI limit that explains everything

URL: https://whizi.io/resources/what-is-a-context-window/
Updated: August 2026

Quick answer: A context window is the total amount of text a model can hold in view at once: your instructions, every message in the conversation, any uploaded files, and the reply being written. When the total exceeds it, the oldest parts drop out. Quality also degrades well before the advertised limit.

A plain-English explanation of context windows and tokens, why AI forgets things mid-conversation, why bigger windows are not always better, and how to work within them.

#### The plain explanation

A context window is the total amount of text a model can hold in view at one time. Everything the model knows about your current conversation has to fit inside it: your instructions, every message either of you has sent, every file you uploaded, and the answer it is about to write.

A useful mental picture: the model has no memory of your conversation at all. Each time you send a message, the entire conversation so far is handed to it fresh, it reads all of it, and it writes the next reply. The context window is the size of the desk that transcript has to fit on. When the desk is full, something has to come off it.

That is why an AI can seem sharp for twenty messages and then start contradicting itself, forgetting a constraint you set at the beginning, or asking for a file you already gave it. It did not get confused. The early part of the conversation slid off the desk.

One distinction to get right immediately, because it causes a lot of confusion: **the context window is not the same as memory.** Products like ChatGPT and Claude have a separate memory feature that stores facts about you between conversations and quietly inserts them into new ones. That is a product feature built on top of the model. The context window is a hard property of the model itself, and no memory feature makes it bigger.

#### Tokens, and how to estimate them

Context windows are measured in tokens rather than words, because models do not read words. Token is one of the words the [AI glossary](https://whizi.io/resources/ai-glossary) defines in a sentence, along with the fifty-odd others that surround it. They read chunks: common words are usually one token, longer or unusual words split into several, and punctuation and spaces count too.

The rough conversions worth remembering:

- **1 token is about 4 characters of English**, or about three quarters of a word.
- **1,000 tokens is roughly 750 words**, about a page and a half of normal prose.
- **Code is denser.** Expect closer to one token per three characters because of symbols, indentation, and unusual identifiers.
- **Other languages are less efficient.** Text in languages that are less represented in the tokenizer can take two or three times as many tokens for the same meaning, which is worth knowing if you are paying per token.

That gives you a way to picture the numbers you see in marketing material:

| Context window | Roughly equal to |
| --- | --- |
| 8,000 tokens | A long article |
| 32,000 tokens | A short research paper with notes |
| 128,000 tokens | A 300 page book |
| 200,000 tokens | A dense technical manual, or a mid-size codebase |
| 1,000,000 tokens | Several books, or a year of meeting transcripts |

The important thing this table hides: the limit covers the **whole** conversation, not each message. A 128,000 token window does not mean you can send 128,000 tokens repeatedly. It means the running total of everything, including the model's own replies, has to stay under that number.

#### What actually fills up your window

People are usually surprised by how fast the window fills, because most of what goes into it is invisible. In a typical chat session, the model is reading all of the following on every single turn:

1. **The system prompt.** Instructions the product sends before you type anything: how to behave, what tools exist, today's date, safety rules. Often thousands of tokens, and you never see it.
2. **Your custom instructions or memory.** Anything the product has stored about you and injects automatically.
3. **Every previous message**, yours and the model's, in full. The model's long answers count too, and they are usually the biggest contributors.
4. **Every uploaded file**, or the extracted portions of it. A 40-page PDF is roughly 20,000 to 30,000 tokens.
5. **Tool and search results.** A web search that pulls five pages can add more than your entire conversation so far.
6. **The reply being generated.** Output shares the same budget as input.

This is why a conversation that felt short can be near its limit. You sent eight short messages, but the model wrote eight long answers, you attached two documents, and it ran three searches. The visible part of that is maybe ten percent of the total.

It is also why the fix for a confused conversation is so often "start a new chat". You are not resetting the model's mood. You are clearing the desk.

#### Advertised size versus useful size

This is the part that matters most and gets discussed least. **A model's quality does not stay flat right up to its limit and then fall off a cliff. It degrades gradually, and it starts degrading well before the limit.**

The best documented version of this is often called the lost in the middle effect. Put a specific fact at the start of a long document and the model finds it. Put it at the end and the model finds it. Bury it in the middle of 200 pages and retrieval accuracy drops noticeably. Attention is not evenly distributed across a long input, and the edges get more of it.

It gets harder still when the task requires combining several facts scattered across a long input. Finding one needle in a haystack is a solved problem. Finding four needles and reasoning about the relationship between them is not, and this is exactly the task people use big context windows for.

So treat the advertised number as a maximum capacity, not a comfortable working range. A practical rule of thumb from real use: **you get reliable behaviour up to roughly half the advertised window, and you should verify anything beyond that rather than trust it.** If a model tells you a 300 page contract contains no termination clause, check, particularly if the clause would be in the middle.

The comparison implication is that a model with a one million token window is not automatically better at long documents than one with 200,000. It is better at accepting them. Whether it reasons well across all of it is a separate question, and the only way to know is to test with a document where you already know the answer.

#### What happens when you run out

Different products handle overflow differently, and knowing which one you are using explains a lot of odd behaviour.

| Behaviour | What you see | Where it happens |
| --- | --- | --- |
| Hard error | The request is rejected with a message about length | Most direct API use |
| Silent truncation | Oldest messages are dropped without telling you | Many chat interfaces |
| Rolling summary | Old messages are compressed into a summary | Increasingly common in chat products |
| Retrieval | Only the relevant parts of your documents are fetched per turn | Document and knowledge-base tools |

Silent truncation is the one that causes real problems, because nothing announces it. The symptom is a model that suddenly ignores a rule you set at the start, reverts to a tone you corrected an hour ago, or asks a question you already answered. Nothing has gone wrong with the model. Those instructions are simply no longer on the desk.

Rolling summaries are better but lossy. A summary keeps the gist and drops the specifics, so the constraint "never use the word synergy" survives as "the user has style preferences", which does not help you at all.

#### Practical tactics that actually work

These are ordered by how much difference they make relative to effort.

**Start a new chat when the topic changes.** The single highest-value habit. A long conversation carries the cost of everything before it, including tangents that are no longer relevant. Long conversations are not more informed, they are more expensive and more diluted.

**Put your question after the long material, not before.** If you paste a document and then ask, the question is close to where the answer gets generated, which measurably improves accuracy on long inputs. Paste first, ask second.

**Re-anchor important constraints.** In any conversation over about fifteen exchanges, restate the rules that matter in the message where they matter: "as a reminder, British English, no bullet points, under 400 words". It costs you one line and it survives truncation.

**Send the relevant pages, not the whole book.** If you need the model to check the indemnity clause, give it the indemnity clause and its neighbours. Precision beats volume, it is faster, it is cheaper, and it is more accurate.

**Summarise and restart deliberately.** When a working session gets long, ask for a structured handoff: current state, decisions made, open questions, constraints. Paste that into a new chat. You keep the substance and drop the noise, and you will notice the model gets sharper immediately.

**Use retrieval for anything that does not fit.** If your material is genuinely larger than any window, the answer is not a bigger window, it is fetching the relevant chunks per question. That is what document tools do under the hood.

**Watch for degradation, not just errors.** If answers get vaguer, hedge more, or start ignoring format instructions, you are probably deep in the window. Start fresh before assuming the model has got worse.

#### How much window do you actually need?

Match the window to the job rather than shopping for the biggest number. For the current figure on any specific model, the [context window comparison](https://whizi.io/resources/context-window-comparison) table lists the whole catalogue in tokens and pages.

| Your work | What you need | Why |
| --- | --- | --- |
| Emails, drafting, quick questions | Anything modern | You will never come close to the limit |
| Editing long documents | 100,000 plus | The document plus your conversation about it |
| Contract and policy review | 200,000 plus, and verify | The document plus the reasoning about it, with the caveat above |
| Reasoning across a whole codebase | The largest available | Code is token-dense and cross-file reasoning needs breadth |
| Analysing months of transcripts | The largest available, or retrieval | Often better served by retrieval than by brute force |
| Building a product on an API | Smaller than you think, plus caching | Long prompts are the main driver of both cost and latency |

For most people the honest answer is that context window is not the deciding factor between subscriptions. Writing quality, ecosystem, and price matter more. It becomes the deciding factor in exactly one situation: your work regularly involves feeding in more material than a normal window holds, and in that case the difference is not marginal, it is the difference between possible and impossible.

If you want to test that yourself, the practical approach is to run the same long document through two or three models and check the answers against something you already know. A workspace like [Whizi](https://whizi.io/pricing) makes that easy because GPT, Claude, Gemini, Grok, and DeepSeek sit behind one subscription, and Gemini's very large window is one click from Claude's careful synthesis. See [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side), or read [how to choose an AI model](https://whizi.io/resources/how-to-choose-ai-model) for the wider decision.

Key points:

- A context window is working memory for one conversation, not long-term memory across conversations
- Quality starts degrading well before you reach the advertised limit
- Most "the AI forgot what I said" problems are context problems with simple fixes

Checklist:

- Estimate token counts with the rule that 1,000 tokens is about 750 words
- Remember that system prompts, files, search results, and replies all consume the same budget
- Treat roughly half the advertised window as the reliable working range
- Paste long material first and ask your question afterwards
- Restate critical constraints in long conversations so they survive truncation
- Start a fresh chat with a written handoff instead of extending a long one

FAQ:

Q: What is a context window in simple terms?
A: It is the total amount of text a model can hold in view at once, covering your instructions, the entire conversation so far, any uploaded files, and the reply being written. When the total exceeds the limit, the oldest parts drop out, which is why long chats start forgetting things.

Q: Is a bigger context window always better?
A: No. A bigger window lets a model accept more input, but accuracy degrades gradually as the input grows, especially for facts buried in the middle. A model with a one million token window is not automatically better at long documents than one with 200,000, so test with material where you know the right answer.

Q: How many words is 128,000 tokens?
A: Roughly 96,000 words, or about a 300 page book. The general conversion is that one token is about four characters of English, so 1,000 tokens is around 750 words. Code and non-English text use more tokens for the same content.

Q: Why does ChatGPT forget what I said earlier?
A: Because the conversation grew past the context window and the oldest messages were dropped or compressed to make room. Most chat interfaces do this silently. Restating your key constraints, or starting a new chat with a short summary, fixes it immediately.

Q: Is the context window the same as AI memory?
A: No. The context window is a hard limit on one conversation. Memory is a separate product feature that saves facts about you between conversations and inserts them into new ones. Memory does not increase the window, it uses part of it.

### What is AI? A plain explanation for people who have never used it

URL: https://whizi.io/resources/what-is-ai/
Updated: August 2026

Quick answer: AI is software that learned the patterns of human writing from an enormous amount of text, and uses them to produce new writing that fits your request. You ask in ordinary language and it writes back a draft, a summary, an explanation. It is not looking facts up, which is why it is fluent and sometimes confidently wrong.

AI explained in simple words: what it actually is, how it learned to do this, what it is genuinely good at, what it gets wrong, and exactly what to type first.

#### The short answer, and what people actually mean

AI is software that has read an enormous amount of human writing and learned the patterns in it well enough to produce new writing that fits. You type a request in ordinary language. It writes back: a draft, a summary, an explanation, a plan. That is the whole product. Everything else you have heard is either a detail sitting on top of that, or a story about something else entirely.

The term technically covers older things too, such as the spam filter in your inbox, the face unlock on your phone, and the system deciding which video plays next. Nobody means those in conversation. When a colleague says she used AI to fix a paragraph, or your nephew says he asked AI about his knee, they mean a chatbot: ChatGPT, Claude, Gemini, or one of their relatives. The product-level view of that category lives in [what an AI chatbot is](https://whizi.io/resources/what-is-an-ai-chatbot). This article sits one level above it, on the idea itself. If the vocabulary is the part that puts you off, model and prompt and token and context, [the plain-language glossary](https://whizi.io/resources/ai-terms-explained) is short and unpretentious.

Signing up for one takes an email address and about a minute, and every major app has a free tier, so nothing here needs a card. Before you read further, open any AI app and type this: "Explain what you are and how you work, in four sentences, to someone who has never used AI before." Leave the answer open in another tab. By section four you will have something concrete to test it against, which beats taking anyone's word for it.

#### How it learned to do that

The training process is dull to describe, and that is a good sign. Take a very large pile of text: books, websites, manuals, forum arguments, court filings, recipes. Hide a word. Make the system guess it. Tell it whether the guess was right. Repeat an absurd number of times. Nothing more exotic is happening. What comes out is a system with an extremely fine sense of what usually comes next in human writing.

So when you ask it who won the 1998 World Cup, it is not opening a reference book. It is producing the text that most plausibly follows your question, one small piece at a time. It happens to be right about France, because that fact appears in its training text thousands of times in a consistent form. It is right for the same reason it is fluent, not for some other, sturdier reason. Hold onto that sentence. Almost everything strange about how AI behaves falls out of it.

- **It writes well immediately.** Fluency is exactly what it was trained on, so a first draft arrives in decent shape without any effort from you.
- **It sounds identical when it is wrong.** Nothing inside it separates a fact seen ten thousand times from one it is assembling out of plausible-looking parts. The confidence is the same either way.
- **It has a cutoff.** Training stopped on some date. Ask about last Tuesday and, unless the app runs a live web search for you, you are asking a system that was not there.
- **It is shaky at arithmetic.** Long multiplication is a procedure, not a pattern in prose. Models handle it better now by writing out steps or calling a calculator, but the underlying instinct is still to produce a number that looks correct.
- **Your phrasing steers the answer.** Your words decide which patterns it reaches for, which is why the same question asked two ways comes back at two different levels of quality. [Talking to AI](https://whizi.io/resources/how-to-talk-to-ai) is mostly that single skill.

#### What it is genuinely good at

The tasks it does best share a shape. The form of a good answer is already well established in the world, and you can tell within about ten seconds whether what came back is any good. Rewriting, summarizing, translating, explaining a document, and getting past a blank page: that is most of the writing in an ordinary week.

| What you already have | What to ask for | Why this one works |
| --- | --- | --- |
| An email you have rewritten four times | "Rewrite this so it is warm but still clearly says no" | Polite refusals are one of the most common patterns in written English |
| A letter from an insurer or a landlord | "Explain this in plain language, then list what I have to do" | Turning jargon into ordinary words is pattern work, and you can check it against the letter in front of you |
| A 60-page PDF you were never going to read | "Summarize the argument in eight bullets, then pull out every number" | It reads the whole file in seconds, and you can spot-check the bullets you care about |
| A trip you keep not planning | "Plan four days in Seville for two adults who like food and walking" | Thousands of itineraries sit in the training text, and an imperfect one costs you nothing |
| A message to a client in another language | "Translate this into Spanish, polite but not stiff" | Tone survives here in a way it did not with older translation tools |

Look at what is missing from that table: not one row asks the machine to be the authority. You already hold the material, the judgment, or both. It takes the part you find tedious, the blank page, the second draft, the reformatting, the sixty pages, and leaves the deciding to you. [Everyday tasks worth handing over](https://whizi.io/resources/ai-for-everyday-tasks) has a longer list. One habit to build early: when the first answer misses, say what is wrong in one short sentence instead of starting over. "Too formal, and cut the last paragraph" fixes more than a rewritten request usually does.

#### What it is bad at, for exactly the same reason

Now the honest part, because you will meet this in your first week anyway. It will tell you something false in the same calm, well-organized voice it uses for everything else. Not rarely, and not only about obscure topics. It will invent a section number in a lease, a study with believable authors and a believable year, a product feature that does not exist, a quotation nobody ever said. There is no tell: no hedge in the wording, no wobble in the tone. The word for this is hallucination, and [why AI gets things wrong](https://whizi.io/resources/why-ai-gets-things-wrong) walks through the failure modes one at a time, because recognizing which kind of wrong you are looking at is most of the defense.

Three further limits, all growing from the same root. Recency: unless the app is searching the web live, it does not know the past few months, and it will discuss them anyway using older material. Counting and arithmetic: it can reason its way through a calculation and still hand you a total that is off by a digit, so verify any number you would act on. Anything specific to you: your contract, your dose, your local rules, your deadline. It has read a million documents that resemble yours and not one of yours.

The rule worth keeping is this. Do not use it where being 95 percent right is worthless. Medication, legal deadlines, tax figures, the amount on an invoice, the spelling of a name in a speech you are giving. In those cases nearly right is not partial credit, it is simply wrong, and the fluent presentation makes the error harder to catch than a bad Google result would be. Everywhere else, 95 percent is a strong first draft that you then spend four minutes correcting. If your hesitation is less about accuracy and more about what happens to the things you type, that is a fair and separate question, and [is AI safe to use](https://whizi.io/resources/is-ai-safe-to-use) treats it as one.

#### Your first five minutes

Reading about this has a low ceiling. Open an app and type one of the four below, adapted to something real from your own week. Real matters: a made-up test question gets you a made-up-feeling answer, and teaches you nothing about whether this is useful to you.

1. **The awkward email.** "Here is an email I need to send. Rewrite it so it is polite, about half as long, and still clearly says no." Paste your draft underneath.
2. **The letter you dread.** "Explain this letter in plain language, then tell me exactly what I have to do and by when." Paste or upload the thing from the insurer, the landlord, or the tax office.
3. **The plan you keep postponing.** "I have four days in Lisbon in October with one teenager and roughly 700 dollars for food and activities. Plan it day by day, then tell me what you assumed about us."
4. **The question you feel silly asking.** "Explain how a mortgage rate is actually set, as if I am intelligent but have never bought a house. Then give me the three questions to ask a lender."

A good answer does the task instead of describing how it would do the task. It holds the format you asked for. Where you left something out, it either asks you or names the assumption it made. If what comes back is a padded, over-headed, essay-shaped thing, that is usually a symptom of a vague request rather than a weak model, so add who it is for, how long you want it, and what you plan to do with it. Models also differ more than the marketing suggests, in voice and in how carefully they follow a long instruction, so the fastest way to find your fit is to send one prompt to several of them and compare. That is what Whizi is for: one subscription that includes GPT, Claude, Gemini, and image models in one workspace, instead of paying each provider separately. [Start a trial](https://whizi.io/register), run your own four prompts, and keep whichever answer you would have been willing to send.

Key points:

- When someone says AI in 2026 they almost always mean a chatbot you type into, not a robot and not a private database of answers
- These systems work by predicting what text comes next, which explains both why they write so well and why they state wrong things confidently
- Ten minutes of typing real requests teaches you more about AI than an hour of reading about it

Checklist:

- Treat AI as a text predictor, not a lookup tool with verified answers.
- Bring a real task from your own week instead of testing it with trivia.
- Say who the answer is for, how long it should be, and what format you want.
- Check every name, number, date, and quotation before you rely on it.
- Ask for a fix in one sentence instead of starting the conversation again.
- Skip AI entirely for anything where being almost right is worthless.
- Run the same request past two different models before deciding AI is not for you.

FAQ:

Q: What is AI in simple words?
A: AI is software that learned the patterns of human writing from an enormous amount of text and uses them to produce new text that fits your request. You ask in ordinary language and it writes back. It is not consulting a database of verified facts, which is why it is so fluent and also why it is sometimes confidently wrong.

Q: What is the difference between AI and ChatGPT?
A: AI is the category and ChatGPT is one product inside it, built by OpenAI. Claude from Anthropic and Gemini from Google are equally capable alternatives, and there are strong open models behind many other apps. People say ChatGPT the way they say Google for search, but the options genuinely differ in writing style and in how carefully they follow instructions.

Q: How does AI actually work, without the jargon?
A: It was trained by guessing hidden words in a very large pile of text, over and over, until it developed a precise sense of what usually comes next. When you send a message it builds a reply one small piece at a time, choosing text that plausibly continues the conversation. That single mechanism explains its fluency, its confidence, and its mistakes.

Q: Does AI understand what it is saying?
A: Not the way you do. It holds no beliefs and has no awareness of being right or wrong, which is exactly why it never warns you that it is guessing. A useful working picture is an extremely well-read assistant who cannot remember where anything was read, so you keep the judgment and check what matters.

Q: Do I have to pay to try AI?
A: No. Every major app has a free tier that is enough to see what the thing does. Free tiers usually drop you to a smaller model at busy times and cap how many messages you can send, and a single provider's paid plan runs around 20 dollars a month, so check the current pricing page before you subscribe to two of them.

### What is an AI chatbot? A plain-English guide for beginners

URL: https://whizi.io/resources/what-is-an-ai-chatbot/
Updated: August 2026

Quick answer: An AI chatbot is software you talk to in ordinary written language, and it replies with an answer, a draft, a summary, a plan, or an explanation. ChatGPT, Claude, and Gemini are the best-known examples. They predict likely text rather than looking up verified facts, which is why a confident answer can still be wrong.

Understand what an AI chatbot is, how tools like ChatGPT, Claude, and Gemini actually work, what they can and cannot do, and how to try one safely today.

#### The plain-English definition

An AI chatbot is a program you talk to in normal language. You type a request the way you would text a knowledgeable friend, and it replies with an answer, a draft, a summary, a plan, or an explanation. There is no special command language to learn and no menus to memorize. The conversation itself is the interface.

The best-known AI chatbots are ChatGPT from OpenAI, Claude from Anthropic, and Gemini from Google. They belong to a family of technology called large language models. You will see that phrase shortened to LLM in articles, but you do not need the jargon to use them well.

What makes these tools different from the customer-service bots you may have used on airline or bank websites is flexibility. Old-style bots could only follow scripts: press 1 for billing, type "refund" for refunds. Modern AI chatbots handle open-ended requests they have never seen before, such as "rewrite this email so it sounds friendlier but still says no" or "explain this contract clause like I am fifteen."

If you have ever wanted a patient assistant who drafts, summarizes, explains, translates, and brainstorms on demand, that is the practical meaning of an AI chatbot. The rest of this guide explains how they work, what they get wrong, and how to start using one for real tasks in about ten minutes.

#### How AI chatbots actually work, without the math

Under the hood, an AI chatbot is a prediction system trained on enormous amounts of text: books, articles, websites, code, and conversations. During training, the system learned the patterns of how words, ideas, and arguments fit together. When you send a message, it generates a response by predicting, one small piece at a time, what text would best continue the conversation.

That sounds simple, but at very large scale it produces something remarkable: the model can follow instructions, adapt its tone, connect ideas across topics, and work through problems step by step. When you ask it to summarize a report in three bullet points for a busy manager, it is using patterns learned from millions of summaries, reports, and workplace documents.

Two practical consequences follow from this design. First, a chatbot does not look things up in a live database by default. Its knowledge comes from training data with a cutoff date, plus whatever text you paste into the conversation. Some apps add live web search on top, but the core model is working from learned patterns, not a fact table.

Second, because it predicts plausible text rather than retrieving verified facts, a chatbot can state something false with complete confidence. People call these mistakes hallucinations. They are the single most important thing to understand as a beginner, and the reason the verification habit in the last section of this guide matters so much.

#### What you can use an AI chatbot for on day one

You do not need prompt tricks to get value on the first day. You need a real task. These are the ones beginners get the most out of, with a starting phrase you can adapt for each.

- **Writing and rewriting.** "Rewrite this email so it is shorter and more polite, and keep the main request clear." Paste the draft below the instruction.
- **Summarizing.** "Summarize this article in five bullet points for someone who has never heard of the topic." Paste the text or upload the document.
- **Explaining.** "Explain how a fixed-rate mortgage works like I am completely new to this, then give me the three questions I should ask a lender."
- **Planning.** "Plan a four-day trip to Lisbon for two adults who like food and walking, with a rough daily budget in euros."
- **Deciding.** "Here are two job offers. Build a comparison table covering salary, commute, growth, and risk, then tell me what questions I still need to answer."
- **Translating.** "Translate this message into Spanish. Keep it polite but informal, the way a coworker would write it."

Notice the pattern in every example: a clear job, a bit of context, and the shape you want back. That is most of what people call prompt writing. When you are ready to go deeper, the [beginner prompt engineering guide](https://whizi.io/resources/prompt-engineering-beginners) turns this pattern into a reusable framework, and [how to use AI](https://whizi.io/resources/how-to-use-ai) walks through a full first week of tasks. The translation example has more to it than it looks, and [how to use AI to translate](https://whizi.io/resources/use-ai-to-translate) covers holding tone steady across languages.

#### What AI chatbots get wrong

AI chatbots are genuinely useful, and they fail in specific, predictable ways. Knowing the failure modes in advance is what separates people who get value from people who get burned.

| Failure mode | What it looks like | What to do |
| --- | --- | --- |
| Invented details | A citation, statistic, or product feature that does not exist, delivered in a confident tone | Treat every specific fact, name, number, and quote as unverified until you check it |
| Knowledge cutoff | Recent events, prices, laws, and product details are out of date | Ask what it knows after a given date and whether it searched the web for the answer |
| Mirroring your framing | You ask why X is better than Y, and it argues for X | Ask for the strongest case against your current view as a separate question |
| Privacy | Data handling varies by provider, plan, and settings | Keep passwords, financial details, and other private information out of the chat |

They invent details. When a chatbot does not know something, it often produces a confident-sounding answer anyway: a citation that does not exist, a statistic with no source, a feature a product never had. Treat every specific fact, name, number, and quote as unverified until you check it.

Their knowledge has a cutoff. Models are trained on data up to a certain date, so recent events, prices, laws, and product details can be outdated unless the app adds live search. Ask the chatbot directly what it knows about events after a given date and whether it searched the web for the answer.

They mirror your framing. If you ask "why is X better than Y," many models will argue X is better instead of challenging the premise. For decisions that matter, ask for the strongest case against your current view as a separate question.

Privacy deserves a moment of thought. Reputable providers publish policies on how conversation data is handled, and paid plans often include stronger controls. The practical beginner rule: do not paste passwords, full financial details, or other people's private information into any chatbot, and read the privacy settings of the app you choose once, at the start.

#### ChatGPT, Claude, and Gemini: what the names mean

Beginners often think ChatGPT is a synonym for AI chatbot, the way some people say Google for search. It is actually one product among several excellent ones. ChatGPT is built by OpenAI, Claude is built by Anthropic, and Gemini is built by Google. There are also strong open models such as Llama and DeepSeek that power many other apps. The most famous one has its own beginner guide: [what ChatGPT is and how to use it](https://whizi.io/resources/chatgpt).

The differences are real but not what marketing suggests. In day-to-day use, the models differ in writing style, how carefully they follow long instructions, how well they handle documents, coding strength, and how they explain their reasoning. One model may write warmer emails while another is stricter about sticking to a source document you gave it. Our [three-way comparison of ChatGPT, Claude, and Gemini](https://whizi.io/resources/chatgpt-vs-claude-vs-gemini) covers the practical differences task by task.

Here is the catch for a beginner. You cannot know which model fits your work until you have run the same task through more than one of them. But subscribing to each app separately costs more than most people want to spend on an experiment. That is the problem Whizi exists to solve. One subscription gives you the major models in one workspace, so you can ask the same question across models, compare the answers side by side, and settle on the one that fits each kind of task.

A useful mental model as you start: the chatbot app is the kitchen, the model is the chef. Same request, different chef, noticeably different meal. Being able to switch chefs without paying for three restaurants is what makes the learning phase cheap.

#### Try one safely in the next ten minutes

The fastest way to understand AI chatbots is not another article. It is ten minutes with a real task. Here is a safe first session you can run right now.

1. Open a chatbot app and start a new conversation.
2. Paste a real piece of text you are working with: an email you need to answer, a document you need to understand, or a plan you need to make.
3. Ask for one specific thing: "Summarize this," "Draft a reply that politely declines," or "Turn this into a checklist."
4. Ask two follow-ups: "Make it half as long" and "What did you assume that I did not tell you?" The second question teaches you more about how these tools think than any tutorial.
5. Verify anything factual before you use it, and rewrite at least one sentence in your own voice.

That last step matters beyond accuracy. The goal is not to let a chatbot replace your judgment. It is to hand off the blank-page work, the reformatting, and the first draft, while you keep the decisions. People who use AI well describe it as a tireless junior assistant: fast, broadly knowledgeable, occasionally wrong, and always in need of a final review.

When you are ready to compare models on your own tasks instead of taking anyone's word for it, you can [start a Whizi trial](https://whizi.io/register) and run the same prompts across ChatGPT, Claude, Gemini, and more in one place. If you want a structured first week, follow the task list in [how to use AI](https://whizi.io/resources/how-to-use-ai).

Key points:

- An AI chatbot is software that understands normal written language and responds with useful text, so you can ask for help the way you would ask a person
- ChatGPT, Claude, and Gemini are the best-known examples, and each one is better at some tasks than others
- You do not need any technical background: if you can describe what you need in a sentence, you can use an AI chatbot today

Checklist:

- Understand that a chatbot predicts text; it does not look up verified facts by default.
- Start with a real task from your day, not a trivia question.
- Give every request a clear job, some context, and a desired format.
- Ask follow-up questions instead of starting over when the first answer misses.
- Verify names, numbers, dates, and claims before you rely on them.
- Never paste passwords, financial details, or other people's private data.
- Try the same task on at least two models before deciding which to trust.
- Keep your own judgment in charge: the chatbot drafts, you decide.

Steps:

1. Pick one app that gives you access to the major models: Choose an app where you can try ChatGPT, Claude, and Gemini in one place, so you are not locked into the first model you test.
2. Start with a real task, not a test question: Paste in a real email, document, or decision you are working on and ask for specific help with it.
3. Give the chatbot context and a clear job: Say who the output is for, what you want it to do, and what format you want back, such as a list, a table, or a short draft.
4. Ask follow-up questions in the same conversation: Treat the first answer as a draft. Ask for shorter, simpler, more formal, or more detailed versions until it fits.
5. Verify anything important before you use it: Check names, numbers, dates, and claims against a source you trust before sending or publishing anything the chatbot wrote.

FAQ:

Q: Is an AI chatbot the same thing as ChatGPT?
A: No. ChatGPT is one AI chatbot made by OpenAI. Claude from Anthropic and Gemini from Google are equally capable alternatives, and each has different strengths. ChatGPT is simply the most famous example of the category.

Q: Are AI chatbots free to use?
A: Most have free tiers with limits on usage and access to the strongest models. Paid plans typically cost 20 dollars or more per month per provider. Multi-model apps like Whizi bundle several providers into one subscription, which is usually cheaper than paying two or three providers separately.

Q: Are AI chatbots safe to use?
A: For everyday tasks, yes, with two habits: verify factual claims before acting on them, and avoid sharing passwords, financial details, or other people's private information. Review the privacy settings of the app you choose when you first sign up.

Q: Do AI chatbots learn from my conversations?
A: It depends on the provider and your settings. Some providers may use conversations to improve models unless you opt out, and business or paid tiers often exclude your data by default. Check the data controls in the app settings rather than assuming.

Q: Which AI chatbot is best for a beginner?
A: The honest answer is that it depends on your tasks. Writing, research, coding, and document work each favor different models. The practical approach is to run your own tasks on ChatGPT, Claude, and Gemini side by side, then keep whichever wins for each job. Whizi makes that comparison possible under one subscription.

### What do I use when ChatGPT is down? Check in thirty seconds, then switch

URL: https://whizi.io/resources/what-to-use-when-chatgpt-is-down/
Updated: August 2026

Quick answer: Switch to Claude, Gemini, Perplexity, Mistral Vibe, Grok or DeepSeek, which run on separate infrastructure and are rarely down at the same time. First open status.openai.com to confirm it is a real outage rather than your own usage cap, and copy any unsent message out of the box before you refresh.

Tell a real ChatGPT outage apart from your own rate limit, save the work sitting in the tab, and switch to something that is still up without losing the thread.

#### The thirty second check

Open **status.openai.com** in a new tab. An active incident affecting ChatGPT is posted there, usually within minutes of the problem starting, and if you see one you can stop diagnosing. If the page says everything is operational, the problem is almost certainly on your end: your own usage cap, your browser, your network, or your account. Four problems, four different fixes, and guessing wrong costs you the twenty minutes you did not have. One thing before you touch refresh: copy whatever is sitting in the message box, because that is the only part of your work an outage can actually destroy, and the next section covers the rest of it.

People assume an outage far more often than one is happening, because a cap and a crash feel identical from where you sit: you type, nothing useful comes back. But if you are capped, no amount of refreshing helps, and if the service is genuinely down, clearing your cache is wasted effort.

| What you see | What it usually means | What actually fixes it |
| --- | --- | --- |
| Status page shows an active ChatGPT incident | A real outage | Nothing on your end. Switch tools and come back later |
| A message about reaching a limit, or being asked to wait or upgrade | Your own usage cap | Wait for the window to reset, switch to a model you have not exhausted, or use another provider |
| The page loads but every reply errors out, and the status page is green | Often browser or network | Try a private window with extensions off, then a different browser, then your phone on mobile data |
| The whole site fails to load, and the status page is green | Usually network or DNS | Test another site, restart the router, try mobile data |
| Logged out unexpectedly, or told your account is unavailable | An account problem | Sign in again, then check the email on file for anything from OpenAI |

One more check worth ten seconds: load ChatGPT on your phone over mobile data, with wifi off. If it works there and not on your laptop, the problem is local, whatever a crowdsourced outage tracker says. Those trackers count complaints, which spike whenever a popular service has any bad minute anywhere. Treat them as a hint and the official status page as the answer.

#### Before you refresh: do not lose the work in the tab

During an outage the instinct is to keep pressing refresh, and the refresh is what costs you the draft. Before you diagnose anything further, spend twenty seconds on this in order.

1. **Select everything in the message box and copy it.** A long prompt you have been shaping for ten minutes is the thing most likely to disappear. Paste it into a note, a document, anywhere outside the browser tab.
2. **Copy the last useful answer too.** If a reply is half rendered on screen, select what is visible and copy that as well. A partial answer is worth more than a lost one.
3. **Do not close the tab, and do not use the back button.** An unsent message usually survives in the box as long as the page stays open. It rarely survives a navigation.
4. **Do not send the same message repeatedly.** Retrying twenty times does not move you up a queue. It just gives you twenty more chances to hit an error, and on a capped account it can burn what allowance you have left.
5. **Check whether the answer actually arrived.** Some outages break the display, not the generation. Open the conversation in a second tab or on the phone app before you assume the reply is gone.

Then note where you stopped in one line of plain text: the goal, the constraints, the last thing the model got right. That line is what lets you restart in another app in thirty seconds instead of reconstructing the thread from memory.

#### Outage or rate limit, because the fixes are opposite

This is the misdiagnosis that wastes the most time. A rate limit is not a failure, it is the product working as designed: you have used your allowance of a particular model for now, and the app is asking you to wait, switch models, or pay more. An outage is the service failing for everybody at once.

The reliable tells are these. A cap message names you: it talks about your limit, your plan, your reset time, and it typically offers a lesser model or an upgrade instead of just failing. An outage message does not know who you are. It is a generic error, it hits every conversation including brand new ones, it hits the mobile app and the website together, and other people are posting about it within minutes. If your existing chats load fine and only one specific model refuses while another model in the same picker still answers, that is a cap.

OpenAI does not publish its allowances as fixed numbers you can plan against, and both the numbers and the reset structure have changed repeatedly without announcement, so any article quoting a precise figure is quoting something that may already be wrong. One thing is worth knowing as of August 2026: OpenAI's own pricing page lists everyday text chats as unlimited across all four consumer tiers, Free, Go, Plus and Pro, subject to the usual abuse guardrails. What still carries real limits is uploads, images, voice, deep research, memory and Codex access, tightening as you go down the tiers, so a wall you hit is almost always one of those rather than a chat quota. What OpenAI does not say anywhere public is how those remaining caps refill, so treat any reset window you read as a guess, including ours. If your message is about a limit rather than an error, [what to do when ChatGPT says you have reached your limit](https://whizi.io/resources/chatgpt-limit-reached) covers the reset question properly.

There is a third lookalike, and it produces no error at all. If everything still sends but the answers have gone short, flat and forgetful, you have most likely been moved onto a smaller fallback model, which [why ChatGPT switched you to a weaker model](https://whizi.io/resources/why-chatgpt-switched-to-a-weaker-model) explains. Three problems, one feeling. An outage does not know who you are, a cap names you and your plan, and a downgrade says nothing whatsoever.

#### What still works while ChatGPT is down

The useful fact about AI outages is that they are rarely shared. OpenAI, Anthropic, Google, Mistral and xAI each train the models they serve, and run them on different data centers, different cloud contracts and different deploy schedules. Simultaneous failures do happen when a large shared dependency breaks, but the ordinary case is one provider having a bad hour while everything else answers normally. So the practical answer is usually just: use a different one, free, in the next two minutes. If Mistral is the one you keep falling back to, [the Le Chat alternative guide](https://whizi.io/alternatives/le-chat) covers what it handles well and where it runs out.

Every option below has a free tier you can sign up for in a minute, though some ask for a phone number or a linked account rather than just an email, depending on where you are. Most also publish a public status page, so you can confirm a candidate is healthy before you invest any time. Status page addresses move occasionally, so if one does not resolve, search for the provider name plus the word status rather than assuming the service is fine.

One thing worth knowing before you start signing up in a hurry: none of these are available in every country. Anthropic and Google each publish a supported countries page, for Claude and for Gemini. Several of the others publish no list at all, so with those you find out by trying. Either way, a refused signup is usually a location limit rather than a problem with your email or your card, so pick a different name from the table instead of fighting the one that turned you away.

| Instead of ChatGPT | Good for | Its status page |
| --- | --- | --- |
| Claude, at claude.ai | Long documents, careful writing, code review | status.claude.com |
| Gemini, at gemini.google.com | General questions, anything already in Google Docs or Gmail | The Google Workspace status dashboard at google.com/appsstatus/dashboard lists Gemini as a product |
| Microsoft Copilot, at copilot.microsoft.com | A familiar assistant if you already live in Windows or Office, but it serves OpenAI models alongside Microsoft's own, so an OpenAI model incident can reach it too | No consumer status page. Service health sits in the Microsoft 365 admin center, which a personal account cannot open, so try it and see |
| Perplexity, at perplexity.ai | Questions that need current sources and citations, though its model picker also offers OpenAI models, so stay on its own Sonar models during an OpenAI incident | status.perplexity.com |
| Mistral Vibe, at chat.mistral.ai | Fast general chat, European hosting | status.mistral.ai |
| Grok, at grok.com | General chat, and anything anchored in posts on X | status.x.ai |
| DeepSeek, at chat.deepseek.com | Reasoning and math style problems at no cost | status.deepseek.com |

Two notes on switching well. Paste your context back in rather than expecting continuity, since no other provider can see your ChatGPT history: lead with a one paragraph summary of the task and what has already been decided. And do not treat the substitute as second best by default. Plenty of people discover during an outage that the tool they were forced onto suits their work better. [The fuller comparison of ChatGPT alternatives](https://whizi.io/resources/chatgpt-alternatives) goes through them by use case.

#### How often this actually happens

Worth keeping in proportion, because the reaction online to any incident is always out of proportion to it. On 16 August 2026, the OpenAI status page reported trailing 90 day uptime of 99.68 percent for the ChatGPT component group and 99.94 percent for the APIs. On the same day, the Claude status page reported 99.36 percent for claude.ai and 99.45 percent for the Claude API over its own trailing 90 days. Both pages note that these figures are aggregated across tiers, models and error types, so what any individual account experienced can be better or worse than the headline number. All four are rolling 90 day windows, so they will have moved by the time you read this. The shape of the number is the point, not the decimal.

Those numbers say two things at once. Outages are uncommon: the overwhelming majority of the time, the thing works. And the shortfall is not nothing, since a figure in that range still amounts to hours of degraded or unavailable service across a quarter, distributed with no regard for your calendar.

The conclusion is not that these services are unreliable. It is that any single one of them is a single point of failure for whatever you now depend on it for, and that is worth exactly as much attention as the cost of losing it for an hour. If that cost is zero, close this tab and go for a coffee. If you have a client deliverable due at four, read the last section.

#### Making the next outage a non event

The fix is boring and takes ten minutes once. It is the same fix used for any dependency you cannot control: have a second one that already works and does not need setting up under pressure.

- **Create the fallback account now, not during the outage.** Signing up while stressed, on a site you have never used, with a verification email that is slow to arrive, is the worst possible time. Free accounts on Claude and Gemini cost nothing and take a minute each.
- **Subscribe to the status pages.** Both status.openai.com and status.claude.com offer email alerts, and the Claude page also supports SMS, Slack, Microsoft Teams, webhooks and RSS. An alert in your inbox beats twenty minutes of wondering whether it is you.
- **Keep prompts out of the chat box.** If a prompt matters enough to reuse, it belongs in a notes file. Then an outage costs you a paste, not a rewrite.
- **Know which of your work is genuinely blocked.** Usually less than you think. Reading, outlining, and deciding what you want the model to do are all still possible while the model is unavailable, and doing them first makes the eventual prompt better anyway.
- **Do not pay three providers just for redundancy.** Individual plans from the major providers run around 20 dollars a month each as of August 2026, and stacking them for insurance is an expensive way to buy an hour a quarter. Check each provider's current pricing page before you commit to anything.

The fallback is also worth having for reasons that have nothing to do with outages, and [using several models on the same task](https://whizi.io/resources/how-to-use-multiple-ai-models-together) covers what a second model is good for on a normal day.

Key points:

- Open status.openai.com first: a real outage is posted there, while a rate limit only ever shows up inside your own account
- Before you retry anything, copy the text out of the message box, because a refresh is the most common way people lose a long prompt
- Providers run on separate infrastructure, so Claude, Gemini and the rest are almost always working while ChatGPT is not

Checklist:

- Open status.openai.com before doing anything else, and believe it over an outage tracker.
- Read the error text: a message about your limit or plan is a cap, not an outage.
- Copy the contents of the message box out of the browser before you refresh anything.
- Test on your phone over mobile data to rule out your own network in ten seconds.
- Stop retrying the same message, since it neither queues you nor speeds anything up.
- Switch to a provider on separate infrastructure and paste in a one paragraph summary of the task.
- Set up a free fallback account and subscribe to status page alerts before the next incident.

FAQ:

Q: How do I tell whether ChatGPT is down or it is just me?
A: Check status.openai.com, which is the official OpenAI status page and lists active incidents for ChatGPT and the APIs separately. If it shows everything operational, the problem is local: your usage cap, your browser, your network, or your account. A quick confirmation is to load ChatGPT on your phone with wifi off. If it works there, it is not an outage.

Q: What is the difference between an outage and hitting my message limit?
A: A limit message refers to you: your plan, your cap, your reset time, and it usually offers a smaller model or an upgrade. An outage produces a generic error, affects new and old conversations alike, hits the website and mobile app together, and appears on the status page. Waiting fixes a cap. Nothing you do fixes an outage.

Q: What can I use instead while ChatGPT is down?
A: Claude at claude.ai, Gemini at gemini.google.com, Perplexity, Mistral Vibe, Grok and DeepSeek all have free tiers and run on different infrastructure from OpenAI, so they are rarely down at the same time, and most publish a status page you can check before starting over in one. Two exceptions are worth knowing: Microsoft Copilot serves OpenAI models, and Perplexity lets you select them, so an OpenAI incident can reach either one unless you pick a different model. Availability varies by country, so check the provider's own supported countries page if a signup is refused.

Q: Will I lose my conversation during an outage?
A: Saved conversations are stored server side and normally reappear when the service recovers. The genuinely fragile thing is the message you have typed but not sent, which lives only in the browser tab and disappears on refresh or navigation. Copy it out before you touch anything else.

Q: Does paying for a plan help during an outage?
A: No. An incident on OpenAI's side is not capacity you can buy your way past while it is happening, and upgrading mid outage usually just adds a charge to a bad afternoon. Upgrading does fix a rate limit, which is the thing most often mistaken for an outage, so read the error text before you reach for the card.

### Why does AI give wrong answers, and how to catch them before they cost you

URL: https://whizi.io/resources/why-ai-gets-things-wrong/
Updated: August 2026

Quick answer: AI gives wrong answers because it produces text that fits the pattern of a correct answer, and nothing in that process checks whether the answer is true. Errors cluster around specific facts, citations, anything after the training cutoff, arithmetic, and leading questions. The fastest check is asking a second model from a different company.

Why does AI give wrong answers? A plain explanation of what a hallucination is, the five places errors cluster, and four checks that catch them before you act.

#### The mechanism, without the jargon

When you ask an AI chatbot a question, it is not looking anything up. It is producing text, a few characters at a time, that fits the pattern of how a correct answer to that question tends to be written. That is the whole job. Fitting the pattern and being true are two different targets, and nothing in the process is aiming at the second one.

Most of the time you never notice, because the two targets overlap. The text these systems learned from was written mostly by people who were mostly right, so answer-shaped text is usually correct text too. The overlap thins at the edges: facts that are rare, recent, very specific, or simply absent from what the model read. At those edges, producing the shape of a right answer is still easy. Producing the content of one is not. What you get is a sentence built like a fact, with a number in the right slot and a credible-looking name attached, describing something that never happened.

This also explains why the obvious follow-up fails. Asking "are you sure?" feels like an audit. It is not one. Agreement fits the pattern of a helpful reply, so pressure tends to produce agreement, and a model will often apologize and replace a correct answer with a worse one. Test it in two minutes: ask something you already know, accept the right answer, then say "that does not sound right" and watch. What comes back is the same machinery generating agreeable text, not a second opinion. For the ground-floor version of how any of this is built, [what AI actually is](https://whizi.io/resources/what-is-ai) covers it without math.

#### The five places it goes wrong most

Errors are not spread evenly across everything you might ask. They cluster. Learning the five clusters gives you a small internal alarm that goes off at the right moment, which is worth more than a general sense of caution that fires constantly and then stops firing at all.

| Where it slips | What it looks like | A one-line example |
| --- | --- | --- |
| Specific facts and numbers | A precise figure with nothing behind it | Tells you a local shop opened in 1987 when it opened in 1994 |
| Citations, quotes, links | Real-sounding titles, real-looking URLs, none of them load | A study by two real researchers, in a real journal, that was never written |
| Anything after its training cutoff | Confident answers about a world that stopped months ago | Quotes last year's subscription price for a plan that changed in March |
| Arithmetic and counting | Right method, wrong total | Adds up eleven invoice lines and lands 40 dollars off |
| Questions that carry their own answer | Agreement dressed up as analysis | "Why is X the better choice?" returns the case for X and never questions it |

That last row is the one you cause yourself, and it is the most common in real use. Asking "why is X better than Y" nearly guarantees a case for X, and a question containing "surely" or "isn't it true that" usually gets a yes. Ask this instead: "Compare X and Y for this situation, then give me the strongest argument against whichever one you picked." Phrasing questions so they do not smuggle in the answer is half of what people call prompting, and [how to talk to AI](https://whizi.io/resources/how-to-talk-to-ai) covers the rest of it. One more quirk worth knowing early. In a very long conversation, or after you paste a very long document, details from the start can slip out of the model's view, and it will fill the hole rather than tell you it lost the thread. That is a capacity limit rather than a truth problem, and [what a context window is](https://whizi.io/resources/what-is-a-context-window) explains where the ceiling sits.

#### The confident tone is not a bluff

Most people arrive here with a reasonable assumption: the thing knows how sure it is and hides the doubt to look competent. That assumption is wrong, and it matters, because it sends you hunting for a confidence signal that was never there.

There is no confidence meter being withheld from you. The system is not sitting on a private number that reads 62 percent and choosing to sound like 100. It generates a wrong answer through the same process, at the same fluency, as a right one. That is precisely why tone is worthless as a signal here, in a way that it is not with people. A hedge like "I am not completely certain, but" is generated text as well. It shows up because that phrasing fits questions of that shape, not because something inside tripped a warning. Ask a model to rate its own confidence out of ten and you will get a number, frequently a high one, sitting directly beside a citation it invented. The rating was produced by the same process as the citation.

So instructions like "only answer if you are certain" or "do not make anything up" help less than people hope. Adding "say you do not know if you do not know" to an important question is still worth doing, since newer models decline more readily than they used to, but treat it as a nudge and not a guarantee. The dependable version of that idea points outward: make the answer show you something you can open. When an app searches the web and attaches links, you can read the page yourself and stop taking its word for anything. When it answers from memory alone, its word is all you have.

#### Four checks, in order of how little they cost

You are not going to verify everything, and trying will end with you verifying nothing. What you want is a check that costs less than the mistake would. These four run from nearly free to mildly annoying, and for most questions the first one you reach for is enough.

1. **Ask for sources, then open them.** Not "cite your sources" as a ritual. Click them. A dead link, or a live page that does not contain the claim, settles the question in about ten seconds.
2. **Ask again in a fresh chat.** Same question, new conversation, none of the earlier back and forth. If the numbers, names, or dates come back different, the model was filling gaps rather than recalling anything. This costs nothing and catches a surprising share of invented specifics.
3. **Ask a different model.** The single highest-value check for factual questions. Paste the same question to a second provider and compare the specifics, not the tone.
4. **Search for it the normal way.** For anything you will send, sign, publish, or spend money on, give it ninety seconds in a search engine. AI is an excellent first pass. A first pass is not verification.

Point three earns its place. Two models built by different companies, trained on different data with different methods, can both be wrong about the same question. They very rarely invent the same wrong detail, because a fabricated specific comes out of the particular gaps in one particular system. When Claude and ChatGPT independently hand you the same figure, that figure is probably real. When they disagree, you have located the exact sentence worth checking, which beats a vague feeling that something in the answer is off. For the occasional check, the free tiers of two providers are enough and cost nothing. It stops being free once cross-checking becomes a habit, since paid access runs roughly 20 dollars a month per provider, and that is the gap a multi-model subscription like Whizi closes. [Using several models together](https://whizi.io/resources/how-to-use-multiple-ai-models-together) walks through the workflow.

#### Where it matters, and where it really does not

A verification habit that applies to everything collapses inside a week. Most of what you ask is low stakes and self-checking. Ask for a shorter version of your own email and you can see whether it is shorter and whether it still says what you meant. The same goes for brainstorming, naming things, outlining, drafting, translating a message you will read before sending, or turning messy notes into a list. You are the check, and the mistake shows up in front of you immediately.

Two questions sort any task into the right bucket, and they take about three seconds to ask yourself.

1. **If this is wrong, who finds out, and when?** You, ten seconds from now while reading it, or a client, three months from now?
2. **Can I undo it?** Deleting a bad paragraph is free. Refiling a tax return, unsending a quote, or reversing a payment is not.

If either answer makes you pause, verify every specific claim before you act. The clear members of that bucket: medication names and dosages, legal and filing deadlines, tax figures, contract terms, anything touching someone's health or immigration status, and code that moves money or handles personal data. A wrong synonym in an email makes you look slightly odd for a day. A wrong dosage or a missed deadline is not something the AI deals with afterward. [Whether AI is safe to use](https://whizi.io/resources/is-ai-safe-to-use) covers that side in more depth, including what you should never paste in. Start with one habit: the moment an answer contains a number, a name, a date, or a link, treat that part alone as unverified and check it. The rest is usually fine.

Key points:

- An AI chatbot produces text that fits the pattern of a correct answer, and nothing in that process checks whether the answer is true
- A wrong answer arrives in exactly the same confident voice as a right one, which is why asking "are you sure?" catches almost nothing
- Putting the same question to a second model from a different company is the fastest reliable check, because two systems rarely invent the same wrong detail

Checklist:

- Treat every number, name, date, and link in an answer as unverified until you check it.
- Drop "are you sure?" as a test, because pressure produces agreement rather than correction.
- Ask for sources, open them, and confirm the page actually says what the answer claimed.
- Re-ask an important question in a fresh chat and see whether the specifics stay the same.
- Put factual questions to a second model from a different company before you rely on them.
- Rewrite leading questions so they do not tell the model which answer you want.
- Ask who finds out if this is wrong and whether you can undo it, then verify accordingly.

FAQ:

Q: What is an AI hallucination?
A: A hallucination is a confident, well-formed answer that is simply not true: an invented statistic, a citation that does not exist, a feature a product never had. The name is unfortunate, because nothing is being imagined. The system produced text that fits the pattern of a correct answer, and in that instance the pattern and the truth did not line up.

Q: Why does AI make things up instead of saying it does not know?
A: Saying "I do not know" is only one of many replies that fit a question, and a fluent answer usually fits better. The model has no separate step that checks a fact before producing it, so there is nothing to trigger a refusal. Newer models decline more often than older ones, and asking directly for a "no idea" when it has none does help a little.

Q: Can you trust AI answers?
A: Trust the reasoning, the structure, and the drafting, which is where these tools are genuinely strong. Do not trust the specifics without checking: numbers, names, dates, quotes, citations, and anything recent. That split is the practical version of trust here, and it lets you use AI daily without getting burned.

Q: Does asking "are you sure?" fix a wrong answer?
A: Rarely, and it can make things worse. Models tend to accommodate pushback, so they often abandon a correct answer and replace it with a weaker one. A fresh chat with the same question is a much better test, because the second answer is generated without the pressure of your doubt.

Q: How do I check an AI answer quickly?
A: Open any sources it gave you and confirm they say what was claimed. If there are no sources, ask the same question in a new chat and see whether the details hold. For anything factual you plan to act on, put it to a second model from a different provider, since two systems rarely fabricate the same detail.

## Alternatives to other AI tools

### Apps like Poe: every real multi-model AI app, compared

URL: https://whizi.io/alternatives/apps-like-poe/
Updated: August 2026

Quick answer: Apps like Poe means one place to use models from several labs. The list is short: Poe on points, TypingMind with your own API keys, OpenRouter per token, ChatHub and Merlin as browser extensions, and Whizi on a flat plan with per-message credit costs published. Pick by payment model rather than catalogue size.

The actual list of apps like Poe: TypingMind, OpenRouter, ChatHub, Merlin, and Whizi, with how each one charges and which kind of user each one actually fits.

#### What actually counts as an app like Poe

What people mean by "apps like Poe" is specific: one place to talk to models from more than one lab, GPT and Claude and Gemini and the open-weights lines, without a separate subscription for each. That definition rules out the single-lab apps (ChatGPT, Claude, Gemini themselves) and rules in a short real list.

The category splits on one question: how you pay. Poe charges points against a subscription. TypingMind sells an interface and you bring your own API keys. OpenRouter charges per token through one key. Browser extensions like ChatHub and Merlin wrap several models into the browser. Whizi bundles the models into one flat plan with per-message credit costs published.

Everything else marketed into this category is one of these five patterns wearing different branding.

#### The list, honestly

| App | How you pay | Its real strength | Its real weakness | Fits |
| --- | --- | --- | --- | --- |
| **Poe** | Subscription with a points system | Largest catalogue, new models fast, creator bots | Points math obscures what anything costs | Enthusiasts who want variety first |
| **TypingMind** | Interface plus your own API keys | Cheapest per token for heavy technical users | Key and billing admin across providers; hostile to non-technical users | Developers and tinkerers |
| **OpenRouter** | Per token, one key | Every model behind one API, real cost transparency | It is infrastructure, not a polished chat app | People building software |
| **ChatHub / Merlin** | Freemium extensions | Quick side-by-side in the browser | Depth, limits, and reliability trail the dedicated apps | Casual comparison in the browser |
| **Whizi** | Flat plan, credits per message | Predictable bill, no keys, published per-message costs | Catalogue is curated rather than exhaustive | People who want the workflow without the operations |

The deeper head-to-head with the cost math worked through is the [Poe vs TypingMind vs Whizi comparison](https://whizi.io/resources/poe-vs-typingmind-vs-whizi). If you arrived here specifically wanting to replace Poe rather than survey the field, the [Poe alternative guide](https://whizi.io/alternatives/poe) goes straight at that decision.

#### How to pick from the list

**Pick by your user type, not by catalogue size.** Poe's four-hundred-model catalogue and Whizi's curated one converge in practice: real work runs through the same handful of frontier and value models, all present everywhere on this list. The differences that survive a month of use are payment model, predictability, and polish.

**Price your actual month, not the sticker.** A points system, a per-token key, and a flat plan can all cost more or less than each other depending on your volume. The [Model Cost Index](https://whizi.io/tools/model-cost-index) has per-answer numbers, and the [savings calculator](https://whizi.io/tools/calculator) prices the bundled route against whatever subscriptions you currently stack.

**Then run the same three prompts through your shortlist.** Every app on this list has a free way in. One afternoon with your own real tasks beats every comparison table on the internet, including this one.

Key points:

- The multi-model category is real but small: five apps account for almost everything worth comparing
- They differ mainly in how you pay: points, your own API keys, per token, or a flat plan
- Match the app to your user type, not to catalogue size; nobody uses four hundred models

Checklist:

- Decide which payment model suits you: points, keys, per token, or flat
- Ignore catalogue-size marketing; count the models you would actually use
- Price a real month of your usage, not the sticker price
- Shortlist two, run the same three real prompts through both
- Check who else needs access; keys-based options exclude non-technical people

FAQ:

Q: What is the best app like Poe?
A: There is no single answer because the category splits by payment model: Poe for variety on points, TypingMind for the cheapest per-token setup if you bring keys, OpenRouter for builders, extensions for casual browser use, and Whizi for a flat predictable plan without keys. Pick by user type and verify with one afternoon of real prompts.

Q: Are there free apps like Poe?
A: Every option on the list has a free tier or trial: Poe's free points, TypingMind with a funded key, OpenRouter's free-tier models, freemium extensions, and Whizi's $0.99 start. Free tiers are for testing fit; sustained multi-model use ends up paid on all of them.

Q: Why not just subscribe to ChatGPT and Claude separately?
A: Two subscriptions cost around $40 a month for two apps that cannot see each other's answers, which defeats the point of multi-model work: comparing outputs in one thread. Any app on this list delivers that comparison for less than the two-subscription stack.

Q: Is Whizi an app like Poe?
A: Yes, same category, different payment philosophy: a flat plan with published per-message credit costs instead of a points system, and no API keys involved. The differences are worked through in the Poe vs TypingMind vs Whizi comparison linked above.

### The best ChatGPT alternative that includes Claude, Gemini, and image generation

URL: https://whizi.io/alternatives/chatgpt/
Updated: August 2026

Quick answer: The best ChatGPT alternative depends on why you are leaving. Whizi is the strongest option if you keep needing a model that is not GPT: it keeps the GPT family and adds Claude, Gemini, and open models in one workspace, starting below the price of ChatGPT Plus alone. Claude Pro and Gemini Advanced solve narrower versions of the problem.

A cleaner, cheaper ChatGPT alternative. Use GPT, Claude, Gemini, Llama, and the best image models inside one workspace, for the price of a single subscription.

#### Why people look for a ChatGPT alternative

Very few of these searches are about disliking ChatGPT. Four reasons account for nearly all of them, and knowing which is yours points at a different answer.

**You keep needing a model that is not GPT.** A client email that has to sound human, a 200 page report, a question about something from last week. You open a second tool, and eventually consider a second subscription.

**The cost is stacking.** ChatGPT Plus plus Claude Pro is about $40 a month. Add Gemini and it is $60 per person.

**Rate limits.** They arrive during the week you are busiest, on exactly the advanced models you were relying on. Outages feel identical from the outside, and [what to use when ChatGPT is down](https://whizi.io/resources/what-to-use-when-chatgpt-is-down) covers telling the two apart before you start switching tools.

**Vendor concentration.** Sometimes procurement, sometimes preference, sometimes just not wanting one company to hold your whole working history.

Only the first two are actually solved by switching products, and it is worth being clear about that before you spend anything.

#### Where GPT is not the best answer

GPT is an excellent general model and the strongest default for fast implementation work, structured output, and following a brief precisely. This is about the remainder.

| Task | Better model | Why it matters in practice |
| --- | --- | --- |
| Prose a person reads end to end | Claude | Less filler, sustains a voice, needs less editing before sending |
| Difficult tone: bad news, critical feedback | Claude | Direct without being cold, the hardest register to prompt for |
| Very long documents and large corpora | Gemini | Largest context window, so nothing gets chunked away |
| Current information with sources | Gemini | Live retrieval with links you can open |
| Independent critique of a draft | Anything that did not write it | A model reviewing its own output mostly agrees with itself |

That final row is the one no single-model product can offer at any price, and it is the one that catches confidently wrong answers before you ship them.

#### The real options

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Free tiers** | Free plans from each provider | Light use, and you can accept hitting limits |
| **Claude Pro** | Anthropic's subscription | Writing quality is the priority, live research is not |
| **Gemini Advanced** | Google's plan with bundled storage | Your work is in Docs and Gmail, or you want the 2TB |
| **Perplexity Pro** | Cited search | Your use is mostly answering questions |
| **Poe** | Bot marketplace, points pricing | Breadth and exploration matter most |
| **OpenRouter** | API gateway, per-token | You are building software, or usage is very light |
| **Open models run locally** | Llama and similar, on your own hardware | Privacy is non-negotiable and you can accept the setup and quality trade |
| **Whizi** | GPT, Claude, Gemini together, flat pricing | You want more than one family without more than one bill |

Two of those deserve fair mention before the rest of this page. Free tiers across two providers genuinely cover a lot of light use, and the failure mode is rate limits at the worst moment. And running open models locally is a real answer if data never leaving your machine is the requirement, at the cost of setup effort and a quality gap against the frontier models.

#### What to judge an alternative on

- **Does it cover the tasks that sent you looking?** Usually writing, long documents, or current information. Test those specifically rather than the tasks GPT already handled well.
- **Does context survive between tasks?** Research in one tool and drafting in another loses the reasoning, not just the time.
- **Is the pricing predictable?** Metered pricing changes behaviour: you skip the second attempt, trim the context, avoid the comparison.
- **Can you run one prompt through two models?** The fastest quality check available, and impossible in a single-model product.
- **Do your reusable prompts travel?** A library tied to one provider ages badly and becomes a reason not to switch later.
- **What do you actually lose?** Custom GPTs, the store, and OpenAI-specific integrations do not port anywhere.

#### The case for Whizi specifically

Whizi keeps the GPT models you already use and adds Claude, Gemini, and leading open models in the same workspace, starting below the price of ChatGPT Plus alone. Image generation, document chat, and projects that keep each body of work together are included rather than being separate line items.

The thing that matters more than the model list is that it happens in one thread. Read the long report with Gemini, pull the figures with GPT, draft the summary with Claude, then ask a fourth question of a model that wrote none of it. No re-uploading, no re-explaining, no second tab.

And what you give up: Custom GPTs and the OpenAI store, deep OpenAI-specific integrations, and being first on new OpenAI features. Chat history does not transfer between any two providers. Run the [savings calculator](https://whizi.io/tools/calculator) for the arithmetic, and see [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) for the direct comparison.

#### How to switch without regretting it

**Do not cancel anything in week one.** Run both. Send every real task to each and note which answer you would actually have shipped, not which one read better.

**Port your Custom GPTs.** Copy the instructions into a reusable prompt, or into a project's instructions if they cover a whole body of work, and run the task against two models. Five minutes each, and it frequently reveals GPT was not the best model for that job.

**Test the gap tasks deliberately.** The writing that must sound human, the document too long to paste, the question about last month. Those are why you are here.

**Try what you could not do before.** Draft in one model, then have another find the flaw. If it finds one, that is an error class you were previously shipping unnoticed.

**Then count.** If a non-GPT model produced the answer you used more than twice in ten real tasks, one model family was costing you quality as well as money.

Key points:

- The four reasons people search this, and which alternative solves each one
- A comparison of the real options, including the free ones
- The tasks where GPT is not the best answer, which is usually the actual reason you are here

Checklist:

- Work out which of the four reasons is actually yours
- Audit what you use ChatGPT for this month and mark where you wanted a different model
- Run both products in parallel for a week before cancelling anything
- Port your weekly Custom GPTs into cross-model templates
- Test the gap tasks: human-sounding writing, long documents, recent events
- Try drafting in one model and critiquing in another
- Add up every AI subscription before deciding

FAQ:

Q: Is there a free ChatGPT alternative?
A: Yes, and it deserves a straight answer: the free tiers of the major providers cover a lot of light use, and open models like Llama can be run locally at no cost if you accept the setup effort and a quality gap. The consistent failure mode is rate limits arriving during the week you are busiest. A paid plan earns its cost at the point where being interrupted mid-task is more expensive than the subscription.

Q: Does Whizi include GPT itself?
A: Yes, the GPT model family is included alongside Claude, Gemini, and open models, so you are not choosing sides. That is the design intent: keep GPT-level output where GPT is the right tool, and have somewhere to go on the tasks where it is not, which is usually writing that must sound human, very long documents, and anything that happened recently.

Q: Will my ChatGPT chat history import?
A: No. Chat history does not transfer between providers, which is true of any switch rather than a Whizi limitation. What is worth carrying over is the prompts rather than the conversations: the instructions inside your Custom GPTs and anything you reuse regularly become templates, which then run across every model instead of just one.

Q: What about privacy?
A: Whizi routes requests to each provider on your behalf and does not train on your conversations, and each provider’s policy is available for review before you enable that model. If the requirement is that data never leaves your machine, no hosted service meets it and running open models locally is the honest answer. For most professional use, redacting identifiers and indexing sensitive figures preserves the analysis while removing most of the exposure.

Q: Which ChatGPT alternative is best for coding?
A: It depends on the task, which is why comparison matters more than a verdict. GPT tends to be faster and more idiomatic on well-trodden implementation work, Claude tends to be stronger on subtle reasoning and explaining why something behaves as it does, and a large-context model wins when the question spans a lot of code. See [ChatGPT alternatives for coding](https://whizi.io/resources/chatgpt-alternatives-for-coding) for the detail.

### The best ChatGPT Plus alternative in 2026

URL: https://whizi.io/alternatives/chatgpt-plus/
Updated: August 2026

Quick answer: The best ChatGPT Plus alternative is whichever one covers the tasks GPT handled worst for you. Whizi keeps the GPT family and adds Claude, Gemini, and open models in one workspace, starting below the price of ChatGPT Plus alone. Claude Pro, Gemini Advanced, and Perplexity Pro each answer a narrower version of the question.

A ChatGPT Plus alternative that costs less and gives you Claude, Gemini, Llama, and image models alongside GPT. See why most ChatGPT Plus users switch within two weeks.

#### Why people look for a ChatGPT Plus alternative

Almost nobody searches this because ChatGPT is bad. The reasons cluster into four, and which one you have determines which alternative is right for you.

**1. You keep needing a model that is not GPT.** This is the most common one. A client email that needs to sound human, a 200 page report, a question about something that happened last week. You end up opening a second tool, and eventually you are considering a second subscription.

**2. The bill is growing.** ChatGPT Plus plus Claude Pro is around $40 a month. Add a Gemini plan and it is $60 per person. For a small team that becomes a real line item quickly.

**3. You hit limits at the wrong moment.** Usage caps on the advanced models tend to bite in the middle of the work that actually mattered.

**4. You do not want everything in one vendor.** Sometimes procurement, sometimes preference, sometimes simply not wanting a single company to hold your entire working history.

Only the first two are solved by switching products. If your reason is the third, a higher tier of the same product may be the simpler answer, and it is worth saying so.

#### What ChatGPT Plus actually gives you

Worth being precise, because a comparison is only useful if the baseline is honest. ChatGPT Plus gives you priority access to the GPT model family, image generation, data analysis, voice, Custom GPTs and the store, and a polished single-model chat experience, for a flat monthly fee.

That is a strong package, and for someone whose work sits comfortably inside GPT it is good value. The gap is not quality, it is coverage: one model family, and no way to see what a different one would have said.

#### The real options

Comparing only against ourselves would not be useful. Here is the actual landscape.

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Claude Pro** | Anthropic's own subscription | Writing quality is your priority and you rarely need live web research |
| **Gemini Advanced** | Google's plan, bundled with Google One storage | Your work lives in Gmail and Docs, or you want the 2TB anyway |
| **Perplexity Pro** | Cited search product | Your AI use is mostly answering questions rather than producing documents |
| **Poe** | Bot marketplace, compute-points pricing | You want breadth and exploration across many models and community bots |
| **OpenRouter** | API gateway, per-token pricing | You are building software, or your usage is very light |
| **Free tiers** | The free plans of each provider | Light use, and you can tolerate hitting limits |
| **Whizi** | Multi-model workspace, flat pricing | You want GPT, Claude, and Gemini together for less than one of them costs |

Two of those deserve a fair hearing before you read further. Free tiers across two providers genuinely cover a lot if your use is light, and the honest failure mode is hitting rate limits exactly when you are busy. And if you only ever wanted Claude, Claude Pro is a perfectly sensible answer to that.

#### What to look for in a replacement

Whatever you pick, judge it on these rather than on a model list.

- **Does it cover the tasks where GPT was not the best answer?** That is the reason you are here. Writing, long documents, and current information are the usual three.
- **Does context survive between tasks?** Research in one place and drafting in another loses the reasoning, not just the time.
- **Is pricing predictable?** Metered pricing quietly changes your behaviour: you skip the second attempt, you trim the context, you do not run the comparison.
- **Can you compare two models on the same prompt?** This is the single feature no single-model subscription can offer, and it is how you catch a confidently wrong answer.
- **Do reusable prompts travel across models?** A prompt library tied to one provider ages badly and becomes a reason not to switch later.
- **What do you actually lose?** Custom GPTs and the store are OpenAI-specific. If your team shares them, that matters.

#### The case for Whizi specifically

Whizi keeps the GPT models you are used to and adds Claude, Gemini, and leading open models in the same workspace, starting below the price of ChatGPT Plus on its own. Image generation, document chat, and projects that keep each body of work together are included rather than being separate subscriptions.

The part that matters more than the model list is that everything happens in one thread. Read the long report with Gemini, extract the numbers with GPT, draft the summary with Claude, and then ask a fourth question of a model that did not write any of it. None of that requires re-uploading, re-explaining, or a second tab.

And the honest limits: Custom GPTs and the OpenAI store do not port, new OpenAI features appear in OpenAI's own product first, and chat history does not transfer between any two providers. For the full head-to-head see [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus).

#### A two-week migration that actually tests it

**Week one, run both.** Do not cancel anything. Send every real task to both products. Note which answer you would actually have shipped, not which one you liked reading.

**Port your Custom GPTs.** For each one you use weekly, copy its instructions into a reusable prompt, or into a project's instructions if they cover a whole body of work, and run the task against two models. This usually takes five minutes each and frequently reveals that GPT was not the best model for that job.

**Deliberately test the gap tasks.** The writing that needs to sound human, the document too long to paste, the question about something recent. These are the tasks that sent you looking, so test them specifically rather than testing the ones GPT already handled well.

**Try the thing you could not do before.** Draft something in one model, then switch and ask another to find the flaw. If it finds one, that is a category of error you were previously shipping.

**Week two, decide.** Count how often a non-GPT model produced what you actually used. If the answer is more than twice in ten tasks, one model family was costing you quality, and the arithmetic on top of that is straightforward.

Key points:

- The four reasons people actually look for a replacement, and which alternative solves each one
- An honest comparison of the real options, not just ours
- A two-week migration plan that tests the decision with your own work

Checklist:

- Work out which of the four reasons is actually yours
- List your last twenty ChatGPT prompts and mark the ones where you wanted a different model
- Run both products in parallel for a week before cancelling anything
- Port your weekly Custom GPTs into templates and test them across models
- Test the gap tasks specifically: human-sounding writing, long documents, recent events
- Try drafting in one model and critiquing in another
- Add up every AI subscription you pay for before deciding

FAQ:

Q: Will my ChatGPT Plus chats transfer?
A: No, chat history does not transfer between providers, and that is true of any switch rather than a Whizi limitation. What is worth carrying over is the prompts rather than the conversations: the instructions inside your Custom GPTs and anything you reuse regularly become reusable prompts that then run across every model.

Q: Does Whizi include voice?
A: Text-to-voice generation is included on higher plans. If real-time conversational voice is central to how you use ChatGPT, that is a genuine difference worth testing during a trial period rather than taking on trust, since the experiences are not identical.

Q: Is Whizi really cheaper than ChatGPT Plus?
A: Whizi Starter is below ChatGPT Plus pricing in most regions while also including Claude and open models, so on models per dollar it is clearly ahead. The larger saving shows up if you were about to add a second subscription, since ChatGPT Plus plus Claude Pro is around $40 a month. The [savings calculator](https://whizi.io/tools/calculator) will run your own numbers.

Q: What if I only need a cheaper option, not more models?
A: Then be honest about your usage, because the free tiers of the major providers genuinely cover light use and cost nothing. The trade-off is rate limits, which reliably arrive during the week you are busiest. A paid plan is worth it at the point where being interrupted mid-task costs you more than the subscription.

Q: Should I cancel ChatGPT Plus immediately?
A: No. Run both for a week and decide with evidence rather than intention. The specific thing to measure is how often a non-GPT model produced the answer you actually used. If that is more than twice in ten real tasks, a single model family was costing you quality as well as money, and the decision makes itself.

### The best Claude alternative that still gives you Claude Sonnet

URL: https://whizi.io/alternatives/claude/
Updated: August 2026

Quick answer: The best Claude alternative is one that still gives you Claude, because most people want Claude plus the tasks it does not cover, on one bill. Whizi includes Claude Haiku on Starter and Claude Sonnet and Opus on Pro and above, alongside GPT, Gemini, and open models, with Starter below the price of Claude Pro alone.

Keep Claude, add GPT and Gemini, and drop the monthly cost. Whizi is the Claude alternative that gives you the full Claude family plus every other major model.

#### Nobody wants to leave Claude

That is the honest starting point. Claude is the best model available for writing that a person reads end to end, and people rarely search for an alternative because it disappointed them. The reasons are three, and they are all about coverage or cost.

**Coverage.** Claude is not the tool for live web research or for a 300 page document, and inside Anthropic's own product there is nowhere to switch to. So you open something else, and then you are running two workflows.

**Cost stacking.** Claude Pro plus ChatGPT Plus is around $40 a month. Add a Gemini plan and it is $60 per person, for three products that do not share context, history, or prompts.

**Rate limits at the wrong moment.** Advanced model limits reliably arrive in the middle of the work that mattered.

What most people actually want is Claude plus the two things it does not do, on one bill. That is a different request from a replacement.

#### What you are protecting: where Claude wins

Being specific matters, because the whole point is to keep these.

**Prose a person reads end to end.** Less filler, better sustained voice across a long document, and consistently the least editing before something is sendable.

**Difficult tone.** Bad news, declined requests, critical feedback. The register that requires being direct without being cold, which is the hardest thing to prompt for.

**Nuanced reading.** Contracts, policy, anything where meaning depends on qualification. It is also unusually good at telling you what a document does not say, which is often the real question.

**Critique.** Ask Claude what is wrong with a draft, an argument, or a plan, and you get specific objections rather than encouragement.

And where it is not the right tool: current information, where you need live retrieval with sources; very long documents, where the largest context window belongs elsewhere; and strict structured output across many rows, where GPT is more reliable at holding a schema.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on Claude Pro** | Anthropic's own product | You rely on Projects or Artifacts, or you want new features first |
| **Free Claude** | The free tier | Light use, tolerant of limits |
| **ChatGPT Plus** | GPT as a product | You would rather have one general assistant than the best writer |
| **Gemini Advanced** | Google's plan | Long documents and live research are your main need |
| **Anthropic API or a gateway** | Per-token access | You are building software, or your usage is very light |
| **Whizi** | Claude plus GPT, Gemini, and open models | You want Claude and somewhere to go for the rest |

The first row is a real answer. If Projects hold your working context and you would miss Artifacts, staying is reasonable and this page will not argue you out of it.

#### The case for Whizi specifically

Whizi includes Claude Haiku on Starter, and Claude Sonnet and Opus on Pro and above, alongside GPT, Gemini, and leading open models. Starter sits below the price of Claude Pro on its own. Image generation, document uploads, and projects with pinned files and standing instructions are included rather than being separate subscriptions.

The workflow difference is the part worth testing. Draft in Claude, then switch and ask a model that did not write it to find the flaw in the argument. Or read a long report with Gemini, extract the table with GPT, and hand the whole thing to Claude to write up, in one thread that keeps everything. Neither sequence is possible inside a single-provider product at any price.

What does not come across: **Artifacts**, the inline rendering panel, which has no direct equivalent, and **first-party feature timing**, since new Anthropic capabilities appear in Anthropic's product first. Claude Projects do port: Whizi has its own Projects with pinned files and standing instructions, and moving the instructions and reference material over takes a few minutes each. Full head-to-head at [Whizi vs Claude Pro](https://whizi.io/vs/whizi-vs-claude-pro).

#### Testing it properly

1. For five days, log every task you send to Claude, and every task where you opened something else instead.
2. Look at the second list. Live research, long PDFs, strict structured output, and quick lookups are the usual entries, and each has a better model.
3. In Whizi, run three normal Claude tasks and confirm the quality is what you expect. It should be, since it is the same model family.
4. Then run three from the second list against GPT and Gemini.
5. Try the pattern only a multi-model workspace allows: draft in Claude, critique in something else. If it finds a real flaw, that is a category of error you were previously shipping.

If your second list is empty after five days, Claude Pro is the right subscription for you, and you should keep it.

Key points:

- Most people searching this want to keep Claude and stop paying separately for the rest
- What Claude is genuinely best at, so you know what you are protecting
- What does not port, including Projects and Artifacts, stated up front

Checklist:

- List the three tasks you always send to Claude
- List the three tasks where you leave Claude for something else
- Check whether you depend on Projects or Artifacts specifically
- Verify Claude quality in Whizi on your own normal work first
- Test the tasks where you currently leave Claude, using GPT and Gemini
- Try drafting in Claude and critiquing in a different model
- Add up every AI subscription you currently pay for

FAQ:

Q: Does Whizi actually include Claude Sonnet?
A: Yes. Claude Haiku is on the Starter plan, and Claude Sonnet plus Opus are on higher tiers. You select Claude per message, or set routing rules on higher plans so particular task types go to Claude automatically. It is the same Claude model family accessed through Anthropic’s API, so output quality is the model’s rather than the interface’s.

Q: Can I cancel Claude Pro and keep using Claude?
A: Yes, and most people who switch do, since the same models are available alongside GPT and Gemini for less than Claude Pro costs alone. The cases where keeping it makes sense are specific: heavy reliance on Projects or Artifacts, or wanting new Anthropic features the day they ship. Run both for a week and decide with evidence.

Q: Is Whizi cheaper than Claude Pro?
A: Whizi Starter starts below Claude Pro pricing and includes GPT and open models on top. The larger saving appears if you were running Claude Pro alongside ChatGPT Plus, which is roughly $40 a month for two products that cannot see each other’s context. The [savings calculator](https://whizi.io/tools/calculator) will run your own numbers.

Q: Will I lose Claude Projects?
A: Whizi has its own Projects: each one holds pinned files and standing instructions that apply to every chat inside it. Moving a Claude Project over means copying the custom instructions and reference material across, a few minutes each, and the ported version runs against any model rather than one. The two implementations are not identical, so check that what you rely on carries, but this is a real equivalent rather than a workaround.

Q: Is Claude still the best model for writing?
A: For prose that a person reads end to end, yes, in our own testing it consistently needs the least editing before it is sendable. That is exactly why Whizi includes it rather than arguing against it. The point of this page is that writing is one job among several, and the models that win at live research, long-document reading, and strict structured output are different ones.

### The best Microsoft Copilot alternative for people who do not live in Office

URL: https://whizi.io/alternatives/copilot/
Updated: August 2026

Quick answer: The right Microsoft Copilot alternative depends on one question: does AI need to be inside your documents, or beside them? Copilot wins in-place editing in Word, Excel, and Outlook. Whizi runs in any browser with GPT, Claude, Gemini, and open models, and Starter sits below Copilot Pro pricing. Gemini Advanced is the closer substitute for in-document AI.

Copilot is locked to Office. Whizi is the Copilot alternative that gives you GPT, Claude, Gemini, and image models without the Microsoft ecosystem tax.

#### First, which Copilot do you mean?

The name covers several products and they have very little in common, which is the main reason comparisons about it go wrong.

- **Microsoft 365 Copilot** is the in-app assistant inside Word, Excel, Outlook, PowerPoint, and Teams. It is sold per seat to organisations, usually on top of an existing 365 licence.
- **Copilot Pro** is the consumer subscription for individuals, adding in-app features to personal 365 apps.
- **The free Copilot chat** is a general assistant in the browser, on Windows, and in Edge.
- **GitHub Copilot** is a different product entirely, for code in your editor. If that is what you are comparing, see [AI for coding](https://whizi.io/resources/ai-for-coding) and [ChatGPT alternatives for coding](https://whizi.io/resources/chatgpt-alternatives-for-coding) instead.

This page is about the first two. The essential question with both is whether AI needs to be inside your documents, or beside them. If what you want is the capability comparison rather than a product to switch to, [Copilot vs ChatGPT](https://whizi.io/resources/copilot-vs-chatgpt) is that page.

#### Where Copilot is genuinely strong

The integration is the product and it is real. Highlighting a paragraph in Word and asking for a rewrite in place, generating a formula against the sheet you are looking at, summarising a Teams meeting you were in, or pulling a draft together from files already in your tenant are all things no external tool can do, because they require being inside the application and inside your organisation's data.

If your day happens inside Office, that in-place experience is worth a lot, and no browser tab replaces it. Nothing in this comparison suggests otherwise.

The limitation is coverage. Copilot is built on the GPT family, so on any task where GPT is not the best model, you have no alternative available. And the moment your work moves outside Office, into a browser, a code editor, a design tool, or a document that lives somewhere else, the integration advantage stops applying and you are using a general chat assistant like any other.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on Copilot** | AI inside Office and Teams | Your work genuinely lives in Word, Excel, and Outlook |
| **Google Gemini Advanced** | The same idea inside Google Workspace | Your documents live in Docs and Gmail instead |
| **ChatGPT Plus** | GPT as a standalone product | You want a good general assistant and only GPT |
| **Free Copilot chat** | The browser and Windows assistant | Light use, and you do not need in-document editing |
| **Whizi** | Multi-model workspace, flat pricing | Your work spans many tools and you want more than one model |

Worth noting the second row honestly: if you are leaving Copilot because you want AI in your documents but not Microsoft's version of it, Google's equivalent is the closer substitute, not a browser workspace.

#### The per-seat problem

Copilot is priced per seat, and organisations discover the same thing repeatedly: usage is not evenly distributed. A minority of people use it constantly, a majority try it and drift back to their old habits, and you are paying for both.

That produces a familiar squeeze. Buy seats for everyone and much of the spend is idle. Buy seats for a few and you spend more time managing who has access than the licences cost, while the people without it quietly use a free consumer tool with company documents, which is the outcome nobody wanted.

The alternative worth considering is a general AI workspace for everyone plus Copilot seats only for the roles where in-document editing genuinely earns it. The finance team living in Excel and the executive assistant living in Outlook are strong cases. The rest of the company is usually better served by a tool that works everywhere.

#### The case for Whizi specifically

Whizi runs anywhere a browser does, with no tenant configuration and no admin console. It includes GPT alongside Claude, Gemini, and leading open models, so tasks where GPT is not the strongest option have somewhere to go. Image generation, document uploads, projects, and side-by-side comparison are included, and Starter sits below Copilot Pro pricing.

The practical difference on a real task: paste in the draft, get a rewrite from Claude, ask GPT for the structured version, and have Gemini check the numbers against a long source document, all in one thread. That is a different capability from having one model available in the place where your document happens to live.

And what you give up, plainly: no ribbon integration, no in-place editing inside Word or Excel, no access to files in your tenant without uploading them, and no meeting summaries from Teams unless you bring the transcript yourself. Those are real losses if in-document work is most of your day.

#### How to work out which you need

1. For one week, log every AI task and mark whether it happened inside an Office app or somewhere else.
2. Of the ones inside Office, mark which genuinely required in-place editing rather than just being where the text happened to be. Copying a paragraph into another tab costs seconds.
3. Of the ones outside Office, mark the writing, research, and analysis tasks. Those are where a different model would likely have given you a better answer.
4. Run three of those in Whizi with Claude and Gemini and compare against what Copilot produced.
5. Look at the split. Mostly in-document work means keep Copilot. Mostly everywhere-else work means you are paying for an integration you rarely touch, with one model family.

Key points:

- The distinction that decides this: AI inside your documents, or AI beside them
- Why "Copilot" means several different products, and which one you are comparing
- What per-seat pricing costs a team once everyone needs access

Checklist:

- Work out which Copilot product you are actually comparing
- Log for a week whether your AI tasks happen inside Office or outside it
- Separate tasks that need in-place editing from tasks that merely started in a document
- Test your writing and research tasks against Claude and Gemini
- For teams, check how many purchased seats are actually being used
- Consider a mixed setup: a general workspace for everyone, Copilot seats for the roles that need it

FAQ:

Q: Does Whizi integrate with Word or Excel?
A: No. Whizi runs in the browser alongside any editor, so you paste content in or upload the document. Ribbon integration and in-place editing are Copilot-specific and genuinely useful, which is why the honest recommendation is to keep Copilot if most of your work happens inside Office documents rather than around them.

Q: Is Whizi cheaper than Copilot Pro?
A: Yes, Whizi Starter is below Copilot Pro pricing while also including Claude, Gemini, and open models rather than just the GPT family. For organisations, the larger question is usually seat utilisation, since per-seat AI licences are consistently used heavily by a minority and barely at all by the majority, and you pay for both.

Q: Can I use both?
A: Yes, and for many teams that is the sensible arrangement: Copilot seats for the roles that genuinely live in Office documents, and a general workspace for everyone else. It usually costs less than universal Copilot seats and it avoids the outcome where people without a licence quietly paste company documents into a free consumer tool.

Q: Is this about GitHub Copilot?
A: No, GitHub Copilot is a separate product for code inside your editor, and it is not what this page compares. If that is what you are evaluating, the relevant reading is [AI for coding](https://whizi.io/resources/ai-for-coding) and [ChatGPT alternatives for coding](https://whizi.io/resources/chatgpt-alternatives-for-coding), since editor integration and chat-based reasoning about code are different tools that most developers end up using together.

Q: Does Copilot use the same models as ChatGPT?
A: Copilot is built on the GPT model family, so on many tasks output quality is broadly comparable to other GPT-based products, with differences coming from the surrounding prompts and the context it can reach inside your tenant. The gap is coverage rather than quality: with one model family, there is nothing to switch to on tasks where a different model would be better.

### The best DeepSeek alternative for each reason people leave

URL: https://whizi.io/alternatives/deepseek/
Updated: August 2026

Quick answer: Which DeepSeek alternative you need depends on your complaint: reliability, data location, or text-only models. If the free app has not let you down, you need nothing. Whizi carries DeepSeek models from the Pro plan up without a DeepSeek account, next to GPT, Claude, and Gemini in one conversation.

People leave DeepSeek for reliability, data location, or multimodal work. Whizi is the DeepSeek alternative that keeps the price discipline and adds GPT, Claude, Gemini, and images.

#### First, which DeepSeek are you leaving?

DeepSeek is three products wearing one name: a free chat app, a very cheap hosted API, and open weights anyone can run or host. Most "DeepSeek alternative" searches are about the first one, and they come from three distinct complaints.

**Reliability.** The hosted service has had visible periods of degradation and signup pauses under load. A free product with no guarantees is a fine experiment and a bad dependency.

**Data location.** The app and the hosted API run on DeepSeek's servers, and for plenty of companies and countries that is the entire conversation. The models themselves are open, so this is a complaint about the hosting, not the weights.

**Text only.** DeepSeek's text models are text models. The moment your work involves images, in or out, you need something else, and no prompt fixes that.

Each complaint has a different correct fix, which is why generic alternative lists are useless here.

#### The alternatives, honestly

| Alternative | What it fixes | What it costs you |
| --- | --- | --- |
| **Stay on the free app** | Nothing is broken for casual use | Reliability and data location stay as they are |
| **DeepSeek API** | Reliability improves, cost stays tiny | Same servers, and you are building plumbing yourself |
| **The open weights elsewhere** | Data location, fully | Setup and hosting are on you or a provider you vet |
| **ChatGPT or Claude subscriptions** | Polish, multimodal, reliability | Ten times the cost per answer, or more |
| **Whizi** | DeepSeek models without a DeepSeek account, next to the frontier models | A paid plan, though below any single flagship subscription |

The first row is a real answer. If DeepSeek is your free tool for casual questions and it has not let you down, you do not need an alternative; you need nothing. Switching costs only make sense when one of the three complaints above is actually yours.

#### Keep the discipline, whatever you choose

The genuinely important thing about DeepSeek was never the app. It was the proof that competitive reasoning and coding can cost a tenth of flagship pricing, sometimes less. That number should permanently change how you buy AI.

The discipline that follows: match the model to the job instead of sending everything to the most expensive assistant you pay for. Summaries, extraction, translation, routine drafts, and first-pass code do not need a flagship. Hard reasoning, high-stakes writing, and anything a client sees might.

This only works in practice if switching models is one click rather than a separate app and a separate bill. That is the actual feature to shop for in any DeepSeek alternative: not "which model is best" but "how cheaply can I route the easy 80 percent."

The [AI Model Cost Index](https://whizi.io/tools/model-cost-index) keeps the per-answer numbers current if you want to see how far apart the tiers really are.

#### The case for Whizi specifically

Whizi carries DeepSeek models from the Pro plan up, without a DeepSeek account, next to GPT, Claude, and Gemini in the same conversation. Route the routine work to the cheap model, escalate the hard question to a flagship, and compare their answers when it matters, on one bill.

That covers all three complaints at once: hosting and uptime are Whizi's problem rather than a free service's best effort, you are not creating a DeepSeek account or sending your chats to its app, and when the job needs images or a model with eyes, the multimodal models are in the same picker.

The honest trade: Whizi is a paid product, and if a free chat app has been meeting your entire need, paying anything is a step up in cost. The plan still lands below a single ChatGPT Plus or Claude Pro subscription, which is the comparison that matters once you are paying for AI at all.

Key points:

- The three reasons people actually search for a DeepSeek alternative, and each one has a different right answer
- How to keep the thing DeepSeek got right, which is refusing to overpay per answer
- When you do not need an alternative at all, stated plainly

Checklist:

- Name your actual complaint: reliability, data location, or text-only
- If the free app meets your need, keep it and spend nothing
- Stop sending easy jobs to expensive models, whatever tool you use
- Check the per-answer cost gap in the Model Cost Index before paying for anything
- If data location is the issue, make sure the alternative actually hosts elsewhere
- Test a cheap model against a flagship on your own work, not on benchmarks

FAQ:

Q: Can I use DeepSeek models in Whizi?
A: Yes, from the Pro plan up, without creating a DeepSeek account. The models sit in the same picker as GPT, Claude, and Gemini, so the cheap-model-for-easy-jobs discipline is one click instead of a separate app.

Q: Is DeepSeek good enough that I do not need an alternative?
A: Often, yes. On reasoning and code it is genuinely competitive with far more expensive models. If you use it casually, nothing is broken, and none of the three complaints apply to you, staying on the free app is the correct and cheapest answer.

Q: What about the privacy concerns?
A: Be precise about what the concern is. The app and hosted API run on DeepSeek's servers, which is a hosting question. The weights are open, so the same models served by other providers do not involve DeepSeek's infrastructure at all. If data location is your issue, the fix is choosing where the model is hosted, not avoiding the model.

Q: Is DeepSeek cheaper than using Whizi?
A: The free app costs nothing, so yes, and if it covers your need, use it. The comparison changes once you need reliability, images, or a stronger model for hard tasks, because then you are comparing against a flagship subscription at $20 a month, and a Whizi plan comes in under that with the cheap models still included.

Q: Which model should replace DeepSeek for image work?
A: None of DeepSeek's text models handle images, so this is about adding a capability rather than replacing one. In Whizi the multimodal and image-generation models sit next to the text models, so the answer becomes a model picker choice per message instead of another subscription.

### The best Gemini alternative that still includes Gemini 3 Pro

URL: https://whizi.io/alternatives/gemini/
Updated: August 2026

Quick answer: Most people searching for a Gemini alternative want to keep Gemini and stop paying for the bundle around it. Whizi includes the Gemini family on Pro and above with GPT, Claude, and open models, starting below Gemini Advanced pricing and with no 2TB of storage attached. Stay on Gemini Advanced if you edit inside Docs and Gmail.

Keep Gemini, add GPT and Claude, and drop Google’s single-provider lock-in. Whizi is the Gemini alternative that gives you every model in one workspace.

#### Why people search for a Gemini alternative

Rarely because the model disappoints. Gemini is excellent at the things it is excellent at. The reasons cluster into three.

**The bundle.** Gemini Advanced arrives as part of Google One AI Premium, with 2TB of cloud storage attached. If you would not otherwise buy that storage, a meaningful share of the subscription is paying for capacity you will never fill.

**Coverage.** Gemini is the strongest option for long documents and current information, and it is usually not the strongest for prose a person reads end to end or for strict structured output. Inside Google's product there is nothing to switch to.

**Ecosystem gravity.** Some people simply do not want their documents, email, storage, and AI all held by one company, whether for procurement reasons or preference.

The important pattern: almost nobody wants to stop using Gemini. They want to keep it and stop paying for everything wrapped around it.

#### What Gemini is actually best at

Worth being specific, because the point of an alternative is to keep the strengths.

**Very long documents.** The largest context window of the major models. A 300 page filing, a full contract set, or thirty research papers can be genuinely present in one conversation rather than processed in pieces. Cross-document questions such as "where do these disagree" only work properly when everything is in context at once.

**Current information.** Tight web integration makes it the most reliable choice for anything that happened recently, with links you can open and check.

**Multimodal work.** Images, video, and text together in one conversation is an area where it leads.

And where it usually is not the best pick: prose that a human reads end to end, where Claude needs less editing before it is sendable; and strict structured output, where GPT holds an exact schema more reliably across many rows. Those two categories are most of what people do after the research is finished.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on Gemini Advanced** | Gemini plus Workspace integration and 2TB storage | Your work is inside Docs and Gmail, or you want the storage anyway |
| **Free Gemini** | The free tier | Light use, and you can accept limits on the advanced models |
| **ChatGPT Plus** | GPT as a product | You want a strong general assistant and only one family |
| **Claude Pro** | Claude only | Writing quality is your priority and you rarely need live research |
| **Whizi** | Gemini plus GPT, Claude, and open models, flat pricing | You want Gemini's strengths and somewhere to go for the rest |

The first row deserves fair weight. If your documents live in Google Docs and you want AI in the sidebar while you edit, no external workspace replicates that, and it is the single best reason to stay.

#### The storage question

This is the whole value calculation for a lot of people, so it is worth answering directly rather than skipping past it.

Ask yourself whether you would buy 2TB of Google storage if no AI were attached. If you are already on a paid Google One tier, or you have a large photo library, or a household depends on Drive, then the answer is yes and the AI is arriving at a genuine discount. Gemini Advanced is good value in that case and this page is not going to pretend otherwise.

If you are on the free tier using a fraction of the allowance, you are paying a subscription price where part of it buys storage you will not use, in exchange for one model family.

#### The case for Whizi specifically

Whizi includes the Gemini model family on Pro and above, alongside GPT, Claude, and leading open models, starting below Gemini Advanced pricing with no storage bundled in.

What that changes in practice is the second half of every task. Read the 200 page report with Gemini, extract the figures into a table with GPT, draft the summary with Claude, then ask a fourth question of a model that wrote none of it. One thread, no re-uploading, no re-explaining. That sequence is not possible inside any single-provider product at any price.

The honest trade-offs: no sidebar integration in Docs or Gmail, no access to files already in your Drive without uploading them, and no bundled storage. If most of your day is editing documents in place inside Google's tools, those losses are real and Gemini Advanced may be the better answer for you. Full head-to-head at [Whizi vs Gemini Advanced](https://whizi.io/vs/whizi-vs-gemini-advanced).

#### Testing it in a week

1. Log every AI task for a week and mark whether it happened inside a Google app or in a separate tab.
2. Of the separate-tab tasks, mark the writing and structured extraction ones. Those are where a different model would probably have done better.
3. Run three of them with Claude and GPT and compare against what Gemini gave you.
4. Check your actual Google storage usage and decide whether you would pay for 2TB alone.
5. Try the long-document sequence end to end in one thread: read, extract, draft, critique.

The result is usually unambiguous. If you rarely leave Docs and Gmail, the integration is worth the bundle. If you work in a browser tab anyway, you are paying for an integration you do not touch and storage you do not need, while missing two model families you would use daily.

Key points:

- Most people searching this want to keep Gemini and stop paying for the bundle around it
- What Gemini is genuinely best at, and the tasks where you should be reaching for something else
- The storage question, answered honestly, because it decides the value calculation

Checklist:

- Decide whether you would pay for 2TB of storage with no AI attached
- Track for a week whether your AI work happens inside Google apps or outside
- Identify the writing and extraction tasks where Gemini is not the strongest choice
- Test three of those against Claude and GPT
- Try one long document end to end: read, extract, draft, critique, in a single thread
- Check whether your writing lives in Google Docs, since that is the real lock-in

FAQ:

Q: Is Gemini in Whizi the same model as in Gemini Advanced?
A: It is the same Google model family, accessed through the provider API, so capability on long documents and research is equivalent. What differs is everything around it: Gemini Advanced adds Workspace integration and bundled storage, while Whizi puts the same models next to GPT and Claude with switching, comparison, and projects that keep each body of work together.

Q: Can I still do long-document work without Gemini Advanced?
A: Yes, that capability comes from the model rather than from the subscription, and the Gemini family in Whizi carries the same large context window. The practical gain is what happens next: handing the same document to Claude for the write-up or GPT for structured extraction without re-uploading, which is usually the second half of any long-document task.

Q: Does Whizi work with Google Docs and Drive?
A: Not natively. You upload documents or paste content rather than editing in place, which is a small friction when the document is an input to research and a larger one when the document is the artefact you are working on. Sidebar integration inside Docs and Gmail is specific to Google’s product and is the strongest reason to keep Gemini Advanced.

Q: Is Whizi cheaper than Gemini Advanced?
A: Whizi Starter is below Gemini Advanced pricing and includes GPT and Claude alongside Gemini, so on models per dollar it is clearly ahead. The fair caveat is the bundled 2TB of Google One storage: if you would buy that anyway, Gemini Advanced represents good value and the comparison is much closer than a model list suggests.

Q: What about the free Gemini tier?
A: It is a legitimate option and worth saying plainly. Free tiers across providers cover a lot if your usage is light, and the honest failure mode is hitting limits on the advanced models during the week you are busiest. A paid plan earns its cost at the point where being interrupted mid-task is more expensive than the subscription.

### The best Grok alternative when X is not the point

URL: https://whizi.io/alternatives/grok/
Updated: August 2026

Quick answer: Whether you need a Grok alternative depends on which part you use. Nothing replaces live X data, so stay if that is the job. If you subscribe for the reasoning and the direct tone, Whizi carries Grok models on its higher tiers next to GPT, Claude, and Gemini in one conversation.

Grok bundles live X data, a direct personality, and strong reasoning models. Whizi is the Grok alternative that gives you the models next to GPT, Claude, and Gemini in one plan.

#### What a Grok subscription is actually selling

Three different things, and most subscribers only use one of them.

**Live access to X.** Grok can pull from posts as they happen, not from articles indexed hours later. For anything moving in real time, that is a capability the other assistants structurally do not have.

**A less hedged personality.** Grok is tuned to refuse less, commit to an angle, and skip the throat-clearing. For satire, punchy drafts, and questions where you want an opinion rather than a survey, that is a real feature.

**Strong reasoning models.** Recent Grok releases have posted strong numbers on the harder reasoning and math benchmarks. This part has nothing to do with X at all; it is a frontier model family like the others.

The alternative question splits the same way. If you subscribe for the live X firehose, the honest answer is that nothing else replaces it. If you subscribe for the second or third thing, you have options, because the reasoning and much of the tone travel anywhere the models are served.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on Grok or X Premium+** | The native product, wired into X | Live X data is the job, or you were paying for X Premium+ anyway |
| **Free Grok** | The free tier with limits | Occasional use, mostly curiosity |
| **ChatGPT Plus** | One general assistant, better tooling | You want polish, files, and automations, and only need one model family |
| **Whizi** | Grok models next to GPT, Claude, and Gemini | You want the reasoning without another subscription, and a second opinion beside it |

The first row is a real answer. Grok is sold in overlapping ways: free with limits, a dedicated subscription at around $20 a month with a heavier tier above it, and bundled into X Premium+, which makes it effectively free for someone who was subscribing to X anyway. If that last case is you, keep it; a bundled product you already pay for beats any alternative on price.

#### What carries over, and what stays on X

Using Grok models through a workspace gets you the reasoning, the speed, and most of the voice. It does not get you the native X integration; the live firehose belongs to the native product, full stop. Anyone selling a Grok alternative who skips that sentence is not being straight with you.

The personality travels better than people expect, because most of it can be asked for. If what you liked was the directness, this prompt gets it from any strong model:

`Answer directly. Commit to a position and defend it. No balanced-overview framing, no "it depends" unless it genuinely does, and if it does, say on what. One caveat maximum, at the end.`

The part of Grok's tone you should not carry over is the part that has caused its public incidents. Grok has shipped answers that were offensive or badly wrong in ways the other assistants filter out. Fewer refusals is an advantage right up until it appears in a deliverable you did not proofread.

#### The case for Whizi specifically

Whizi carries Grok models on its higher tiers alongside GPT, Claude, and Gemini, which changes how you use them. Grok's reasoning becomes one opinion among several instead of the only voice in the room: run the hard question through Grok, then ask Claude what the answer missed, in the same thread.

The arithmetic is the other half. A Grok subscription and ChatGPT Plus on the same card is around $40 a month for two assistants that cannot see each other's answers. One Whizi plan is less than either, and the models share a conversation. If your subscriptions have quietly stacked, run the [AI subscription savings calculator](https://whizi.io/tools/calculator) with your actual list.

The honest trade: no live X data, and the Grok models sit behind Whizi's higher tiers rather than the entry plan. Full head-to-head at [Whizi vs Grok](https://whizi.io/vs/whizi-vs-grok).

#### The test before you switch anything

Open your Grok history and look at your last twenty questions. Count how many actually needed live X data: breaking news, sentiment right now, what people are saying about a launch this hour.

If that count is high, stay. The firehose is the product, and no alternative has it.

If that count is nearly zero, you have been paying for a frontier model with a personality, and the question becomes which workspace serves it best alongside the other models you already use. That is the case Whizi is built for. When you decide to stop the billing, the steps are in [how to cancel SuperGrok](https://whizi.io/resources/how-to-cancel-supergrok).

Key points:

- Grok is really three products in one subscription; the alternative depends on which one you actually use
- What carries over to a workspace and what stays on X, stated plainly
- The arithmetic when a Grok subscription lands on the same card as ChatGPT Plus

Checklist:

- Count how many of your last twenty Grok questions needed live X data
- If you pay for X Premium+ anyway, treat Grok as free and keep it
- Carry the directness over with a prompt rather than a subscription
- Proofread anything with attitude before it reaches a client
- Add up what Grok plus ChatGPT Plus actually costs per year
- Test Grok next to Claude and GPT on one real task before deciding

FAQ:

Q: Can I use Grok models in Whizi?
A: Yes, on the higher tiers, next to GPT, Claude, and Gemini in the same conversation. You get the model's reasoning and most of its voice. You do not get the native X integration; live X data stays with the native product, and it is the one thing no alternative replaces.

Q: Is Grok free with X Premium+?
A: It is bundled, so if you already pay for X Premium+ the marginal cost of Grok is zero, and keeping it is the rational move. The alternative question really applies to people considering a dedicated Grok subscription on top of what they already pay for other assistants.

Q: Should I cancel my Grok subscription?
A: Do the twenty-question test first. If your history is mostly live X questions, keep it. If it is mostly general reasoning, writing, and coding, you are paying a dedicated subscription for something a multi-model workspace covers for less, with other frontier models beside it.

Q: Is Grok reliable enough for client work?
A: The models are strong, but Grok is deliberately less hedged and has had public incidents where answers were offensive or badly wrong. For client-facing work, either proofread everything with extra care or draft with a more conservative model and use Grok for the thinking rather than the final wording.

Q: What is the cheapest way to try Grok next to other models?
A: A workspace that carries several model families under one plan. Run the same real task through Grok, GPT, and Claude, compare the answers side by side, and decide based on your work rather than on benchmarks. Whizi starts at $0.99 for the first trial.

### The best Le Chat alternative (and the one reason to stay)

URL: https://whizi.io/alternatives/le-chat/
Updated: August 2026

Quick answer: The best Le Chat alternative depends on your constraint. If EU data residency is binding, stay on Le Chat, because nothing here replaces European hosting. If price or habit sent you looking, Whizi keeps Mistral models in the same picker as GPT, Claude, and Gemini, on an entry plan below the $20 standard.

Le Chat is fast, cheap, and European. Whizi is the Le Chat alternative that keeps Mistral models and adds GPT, Claude, and Gemini when the ceiling starts to show.

#### What Le Chat actually gets right

Le Chat deserves a fairer hearing than most alternatives pages give it. It is genuinely fast, its free tier is generous, its paid tier undercuts the $20 standard, and Mistral is the one frontier-adjacent lab whose models and hosting are European. For individuals and companies with EU data-residency requirements, that last point is not a nice-to-have; it is the requirement.

Mistral's models are also better than their public reputation. The flagship line holds its own on everyday work, the coding line is respected, and the efficiency focus means you rarely wait. None of the reasons to look elsewhere are about Le Chat being bad.

The reasons are narrower, and they are worth naming precisely before paying for anything else.

#### Why people look for an alternative

**The ceiling.** Le Chat serves one lab's models. On the hardest reasoning, the most careful long-form writing, and the deepest research tasks, the frontier trio of GPT, Claude, and Gemini still sets the bar, and no amount of speed closes that gap on the tasks where it matters.

**No second opinion.** The most reliable everyday quality technique, running the same hard question through two model families and comparing, is structurally unavailable inside a single-lab product. This is not Le Chat's fault; it is true of ChatGPT and Claude's own apps too.

**Ecosystem depth.** Integrations, file workflows, and the surrounding tooling are thinner than what has grown around the bigger apps. For light use it never shows; for daily professional use it does.

| Alternative | What it fixes | What it costs you |
| --- | --- | --- |
| **Stay on Le Chat** | Nothing; the price and residency stay unbeatable | The single-lab ceiling stays too |
| **ChatGPT or Claude** | Frontier ceiling | Another $20 subscription, still one lab |
| **Whizi** | Mistral models plus GPT, Claude, and Gemini in one picker | EU-only hosting is not the pitch; check residency needs first |

#### The case for Whizi specifically

Whizi keeps Mistral in the room. The models are in the same picker as GPT, Claude, Gemini, and the cheap open-weights lines, so the routing habit Le Chat users already have, fast cheap model for the everyday, escalate when it matters, works unchanged. The difference is that the escalation path now reaches the frontier.

The cost logic holds at both ends: Whizi's entry plan sits below the $20 standard, and the cost callout above shows the per-answer spread between the labs. You stop paying a premium for variety you cannot access.

The honest trade, stated plainly: Whizi is not a European-hosted product, and if EU data residency is your binding constraint, Le Chat remains the correct answer and you should stay. This page is for everyone whose constraint was price or habit rather than jurisdiction.

#### The test before you switch

Take the three hardest tasks you gave Le Chat last month: the document that came back flat, the analysis that missed the point, the code that needed a second pass.

Run them through a frontier model next to a Mistral model with identical prompts. If the outputs are indistinguishable, your work fits inside Le Chat's ceiling and the cheapest correct answer is to stay. If the gap is obvious, you have your answer, and the question becomes whether you want the frontier as a second subscription or in the same picker as the models you already like.

Key points:

- Le Chat's real pitch is speed, price, and European hosting, and all three are genuine
- The reason people leave is the ceiling: one lab's models, without the frontier trio beside them
- If EU data residency is a hard requirement, the honest answer is to stay, and this page says so

Checklist:

- Name your actual constraint: price, ceiling, or EU residency
- If residency is binding, stay on Le Chat; nothing here overrides that
- Run your three hardest recent tasks through a frontier model before deciding
- Keep the fast-model habit; change only where the escalation goes
- Price any second subscription annually before adding it

FAQ:

Q: Can I use Mistral models in Whizi?
A: Yes, Mistral models are in the Whizi picker alongside GPT, Claude, Gemini, and the open-weights lines, metered per message like everything else. You keep the speed-and-price habit and gain the frontier escalation path in the same conversation.

Q: Is Le Chat as good as ChatGPT?
A: For everyday speed and value, closer than most people assume. On the hardest reasoning and most careful writing, the frontier models are still ahead, and that gap is exactly what a three-task test on your own work will show you in an afternoon.

Q: What is the best European alternative to ChatGPT?
A: Le Chat is the serious answer: European lab, European hosting, competitive pricing. If your requirement is jurisdiction, start and probably end there. If your requirement is capability breadth, that is the one thing a single-lab product cannot offer, whatever the geography.

Q: Is Le Chat Pro worth it?
A: At a price below the $20 standard, yes for anyone who likes the models and lives inside its ceiling. The moment you find yourself pasting Le Chat output into another AI for a second opinion, the arithmetic favors one workspace that holds both models instead.

### The best OpenRouter alternative if you are not building an app

URL: https://whizi.io/alternatives/openrouter/
Updated: August 2026

Quick answer: An OpenRouter alternative is worth looking for only when you are not building software. For a person doing work rather than shipping code, Whizi packages the multi-provider idea as a product: GPT, Claude, Gemini, and open models, with file uploads, projects, and side-by-side comparison, on a flat plan and no API key.

OpenRouter is great for developers routing API calls. Whizi is the OpenRouter alternative for people who just want a polished multi-model chat with Claude, GPT, and Gemini.

#### When a gateway stops being the right tool

OpenRouter is very good at what it is for: one endpoint and one key routing to many providers, with per-token pricing and failover. If you are writing code that calls a model, it remains a sensible choice and nothing here argues otherwise.

People look for an alternative in three situations, and they are all versions of the same realisation.

**You are not building anything.** You wanted access to several models and ended up with an API key, a credit balance, and a basic test interface. The gateway was designed for software to call, and you are a person trying to do work in it.

**Your colleagues need access.** The moment someone in marketing, research, or support needs the same models, an API key is the wrong shape. What follows is usually an internal chat UI that one engineer builds and nobody maintains.

**The metering is affecting how you work.** You trim context to save tokens, skip the comparison you should have run, avoid re-running a prompt with better instructions. Each decision is small and rational, and together they mean cost is quietly setting your quality bar.

#### What you would have to build

The honest way to compare a gateway with a product is to list what the gateway leaves to you, because all of it is buildable and none of it is free.

| Capability | With a gateway | With a finished product |
| --- | --- | --- |
| Chat interface with history | You build and host it | Included |
| File upload, parsing, and OCR | You build the pipeline | Included |
| Long-document handling and chunking | You build the strategy | Handled, with a large-context model |
| Image generation in the same place | Separate integration | Included on Pro and above |
| Projects with pinned files and instructions | You build storage and context assembly | Included |
| Side-by-side model comparison | You build the parallel calls and the UI | Included on Powerhouse |
| Cost predictability | You monitor and cap | Flat plan |

None of that is hard individually. Together it is a product, and it is a product that needs maintaining every time a provider changes an API. If you were building it anyway, the gateway is the right foundation. If you just wanted it to exist, you are pricing weeks of engineering against a subscription.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on OpenRouter** | API gateway, per-token | You are shipping software, or your usage is very light |
| **Direct provider APIs** | One integration each | You use one provider and want the lowest cost per token |
| **ChatGPT Plus, Claude Pro, Gemini Advanced** | One provider each, as a product | You have settled on exactly one model family |
| **Poe** | Bot marketplace, points pricing | Breadth and exploration matter more than workflow |
| **Whizi** | Multi-model workspace, flat pricing | You want the multi-provider access as a finished product |

The first two rows are genuinely correct answers for a lot of people. If your monthly token spend is a few dollars, no subscription will beat that, and saying so is more useful than pretending otherwise.

#### Where per-token pricing flips

The crossover arrives sooner than most people expect, because the operations that make AI genuinely useful are the expensive ones.

Long documents in context are expensive. Image generation is expensive. Running one prompt through two models doubles the cost of the thing you most want to do. And long conversations re-send their entire history with every message, so a thread that has been running all afternoon costs more per message than the one you started this morning.

A single day of serious document work can cost more in tokens than a month of a flat plan. Meanwhile the behaviour a meter encourages, shorter context and fewer attempts, is exactly the behaviour that produces worse output.

For the wider picture on where AI spending goes, see [AI subscription costs](https://whizi.io/resources/ai-subscription-costs).

#### The case for Whizi specifically

Whizi is the multi-provider idea packaged as a product: GPT, Claude, Gemini, and leading open models, with model switching mid-thread, file uploads that handle long documents properly, image generation, projects, and side-by-side comparison. No key, no credit balance, no integration work, and pricing that does not move when you have a heavy week.

The honest limits: it is not an API, so if you need programmatic access at scale, OpenRouter or a direct provider API is still the right tool. The catalogue is smaller, covering the major families rather than the long tail. And for genuinely light usage, per-token billing will be cheaper.

The full comparison is at [Whizi vs OpenRouter](https://whizi.io/vs/whizi-vs-openrouter).

#### Running both, which is usually correct

This is not an either-or for most teams, and the split is by role rather than preference.

Keep the gateway for what you ship: the production application, the background jobs, the data pipeline. Those need per-request control and volume-based pricing.

Use a product for what your team does: writing, research, analysis, document work, and daily model comparison. Those people should not be managing keys, and the cost of building them an interface is far higher than a subscription.

The benefit beyond convenience is that engineering usage stays lean and observable instead of being mixed in with everyone's ad-hoc chatting, which is what makes an API bill impossible to reason about.

Key points:

- The three situations where a gateway stops being the right tool
- Where per-token pricing genuinely wins, and the usage level where it flips
- What you would have to build to match a finished product, priced honestly

Checklist:

- Answer the first question honestly: are you building software or doing work?
- Estimate real monthly token spend including long documents, images, and comparisons
- List what you would have to build to match a finished product, and price that time
- Note every time you trimmed context or skipped a comparison to save cost
- Check whether non-developers on your team need access
- Consider the split: gateway for production, product for the team

FAQ:

Q: Can I use OpenRouter and Whizi together?
A: Yes, and for most teams past a handful of people that is the correct setup. The gateway handles what you ship, meaning production applications and background jobs that need per-request control and volume pricing. The product handles what your team does day to day. It also keeps engineering usage separate from everyone’s ad-hoc chatting, which is what makes an API bill readable.

Q: Is Whizi cheaper than OpenRouter for heavy users?
A: For a single person doing regular work, almost always, because flat pricing amortises exactly the expensive operations: long documents in context, image generation, model comparison, and long threads that re-send their history each message. For light usage, per-token billing is cheaper and no subscription will beat a few dollars a month.

Q: Does Whizi expose an API?
A: Whizi is a product experience rather than an infrastructure layer, so for programmatic access at scale, OpenRouter or a direct provider API remains the right tool. That is the distinction between the two rather than a limitation: one is built for software to call, the other is built for a person to work in.

Q: Which has more models?
A: OpenRouter, considerably, since aggregating a very large catalogue including niche and experimental models is what a gateway is for. Whizi carries the major first-party families plus leading open models and image generation, which covers what most professional work needs. If a specific long-tail model matters to you, the gateway is the correct choice.

Q: Can non-technical colleagues use OpenRouter?
A: There is a basic chat for developers to verify routing, but it is not built as a daily workspace and does not aim to provide document workflows, templates, or team features. Pointing non-technical colleagues at an API gateway usually ends with an internal chat interface that one engineer builds, everyone depends on, and nobody has time to maintain.

### The best Perplexity alternative for real work (not just search)

URL: https://whizi.io/alternatives/perplexity/
Updated: August 2026

Quick answer: The best Perplexity alternative depends on whether your research ends in an answer or in a document. For pure cited search, staying is sensible. Whizi covers the research with web-connected models and then the writing, so Claude drafts the memo in the same thread where the sources are still visible.

Perplexity is great for answers with citations. Whizi is the Perplexity alternative that gives you the same research flow plus Claude, GPT, and Gemini for writing and coding.

#### Two kinds of research, and which one you do

**Answering a question.** You want a fact with a source you can check. What happened, who owns this, what does this term mean, what is the current state of play. Perplexity is built end to end for this loop and does it better than a general chat.

**Building an understanding.** You are reading fifteen sources, extracting comparable facts from each, noticing where they contradict one another, and then writing something that takes a position. This is less like search and more like reading a stack of documents in one place, and it ends in a deliverable.

Most professional research is the second kind wearing the clothes of the first. You start with a question and finish needing a memo. If that describes your week, a search product covers the first half and hands you off for the second.

The test takes two minutes: look at your last ten searches and write down what you did in the ten minutes after each. If the answer is mostly "opened another tool and started writing", you are paying for a handoff.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on Perplexity** | Cited search product | Your AI use is mostly answering questions |
| **Free Perplexity** | The free tier | Occasional lookups |
| **Gemini Advanced** | Google's plan, strong web integration | You want live research inside the Google ecosystem |
| **ChatGPT Plus with browsing** | GPT plus web access | You want one general assistant and only GPT |
| **Whizi** | Cited research plus Claude, GPT, and Gemini | Research usually ends in something you write |

The first row is a real answer. The query-to-answer loop, the follow-up interaction, and the citation presentation in Perplexity are better designed for pure search than a general chat is, and if that is the whole job, staying is sensible.

#### Keeping the citation discipline elsewhere

The habit worth carrying over is demanding sources. It transfers if you make the requirement explicit in the prompt rather than relying on the product to enforce it.

**The research prompt**

`Research [question]. Return findings with a source URL and publication date for each. Where sources disagree, show both rather than reconciling them. Mark anything you cannot source as UNVERIFIED. Do not fill gaps with plausible values.`

**Chase the original**

`For the claim that [specific finding], find the original source rather than an article citing it, and tell me what the original actually measured, over what period, and for what population.`

That second prompt is the one that separates real research from repetition. Statistics degrade as they are passed along, and a figure three citations deep is frequently unrecognisable at the origin.

**The verification pass**

`List every factual claim in the answer above as a table: the claim, whether it came from a source in this conversation or from general knowledge, and the URL. Mark everything from general knowledge as UNVERIFIED.`

Then open the sources yourself. This is not optional in either product. The most common error is not an invented statistic, it is a real one attached to the wrong claim, and no interface catches that for you.

#### The case for Whizi specifically

Whizi covers the research half with web-connected models and then covers the half that follows. Draft the memo in Claude, extract the comparison table in GPT, read the 200 page source document with Gemini, all in the same thread where the research is still visible.

The specific gain is that the drafting model can see the research rather than a summary of it. Which sources disagreed, the caveat attached to the headline number, the figure you decided not to use. That context is what makes a draft as nuanced as your understanding, and it does not survive a tab switch.

It also enables a check that requires both halves in one place: `Check this draft against the sources above. Flag every claim stronger than what the source supports, and every number I have carried over incorrectly.`

The honest trade: the citation presentation is less polished, there are no Focus modes or Collections as named features, and for a quick factual lookup Perplexity is a nicer experience. Full head-to-head at [Whizi vs Perplexity Pro](https://whizi.io/vs/whizi-vs-perplexity-pro).

#### The rule that matters more than the tool

Whichever you choose, the failure mode is the same and it is worth stating bluntly. Models fabricate citations. They produce plausible author names, real-sounding publications, and DOIs that resolve to nothing or to an unrelated paper. A cited interface reduces this substantially but does not eliminate it, and it does nothing at all about the harder error, which is a genuine source attached to a claim it does not support.

Three rules that survive any product change: never cite something you have not opened; resolve every DOI; and check that the source says what the citation claims it says. That last one catches more real errors than the other two combined.

Key points:

- The distinction that decides this: research that answers a question, or research that produces a document
- How to keep the citation discipline without the dedicated product, with the prompts written out
- Where Perplexity remains better, stated plainly

Checklist:

- Look at your last ten searches and note what you did immediately after each
- Count how many ended in a document rather than in the answer itself
- Put the source requirement in the prompt rather than expecting the tool to enforce it
- Chase the original source rather than an article citing it
- Draft in the same thread as the research so the nuance carries over
- Run the check-the-draft-against-the-sources prompt before sending anything
- Open every source, and verify it supports the claim attached to it

FAQ:

Q: Can I get cited answers in Whizi?
A: Yes, using web-connected models, though you ask for it explicitly rather than getting it by default. Put the requirement in the prompt: a URL and date for every finding, both figures where sources disagree, and anything unverifiable marked as such. The presentation is less polished than a dedicated search product, and the verification work is identical either way, since you should be opening sources regardless.

Q: Should I cancel Perplexity Pro?
A: If your research usually ends in something you write, yes, because the handoff between two tools loses the reasoning rather than just the time. If you use it purely as a cited search engine and value the query-answer loop, keeping it is reasonable. Worth doing the arithmetic though: two subscriptions at around $20 each exceeds one plan that covers research, writing, coding, and images.

Q: Is Whizi better for academic research?
A: For the stage after finding sources, yes. Uploading the papers themselves and working across them, building an evidence table with one row per study, extracting into a consistent schema, and drafting with a model that preserves academic hedging are all things a search product does not aim to do. Neither tool removes the obligation to open every source and resolve every DOI, which is where the real risk sits.

Q: Does Whizi have Focus modes or Collections?
A: Not as named features. The equivalent is a reusable prompt: one you keep and paste, that constrains sources, fixes the output format, and encodes the standards you want applied every time. It takes a little more setup than clicking a mode, and in exchange it is fully editable and runs against any model rather than being fixed by the product.

Q: What about the free Perplexity tier?
A: It is a reasonable option for occasional lookups and worth keeping in mind before paying for anything. The limits show up on heavier use and on the stronger models. If your searches are occasional and rarely lead to a document, the free tier plus whatever you already use for writing may be the cheapest correct answer.

### The best Poe alternative for actual work (not just bot-hopping)

URL: https://whizi.io/alternatives/poe/
Updated: August 2026

Quick answer: The best Poe alternative depends on whether you are buying a marketplace or a workspace. If breadth and character chat are why you subscribe, stay. If you use the same two or three models on real work, Whizi gives you GPT, Claude, and Gemini in one thread on flat pricing, with no points wallet.

Poe is a bot marketplace. Whizi is the Poe alternative built for professionals: Claude, GPT, Gemini, and image models in one workspace with real pricing clarity.

#### Why people leave Poe

Three reasons come up repeatedly, and they are all about the same underlying thing: the product is optimised for discovery, and most people stop discovering after a few weeks.

**The points wallet.** Compute points make every message a small decision. Use the expensive model or save the allowance? Run the comparison or trust the first answer? None of those decisions is individually significant, and collectively they mean you are rationing quality on your own work. That cost never appears anywhere you can see it.

**Bot fatigue.** The catalogue is enormous, and most community bots are a system prompt in a costume. After the novelty passes, most people find they use the same two or three, and they are paying for a marketplace they have stopped browsing.

**It is not built around finishing something.** Poe is organised as a list of bots. Work is organised as a task with a beginning and an end, usually involving a document, several models, and an output. Those are different shapes.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on Poe** | Bot marketplace, points pricing | You genuinely explore, or want niche and experimental models |
| **ChatGPT Plus, Claude Pro, or Gemini Advanced** | One provider each | You have settled on exactly one model family |
| **OpenRouter** | API gateway, per-token | You are building software, or your usage is very light |
| **Free tiers** | Free plans from each provider | Light use, tolerant of rate limits |
| **Whizi** | Multi-model workspace, flat pricing | You use two or three models regularly and want them in one thread |

The first row is real. If breadth is why you are on Poe, no focused workspace will match it, and character or persona chat in particular is a first-class use case there and not one Whizi is built for.

ChatHub and Merlin sit in the same category and are compared alongside these in [apps like Poe](https://whizi.io/alternatives/apps-like-poe).

#### What to look for in a replacement

- **Predictable pricing.** If you are leaving because of the wallet, do not replace it with per-token billing, which has the same effect for the same reason.
- **The models you actually use.** Check your own history first. If it is GPT, Claude, and Gemini, a focused workspace covers you. If it includes something exotic, verify before switching.
- **Portable prompts.** A prompt library that runs across models survives model releases. One tied to a specific bot or provider does not.
- **A thread that holds a task.** File uploads, model switching mid-conversation, and history that persists across the whole piece of work.
- **Comparison built in.** The ability to run one prompt through two models is the fastest quality check available, and it is only free of guilt when the pricing is flat.

#### The case for Whizi specifically

Flat monthly pricing, so the meter stops influencing your choices. First-party GPT, Claude, and Gemini plus leading open models. Image generation inside the chat rather than as a separate bot. Projects that keep each body of work, its files and its standing instructions together. Side-by-side comparison on Powerhouse.

The design difference is that a thread is built around a task rather than around a bot. Upload the document, read it with the large-context model, extract with GPT, draft with Claude, and have a fourth model critique the result, all in one conversation that keeps everything.

And the honest trade: a far smaller catalogue, no community bots, no personas, no store. If those are what you value, this is the wrong product and you should stay. The full head-to-head is at [Whizi vs Poe](https://whizi.io/vs/whizi-vs-poe).

#### Migrating your bots

The thing worth carrying over is the prompts, not the chats. For each bot you use regularly:

1. Open it and read its description and instructions. If you built it, you already have the system prompt.
2. Keep that instruction text as a reusable prompt, or put it in a project's instructions so it applies to every chat in that project automatically.
3. Mark what changes per run, for example the topic or the draft, so the reusable part stays obvious.
4. Add what most community bots omit: an explicit output format, and an instruction for what to do with input that does not fit.
5. Run it against two models and see which wins. Bots are pinned to whichever model their author chose, often a while ago.

That final step is the actual upgrade rather than a chore. Your prompt library stops being tied to one model, which means the next model release improves your existing work instead of requiring someone to rebuild a bot.

#### How to decide with a week of data

1. For one week, note which bot answered each message. If three names cover most of it, the marketplace is not what you are paying for.
2. Note where your points go. Points on the major models are exactly what a flat plan covers; points spread across many exotic models are a reason to stay.
3. Write down every time you chose a cheaper model to preserve the wallet. That number is the real cost of metered pricing.
4. Rebuild your top two bots as reusable prompts and run a real task end to end elsewhere.
5. Compare on editing effort and on whether you hesitated before a second attempt.

Key points:

- The compute-points problem, and why it costs you output quality rather than money
- How to tell whether you are buying a marketplace or a workspace
- A migration path that turns your favourite bots into prompts you own

Checklist:

- Track which bots you actually use for a week, not which you have saved
- Check whether your points go to major models or exotic ones
- Count how often you downgraded a model to preserve points
- Rebuild your top two bots as reusable prompts that run on any model
- Verify any niche model you depend on is available before switching
- Run one real task end to end in both products
- Compare on editing effort rather than on catalogue size

FAQ:

Q: Does Whizi have a bot marketplace?
A: No, and it is a deliberate choice rather than a gap. Whizi focuses on first-party models, and the bot equivalent is a reusable prompt: an instruction you can read, edit and run against any model, whereas a bot is fixed to whichever model its author selected, which ages as models are released and retired. For standing instructions, a project applies them to every chat inside it automatically.

Q: Is Whizi cheaper than Poe?
A: For consistent daily use, generally yes, since flat pricing amortises heavy use while a points wallet charges most for exactly the expensive models you want on important work. For light or sporadic use, a points allowance can be cheaper. The cost people miss with points is behavioural: downgrading the model or skipping a second attempt to protect the wallet shows up as weaker output rather than as a number.

Q: Can I recreate my favorite Poe bots in Whizi?
A: Yes, and it usually takes a few minutes each. Copy the bot instructions into a template, add variables for what changes, and specify the output format explicitly. The rebuilt version is better in one concrete way: it runs against any model, so when a newer model turns out to be stronger for that task you switch immediately rather than waiting for someone to update a bot.

Q: Does Whizi have as many models as Poe?
A: No. Poe carries considerably more in total, including experimental and niche models plus thousands of community bots, and that breadth is genuinely its advantage. Whizi carries the major first-party families, leading open models, and image generation, which covers what most professional work needs. If a specific long-tail model is essential to you, check availability before switching.

Q: What about character and persona chat?
A: Poe is the better product for that and it is not close. Character chat is a first-class use case there, with a large community building for it, and Whizi is built for professional work rather than for personas. If that is a significant part of why you subscribe, this is not the right replacement.

### The best TypingMind alternative when the keys become the job

URL: https://whizi.io/alternatives/typingmind/
Updated: August 2026

Quick answer: You do not need a TypingMind alternative for price: your own API keys at list rates win per token for heavy technical use. The switch case is operations. Whizi drops the keys, top-ups, and per-provider billing, bundling GPT, Claude, Gemini, Mistral, and DeepSeek into one flat plan that starts below $16 a month.

TypingMind plus your own API keys is the cheapest way to chat with frontier models. Whizi is the TypingMind alternative for when managing keys, top-ups, and rate limits stops being fun.

#### Credit where it is due

If you run TypingMind with your own API keys, you already understand the economics better than most people writing about AI. You pay provider list rates with no middleman margin, you picked your interface once, and per token you are running one of the cheapest serious AI setups that exists. On raw cost per answer for a heavy technical user, that setup wins, and this page is not going to pretend otherwise.

The reasons people who love that setup still end up leaving it are operational, not economic. They tend to show up between six months and two years in.

#### Why BYOK users eventually look around

**The keys become a part-time job.** One provider account per lab, separate billing dashboards, separate top-ups, separate rate limits, separate status pages. Each is trivial; together they are recurring admin that scales with every new model family you want to try.

**Spend becomes invisible.** Provider billing is per token across several dashboards, so the monthly total is a research project, and one runaway agent loop or oversized context can spike a bill silently. The flat-rate world has the opposite property: predictable, boring, and capped.

**Nobody else in your life will do this.** The moment you want a partner, a teammate, or a client to use the same setup, BYOK collapses. They will not create an OpenAI account, fund an Anthropic key, and paste both into a settings screen. Every BYOK user becomes tech support for their own household.

**The rates you pay are retail.** Your keys buy list price. Bundled products buy wholesale and smooth it across users, which is how a flat plan can undercut what a moderate user pays across three provider accounts, even before counting the admin time.

#### The alternatives, honestly

| Alternative | What it is | Best when |
| --- | --- | --- |
| **Stay on TypingMind** | Your interface, your keys, list rates | Heavy technical use, you enjoy the control, nobody else needs access |
| **OpenRouter** | One key, every model, still per token | You are building software rather than chatting |
| **Poe** | Big catalogue on a points system | You want variety and creator bots more than predictability |
| **Whizi** | Flat plan, no keys, GPT + Claude + Gemini + open models | You want the multi-model workflow without the operations |

The full head-to-head, including where each option is cheapest, is in the [Poe vs TypingMind vs Whizi comparison](https://whizi.io/resources/poe-vs-typingmind-vs-whizi). The one-line version: TypingMind wins on cost per token for heavy technical users, OpenRouter wins for builders, and the bundled workspace wins the moment predictability or other people enter the picture.

#### The case for Whizi specifically

Whizi is the same core workflow you already run, one thread, model picker per message, minus everything around it: no keys, no per-provider accounts, no top-ups, no surprise bills. GPT, Claude, Gemini, Mistral, DeepSeek, and the other open lines are metered in credits on one flat plan, with the per-message costs published on the [Model Cost Index](https://whizi.io/tools/model-cost-index).

For the spreadsheet you are already opening: take your last three months of provider bills, add them up, and put the total next to a flat plan that starts below $16 a month. Heavy API users will still beat that with raw keys. Moderate users usually find the bundle is already cheaper before their admin time costs a cent.

The honest trade: you give up list-rate economics at the extreme high end, granular parameter control, and the tinkering itself, which for some people is the hobby. If that is you, stay, and enjoy it.

Key points:

- TypingMind users are the most model-literate buyers there are; this page respects that
- Bring-your-own-keys is genuinely cheaper per token, and that stays true; the question is what your admin time costs
- The switch case is not price, it is operations: keys, top-ups, limits, and teammates who will never make an API account

Checklist:

- Add up your last three months of provider bills across every key you fund
- Count the minutes spent on top-ups, limits, and key management in the same period
- Price what happens when a teammate or partner needs the same access
- Compare against a flat plan using the Model Cost Index per-message rates
- If heavy technical use and control are the point, stay on BYOK with a clear conscience

FAQ:

Q: Is Whizi cheaper than TypingMind?
A: Per token for a heavy technical user, no: your own keys at list rates win, and that stays true. For moderate use across several model families, the flat plan is usually cheaper once you sum real provider bills, and it is always more predictable. Run your last three months of bills against it rather than guessing.

Q: Does Whizi support bring-your-own-keys?
A: No, and that is the point of the design: the plan includes the models, metered in credits, with no provider accounts to fund or manage. If BYOK control is a requirement, TypingMind and OpenRouter remain the right tools.

Q: What happens to my TypingMind setup if a provider changes its API?
A: That maintenance is yours in a BYOK setup: key rotations, deprecated models, changed endpoints. In a bundled workspace the provider absorbs it, which is one of the invisible costs the per-token spreadsheet never shows.

Q: Can I share a Whizi plan with someone non-technical?
A: Yes, and this is usually the actual switching trigger: the workspace is a normal app with a normal login. No API accounts, no keys, nothing to configure. The multi-model workflow becomes usable by people who would never touch a provider dashboard.

## Whizi compared with other subscriptions

### Whizi vs ChatGPT Plus: which AI subscription is worth paying for in 2026?

URL: https://whizi.io/vs/whizi-vs-chatgpt-plus/
Updated: August 2026

Quick answer: Whizi wins for most professionals. ChatGPT Plus includes only GPT models for around $20 a month, while Whizi Starter costs less and includes GPT, Claude, and open models, with Gemini and image generation on Pro. Choose ChatGPT Plus only if your team depends on shared Custom GPTs.

A clear, no-spin comparison of Whizi and ChatGPT Plus on models, features, pricing, and real workflow coverage, so you can pick the right subscription in minutes.

#### Quick verdict

Pick **ChatGPT Plus** if you only use GPT, you depend on OpenAI-specific features such as Custom GPTs your team has built, and you are content inside one ecosystem.

Pick **Whizi** if you want the same GPT models plus Claude, Gemini, and open models in one workspace, for less than ChatGPT Plus costs on its own.

For most people who write, research, or code professionally, Whizi wins for a specific reason: there is a consistent slice of work, usually a quarter to a third of it, where a non-GPT model is simply better, and covering that slice by adding subscriptions is how people end up paying $40 or $60 a month.

#### Where GPT is not the best answer

GPT is an excellent general model, and this is not a page arguing otherwise. It is the strongest default for fast implementation work, structured output, and following a brief precisely. The question is what happens on the tasks where it is not the best tool.

| Task | Better model | Why it matters |
| --- | --- | --- |
| Prose a person reads end to end | Claude | Less filler, sustains a voice, needs less editing before it is sendable |
| Difficult tone: bad news, critical feedback | Claude | Direct without being cold, which is the hardest register to get right |
| Very long documents and large corpora | Gemini | Largest context window, so nothing has to be chunked away |
| Current information with sources | Gemini | Tight web integration, links you can open |
| Independent critique of a draft | Anything that did not write it | The model that produced something is the worst reviewer of it |

That last row is the one no single-model subscription can deliver at any price. Asking GPT to find the flaw in GPT's answer produces agreement far more often than it produces a catch.

#### Feature comparison

| Feature | ChatGPT Plus | Whizi |
| --- | --- | --- |
| GPT models | Included | Included |
| Claude models | Not included | Included |
| Gemini models | Not included | Included on Pro and above |
| Open models | Not included | Included |
| Image generation | OpenAI image models | Multiple models including Flux and Stable Diffusion |
| Document chat | Yes | Yes, with a large-context model for long files |
| Voice | Yes | Included on higher plans |
| Reusable prompts | Custom GPTs | Cross-model templates with variables |
| Model switching per message | No | Yes |
| Side-by-side comparison | No | On Powerhouse |
| Cost | Around $20 per month | Starts below $20 per month |

#### Pricing, and the way this usually goes wrong

**ChatGPT Plus** is a flat monthly fee for GPT only. That is fine until the day you need Claude for a difficult client email or Gemini for a 200 page report. Then you add Claude Pro, which is another $20 or so, and possibly a Gemini plan, which is another $20 again. Three subscriptions, roughly $60 per person per month, none of which share context or history.

**Whizi Starter** sits below the price of a single one of those and already includes GPT alongside Claude and open models. **Whizi Pro** adds the higher-tier Claude and Gemini models, image generation, and more messages for roughly what one other major subscription costs alone. Run the numbers against your own stack with the [savings calculator](https://whizi.io/tools/calculator).

The cost that does not appear on any invoice is context loss. When research happens in one product and drafting in another, the reasoning does not travel, only the conclusions. The draft ends up thinner than your understanding, and you do not notice because nothing went visibly wrong.

#### What you give up leaving OpenAI's own product

Being straightforward, because the omissions are what make a comparison page useful.

**Custom GPTs and the store.** If your team shares Custom GPTs, or you rely on ones published by others, that ecosystem is OpenAI-specific. Porting your own is a five minute job per GPT, since the underlying instructions become a reusable prompt or a project's standing instructions, but a store of other people's work has no equivalent.

**Deep OpenAI integrations.** Anything built specifically against OpenAI's product surface, including some third-party plugins and connectors, stays there.

**Being first on new OpenAI features.** New capabilities appear in OpenAI's own product before they are available through the API to anyone else.

What you do not give up is the GPT models themselves. Same model family, same output. The difference is the wrapper and everything alongside it.

#### When ChatGPT Plus is still the right answer

ChatGPT Plus is genuinely the better subscription when:

- Your team has standardised on shared Custom GPTs that people use daily.
- You use the OpenAI store or specific integrations built for it.
- You are locked to one vendor for compliance, procurement, or integration reasons.
- You genuinely only use GPT and have never once wanted a second opinion.

For everyone else, the arithmetic and the workflow both point the same way. Test it rather than taking anyone's word for it: the checklist below takes a week and settles the question with your own work.

Key points:

- ChatGPT Plus covers one model family. Whizi covers every major one, generally for less.
- The specific tasks where GPT is not the best answer, which is where the second subscription usually comes from
- What you give up leaving OpenAI's own product, stated plainly

Checklist:

- List the last ten things you asked ChatGPT Plus to do
- Run each one in Whizi with GPT first, then with Claude and Gemini
- Note which answer you would actually have shipped
- Pay attention to the writing tasks and the long-document tasks specifically
- Try the pattern GPT alone cannot do: draft in one model, critique in another
- Check whether you depend on any Custom GPT you did not build yourself
- If a non-GPT model wins twice in ten, one subscription is already costing you quality

FAQ:

Q: Does Whizi really include the same GPT that ChatGPT Plus does?
A: Whizi includes the same GPT model family that powers ChatGPT Plus, accessed through OpenAI’s API, so output quality is the model’s rather than the interface’s. Some product-specific features, notably Custom GPTs and the store, live in OpenAI’s own app and are not replicated. The models themselves are the same.

Q: Can I keep ChatGPT Plus and use Whizi?
A: Yes, and running both for a week or two is the sensible way to decide rather than switching on faith. Send the same real tasks to each and compare what you would actually have shipped. Most people who do this cancel ChatGPT Plus, since the GPT models are present in Whizi anyway and the second subscription is buying an interface rather than a capability.

Q: Does Whizi have Custom GPTs?
A: The equivalent is a reusable prompt plus a project's standing instructions, which cover the same professional use cases with one meaningful advantage: they run against Claude and Gemini as well as GPT, so the work is not tied to whichever model was best when you wrote it. Porting a Custom GPT takes about five minutes. What does not port is the store of GPTs built by other people.

Q: Is Whizi actually cheaper?
A: Whizi Starter is below the price of ChatGPT Plus while including Claude and open models alongside GPT, so on models per dollar it is clearly ahead. The larger saving appears if you were heading toward a second or third subscription. Two plans is roughly $40 a month and three is roughly $60, for products that do not share context, history, or prompts.

Q: What about privacy and data handling?
A: Whizi routes requests to each provider on your behalf and does not train on your conversations, and each provider’s data policy is available for review before you enable that model. For company use, your own internal policy is usually the stricter constraint. The practical habit worth keeping either way is to redact identifiers and index sensitive figures, which preserves the analysis while removing most of the exposure.

### Whizi vs Claude Pro: keep Claude, add every other major model

URL: https://whizi.io/vs/whizi-vs-claude-pro/
Updated: August 2026

Quick answer: Whizi wins unless you depend on Anthropic-specific features. Claude Pro costs around $20 a month for Claude alone. Whizi Starter starts below that and adds GPT and open models, while Whizi Pro carries Claude Sonnet and Opus alongside Gemini. Keep Claude Pro if you rely on Projects or Artifacts.

Side-by-side comparison of Whizi and Claude Pro on models, pricing, and workflow coverage. Keep Claude Sonnet, add GPT and Gemini, and pay less overall.

#### Quick verdict

Pick **Claude Pro** if you use Claude for essentially everything, you rely on Anthropic-specific features such as Projects or Artifacts, and you have no interest in comparing against other models.

Pick **Whizi** if you love Claude for writing and reasoning but keep opening something else for live research, quick structured extraction, or long-document work. That is the common pattern, and paying two subscriptions to serve it is the expensive way to solve it.

The honest test: how often do you finish a Claude conversation and immediately open a different tool? If the answer is more than never, you are already running a multi-model workflow and paying per provider for it.

#### What Claude is genuinely best at

This page is not an argument against Claude. It is the best model available for several things, and knowing which ones makes both products more useful.

**Prose that a human reads end to end.** Less filler, better sustained voice across a long document, and consistently the least editing required before something is sendable.

**Difficult tone.** Delivering bad news, declining a request, giving critical feedback. It handles the register that requires being direct without being cold better than anything else.

**Careful reading of nuanced text.** Contracts, policy, anything where meaning depends on qualification. It is also unusually good at telling you what a document does not say, which is often the actual question.

**Critique.** Ask Claude what is wrong with a draft, an argument, or a plan and you get specific, useful objections rather than encouragement.

None of that changes in Whizi, because it is the same model family. What changes is that when the task is live research, a hundred-page report, or a strict JSON schema across two hundred rows, the better model is one click away instead of one subscription away.

#### Feature comparison

| | Claude Pro | Whizi |
| --- | --- | --- |
| Claude Sonnet | Included | Included on Pro and above |
| Claude Opus | Included, with limits | Included on Pro and above |
| GPT models | Not included | Included |
| Gemini models | Not included | Included on Pro and above |
| Open models | Not included | Included |
| Image generation | Not included | Included on Pro and above |
| Reusable prompts | Projects | Cross-model templates with variables |
| Model switching per message | No | Yes |
| Side-by-side comparison | No | On Powerhouse |
| Artifacts and in-product code preview | Yes | Not equivalent |
| Cost | Around $20 per month | Starts below $20 per month |

#### What you actually give up

Every switch has a cost and pretending otherwise wastes your time.

**Projects.** Claude Pro's Projects hold a persistent knowledge base and custom instructions for a body of work. Whizi has its own Projects, with pinned files and standing instructions that apply to every chat inside, and with the advantage of running against any model. The two implementations are not identical, and porting takes a few minutes per project.

**Artifacts.** Anthropic's side panel for rendering code, documents, and small applications inline is a nice piece of product design with no direct equivalent.

**First-party feature timing.** New Anthropic capabilities appear in Anthropic's own product first. If being early on those matters to you, that is a genuine argument for the first-party subscription.

What you do not give up is the model. Same Claude family, same output quality. The difference is in the wrapper.

#### The arithmetic

Claude Pro is around $20 a month for Claude alone. If you also pay for ChatGPT Plus, you are at roughly $40. Add a Gemini plan for long-document work and you are at $60 per person per month for three products that do not share context, history, or prompts.

Whizi Starter sits below the price of any single one of those and covers GPT, Claude, and Gemini together. Pro adds the higher-tier models, image generation, and more messages for roughly what one other major subscription costs on its own. Run your own numbers with the [savings calculator](https://whizi.io/tools/calculator).

The saving that does not show up on the invoice is context. Three subscriptions means the research you did in one product is invisible to the model drafting in another, so you either re-explain it or lose the nuance. One thread keeps all of it.

#### How to test it properly in a week

1. For five days, log every task you send to Claude and every task where you opened something else instead.
2. Look at that second list. Live research, long PDFs, strict structured output, and quick factual lookups are the usual entries, and they are precisely where a different model is better.
3. In Whizi, run three tasks that Claude currently handles and confirm the quality is what you expect. It should be, since it is the same model family.
4. Then run three tasks from the second list and compare against however you were solving them before.
5. Try the pattern only a multi-model workspace allows: draft in Claude, then switch and ask a different model to find the flaw in the argument. That single habit is the thing Claude Pro cannot do at any price.

If your second list is empty after five days, Claude Pro is the right subscription for you and you should keep it.

Key points:

- The same Claude models, plus GPT and Gemini, generally for less than Claude Pro alone
- An honest account of what Claude does better and where you should still reach for it
- What you give up leaving Anthropic's own product, including Projects and Artifacts

Checklist:

- Log for a week which tasks go to Claude and which send you elsewhere
- Note whether you rely on Projects or Artifacts specifically
- Verify Claude output quality in Whizi on three of your normal tasks
- Test the tasks where you currently leave Claude, using GPT and Gemini
- Try drafting in Claude and critiquing in a different model
- Add up every AI subscription you currently pay for

FAQ:

Q: Is Claude in Whizi the same as in Claude.ai?
A: It is the same Claude model family, including Sonnet and Opus, accessed through Anthropic’s API. Output quality is the model’s, not the interface’s. What differs is the surrounding product: Artifacts are Anthropic-specific, while Whizi adds model switching, side-by-side comparison, and its own Projects that run across every model family.

Q: Will I lose Claude Projects if I switch?
A: They port. Whizi has its own Projects, holding pinned files and standing instructions that apply to every chat inside, so moving a Claude Project means copying the custom instructions and reference material across, a few minutes each. The two implementations are not identical, so check that what you rely on carries. The gain is that the ported version runs against Claude, GPT, or Gemini rather than being fixed to one.

Q: Can I cancel Claude Pro?
A: Most people who switch do, because the same Claude models are available alongside GPT and Gemini for less than Claude Pro costs on its own. The cases where keeping it makes sense are specific: heavy reliance on Projects or Artifacts, or wanting new Anthropic features the moment they ship. Test for a week before cancelling rather than taking a marketing page’s word for it.

Q: Does Whizi have Claude Opus?
A: Yes, on Pro and above, alongside Sonnet. Opus is worth reaching for on genuinely hard reasoning and on long, complex documents, while Sonnet handles the majority of daily work faster. The practical benefit of a multi-model workspace is that when a task is not actually a Claude task, you switch rather than paying an Opus-tier request to do something GPT or Gemini does better.

Q: Is Claude still the best model for writing?
A: For prose a person reads end to end, it remains the strongest choice in our own testing, and it needs the least editing before something is sendable. That is exactly why Whizi includes it rather than arguing against it. The point is that writing is one job among several, and the models that win at live research, long-document reading, and strict structured output are different ones.

### Whizi vs Gemini Advanced: keep Gemini 3 Pro, skip the Google bundle

URL: https://whizi.io/vs/whizi-vs-gemini-advanced/
Updated: August 2026

Quick answer: Whizi wins unless your work lives inside Google apps. Gemini Advanced bundles Gemini with 2TB of storage and sidebar AI in Docs and Gmail for around $20 a month. Whizi starts below that and puts Gemini beside GPT and Claude, without the bundled storage. Gemini Pro sits on Whizi Pro and above.

Compare Whizi and Gemini Advanced (Google One AI Premium) on models, pricing, and workflow coverage. Keep Gemini, add GPT and Claude, and drop the bundle tax.

#### Quick verdict

Pick **Gemini Advanced** if your work lives inside Gmail, Docs, and Drive, and you want AI in the sidebar of the tools you already have open. That integration is the product's genuine advantage and no external workspace can replicate it.

Pick **Whizi** if you want Gemini for what it is best at, which is long documents and current information, alongside GPT and Claude for the tasks where they are better. Same Gemini model family, plus the other two, without paying for cloud storage as part of the deal.

The deciding question is not about models. It is whether the AI needs to be inside your documents or beside them.

#### What each subscription actually contains

| | Gemini Advanced | Whizi |
| --- | --- | --- |
| Gemini Pro | Included | Included on Pro and above |
| Gemini Flash | Included | Included |
| GPT models | Not included | Included |
| Claude models | Not included | Included |
| Open models | Not included | Included |
| Image generation | Google's image models | Multiple models including Flux and Stable Diffusion |
| Gmail, Docs, Drive sidebar | Yes | No |
| Cloud storage | 2TB bundled | Not bundled |
| Model switching per message | Within Gemini | Across every family |
| Side-by-side comparison | No | On Powerhouse |
| Cost | Around $20 per month | Starts below $20 per month |

Two rows explain most of the decision. Gemini Advanced gives you deep integration into one ecosystem. Whizi gives you three model families and the ability to move between them mid-task. Which one is worth more depends entirely on whether your work is already inside Google's tools.

#### The bundle question

Gemini Advanced comes as part of Google One AI Premium, which includes 2TB of cloud storage. Whether that is value or padding is a question only you can answer, and it is worth answering honestly.

Ask yourself: would you buy 2TB of Google storage if AI were not attached? If you are already on a paid Google One tier, or you have a large photo library, or your household runs on Drive, then yes, and the AI is genuinely arriving at a discount. If you are on the free tier and using a fraction of it, you are paying a subscription price where a meaningful share goes to storage you will never fill.

The same logic applies in reverse and is worth stating: if you need the storage anyway, Gemini Advanced is very good value and this page is not going to argue otherwise.

#### Where Gemini genuinely wins, and where it does not

Being specific here matters more than a verdict, because Gemini is excellent at particular things and this comparison is not about whether the model is good.

**Gemini is the best choice for:** very long documents, thanks to the largest context window of the major models; current information, because of tight web integration; and multimodal work across images, video, and text together.

**Gemini is usually not the best choice for:** prose that a person reads end to end, where Claude tends to need less editing; and strict structured output, where GPT is more reliable at holding an exact schema across many rows.

That is precisely the argument for a multi-model workspace. Keeping Gemini for the reading and research it excels at, while sending the drafting to Claude and the extraction to GPT, is a better workflow than using one model for all three because it is the one you pay for. Whizi is where that routing is one click rather than three products.

#### What you give up by leaving the Google bundle

Worth being straight about, because it is the real cost.

**Sidebar AI inside Docs, Gmail, and Drive.** Highlighting a paragraph in a doc and asking for a rewrite in place is a genuinely different experience from copying it into another tab. If most of your writing happens inside Google Docs, this is a meaningful loss.

**Access to your Drive contents without uploading.** Gemini can reach files you already have in Drive. Working outside it means uploading, which is a small friction repeated often.

**One account, one bill, one login.** Household plans and existing Google billing are real conveniences.

The mitigation is straightforward but not free: upload the document rather than referencing it, and paste content in rather than editing in place. For research and analysis, where the document is an input, that costs almost nothing. For editing work where the document is the artefact, it costs more, and if that is most of your day, Gemini Advanced may be the right answer for you.

#### A week-long test

1. For one week, note every AI task and mark whether it happened inside a Google app or in a separate tab.
2. Count how many of the separate-tab tasks were writing, structured extraction, or code. Those are the ones where a different model would likely have produced a better result.
3. Run three of them in Whizi with Claude or GPT and compare against what Gemini gave you.
4. Separately, check your Google storage usage and decide whether you would pay for 2TB on its own.

The pattern is usually clear by the end of the week. People whose work is inside Docs and Gmail feel the integration constantly. People who work in a browser tab anyway are paying for an integration they never touch, plus storage they do not need, while missing two model families they would use.

Key points:

- Both give you Gemini. Only one of them also gives you GPT and Claude
- The Workspace integration is the real reason to keep Gemini Advanced, and it is a good one for some people
- How to work out whether you are paying for storage you would not otherwise buy

Checklist:

- Decide whether you would pay for 2TB of storage if no AI were attached
- Track for a week whether your AI tasks happen inside Google apps or in a separate tab
- Identify the writing and extraction tasks where Gemini is not the strongest choice
- Run three of those in Claude and GPT and compare the editing effort
- Check whether your writing lives in Google Docs, since that is the real lock-in
- Add up what you would pay to get GPT and Claude separately

FAQ:

Q: Is Gemini in Whizi the same as in Gemini Advanced?
A: It is the same Google model family, accessed through the provider API. The models are the same; what differs is everything around them. Gemini Advanced wraps them in Google’s own interface with Workspace integration, while Whizi puts them next to GPT and Claude with model switching, side-by-side comparison, and projects that keep each body of work together.

Q: Does Whizi integrate with Google Docs?
A: Not natively. You upload documents or paste content, which works well when the document is an input to research or analysis and is more friction when the document is the thing you are editing. Sidebar integration inside Docs and Gmail is specific to Google’s own product and is the strongest reason to keep Gemini Advanced if that is where your work lives.

Q: Is Whizi cheaper than Gemini Advanced?
A: Whizi Starter is below Gemini Advanced pricing and includes GPT and Claude alongside Gemini, so on models per dollar it is clearly ahead. The fair caveat is the 2TB of storage bundled with Google One: if you would buy that storage anyway, Gemini Advanced is good value and the comparison is closer than the model list suggests.

Q: Can I keep Gemini Advanced and use Whizi?
A: Yes, and some people reasonably do, keeping the Google bundle for in-document work and using Whizi for tasks where Claude or GPT is stronger. At around $40 a month combined it is worth checking whether the Workspace integration is genuinely earning its share. Most people who try both for a fortnight find they are using one of them for almost everything.

Q: Which is better for long documents?
A: Both use the same Gemini models, so raw long-document capability is equivalent and Gemini is genuinely the strongest option for that task. The difference is what happens next: in Whizi you can hand the same document to Claude for the write-up or to GPT for structured extraction without re-uploading, which is usually the second half of any long-document job.

### Whizi vs Grok: one model with a firehose, or every model in one place

URL: https://whizi.io/vs/whizi-vs-grok/
Updated: August 2026

Quick answer: Whizi wins unless live X data is the job. A Grok subscription is around $20 a month and owns real-time X access, which no workspace replicates; if X Premium+ already bundles it, keep it. Whizi carries Grok models on its higher tiers next to GPT, Claude, and Gemini, in one conversation.

Compare Whizi and a Grok subscription on real-time data, reasoning, tone, and cost. Grok owns live X access; Whizi puts Grok models next to GPT, Claude, and Gemini.

#### Quick verdict

Pick **a Grok subscription** if live X data is the job. Breaking news, sentiment right now, what people are posting about a launch this hour. Grok pulls from posts as they happen, the others read a web that indexes later, and no workspace replicates that. If you already pay for X Premium+, Grok is bundled and effectively free, and this whole comparison is moot: keep it.

Pick **Whizi** if what you want from Grok is the model rather than the firehose. Whizi carries Grok models on its higher tiers next to GPT, Claude, and Gemini, in one conversation, for less than a dedicated subscription. The reasoning travels; the X integration does not.

The test: count how many of your last twenty Grok questions needed data from the last hour. High count, subscribe natively. Near zero, you want the model, not the product.

#### Side by side

| | Grok subscription | Whizi |
| --- | --- | --- |
| Live X data | The defining feature | No |
| Grok models | Yes, always current | Yes, on higher tiers |
| GPT, Claude, Gemini | No | Yes, same conversation |
| Second opinions across model families | No | One click |
| Image generation | Included | Included on paid plans |
| Tone | Direct by default | Direct when you ask for it |
| Files and documents | Supported, workflow is younger | Core workflow |
| Cost | Around $20 a month, heavier tier above | Starts below that, models included by tier |

The tone row deserves a sentence. Grok is deliberately less hedged, which is a real feature for satire and for questions where you want a position. Most of that is promptable elsewhere: tell any strong model to commit to an answer and cap itself at one caveat. What is not promptable is live X access, which is why it sits in the first row.

#### What the second subscription actually buys

The typical person weighing this already pays for ChatGPT Plus or Claude Pro. Adding Grok makes it around $40 a month across two apps that cannot see each other's answers, and the workflow cost is the tab switch: re-explaining context to the second assistant, or more often not bothering.

In one workspace the comparison becomes a habit rather than a project. Ask Grok the hard question, then ask Claude what the answer missed, in the same thread, with the same context. Different model families fail differently, so the disagreement between them is signal you can act on before a mistake ships.

If the subscriptions have already stacked, put your real list into the [AI subscription savings calculator](https://whizi.io/tools/calculator) and look at the annual number before deciding anything.

#### The honest trades

What Whizi does not give you: live X data, and Grok models on the entry plan; they sit on the higher tiers. If either is decisive, the native subscription is the right call and this page just saved you a trial. If it is not, [how to cancel SuperGrok](https://whizi.io/resources/how-to-cancel-supergrok) walks through ending the native billing first.

What the native subscription does not give you: any second model family, and the proofreading backstop that comes from one. Grok has had public incidents of answers that were offensive or badly wrong, and its directness reaches deliverables unfiltered. Where the stakes are client-facing, drafting with a conservative model and using Grok for the thinking is the safer split, and that split needs both in one place.

Key points:

- The one capability a Grok subscription owns outright, stated up front
- What using Grok models inside a workspace actually gets you, and what it does not
- The cost picture when Grok would be your second or third AI subscription

Checklist:

- Count how many of your last twenty Grok questions needed the last hour of X
- If X Premium+ is already on your card, keep bundled Grok and stop here
- Price the two-subscription stack as an annual number, not a monthly one
- Run one hard question through Grok and Claude side by side before deciding
- Keep client-facing wording with a conservative model, use Grok for the thinking

FAQ:

Q: Does Whizi include Grok models?
A: Yes, on the higher tiers, in the same picker as GPT, Claude, and Gemini. You get the model's reasoning and speed. You do not get live X data, which stays exclusive to the native product.

Q: Is a Grok subscription worth it on its own?
A: If your questions genuinely need live X data, yes, nothing else does that job. If they are general reasoning, writing, and code, you are paying a dedicated price for a model family a multi-model workspace serves for less, with alternatives beside it.

Q: Can Whizi match Grok's tone?
A: Mostly. Directness is promptable: ask any strong model to commit to a position, skip the balanced-overview framing, and cap itself at one caveat. The unfiltered edge that occasionally gets Grok into public trouble is the part you likely do not want in deliverables anyway.

Q: What does the comparison cost to run myself?
A: Whizi starts at $0.99 for the first trial, and Grok has a free tier with limits. Run the same three real tasks from your actual week through both before committing to either at full price. Your work is a better benchmark than anyone's chart.

### Whizi vs OpenRouter: a product vs an API gateway

URL: https://whizi.io/vs/whizi-vs-openrouter/
Updated: August 2026

Quick answer: Neither wins outright: they answer different questions. OpenRouter is an API gateway with per-token pricing, built for developers writing code that calls a model. Whizi is a finished workspace with threads, files, projects, and comparison, on a flat monthly plan. Many teams run both: the gateway for production, the product for the team.

OpenRouter is an API router for developers. Whizi is a finished multi-model workspace for professionals. See which one fits your use case.

#### These are not really competitors

OpenRouter is an API gateway. You get one endpoint and one key that route to models from many providers, with per-token pricing and automatic failover. It exists so that software you are building can talk to several model providers without integrating each one separately.

Whizi is an end-user product. You sign in and work: chat threads, file uploads, image generation, projects, model switching mid-conversation, side-by-side comparison. There is no key to manage because there is no code involved.

Comparing them is a bit like comparing a wholesale supplier to a shop. The honest question is not which is better, it is which one your situation calls for, and the answer is usually determined in one sentence: are you building something, or are you doing your work?

#### What each one actually gives you

| | OpenRouter | Whizi |
| --- | --- | --- |
| What it is | API gateway across providers | Finished multi-model workspace |
| Interface | API, plus a basic chat for testing | Full product: threads, files, projects, comparison |
| Pricing | Per token, pay as you go | Flat monthly plan |
| Setup | API key, integration code, billing management | Sign in |
| Model breadth | Very large catalogue, including niche models | Major first-party families plus leading open models |
| Image generation | Not the focus | Included on Pro and above |
| File and document work | You build it | Built in, including large-context reading |
| Prompt templates | You build it | Built in, cross-model, with variables |
| Team use | Whatever you build | Built in on team plans |
| Who it is for | Developers shipping software | Anyone doing knowledge work |

The row that matters most is "you build it". Everything OpenRouter does not provide is buildable, and building it is a project. If you were going to build a chat interface, a file pipeline, a template system, and team sharing anyway, the gateway is the right foundation. If you just wanted those things to exist, you are looking at weeks of work to arrive where a product already is.

#### The pricing comparison people actually want

Per-token pricing is genuinely cheaper for light use. If you send a handful of messages a week, you will spend a couple of dollars and a flat subscription would be poor value. That is a real and honest advantage.

It stops being cheaper as usage rises, and the crossover comes sooner than most people expect, because the things that make AI useful are the expensive things. Long documents in context, image generation, comparing two models on one prompt, and long threads where the whole history is re-sent with every message all consume tokens quickly. A single afternoon of serious document work can cost more in API tokens than a month of a flat plan.

There is also the cost that does not appear on the invoice. Metered billing changes behaviour: you trim context, skip a comparison, avoid re-running a prompt with better instructions. Every one of those is a small quality decision made for accounting reasons.

For a broader breakdown of where AI spending goes, see [AI subscription costs](https://whizi.io/resources/ai-subscription-costs).

#### Using both, which many teams do

The split that works is by role rather than by preference.

**OpenRouter, or a direct provider API, for what you ship.** Your production application, your background jobs, your data pipeline. These need programmatic access, per-request control, and pricing that scales with actual volume. A subscription product is the wrong shape for that.

**Whizi for what your team does.** Writing, research, analysis, document work, and the daily model comparison that everyone from marketing to support benefits from. These people should not be managing an API key, and the cost of building them a decent interface is far higher than the subscription.

This split also has an organisational benefit. Engineering keeps its usage lean and observable, and everyone else gets a tool that does not require them to file a ticket. Trying to serve both from one side of the line is where teams end up either with a homegrown chat UI nobody maintains, or with engineers on a consumer plan that cannot do what their application needs.

#### When to pick which

**Pick OpenRouter if:** you are writing code that calls a model; you need failover across providers or want to route by cost; you need per-request control over parameters; your usage is very light; or you specifically need a niche model from the long tail of the catalogue.

**Pick Whizi if:** you want to use the models rather than integrate them; you want file uploads, image generation, templates, and comparison without building them; your usage is regular enough that predictable pricing beats metered; or the people who need access are not developers.

**Pick both if:** you have an application in production and a team that also does daily knowledge work. That is the common case at any company past a handful of people.

Key points:

- One is infrastructure and one is a product, which makes most feature comparisons beside the point
- When per-token pricing is genuinely cheaper, and the usage level where that flips
- Why many teams run both, and how to split them sensibly

Checklist:

- Answer the first question honestly: are you building software, or doing work?
- Estimate your monthly token spend at your real usage, including long documents and comparisons
- Count what you would have to build to match a finished product, and price that time
- If you are already trimming context to save cost, price a flat plan
- If your team includes non-developers, do not make them manage API keys
- Consider the split: gateway for production, product for the team

FAQ:

Q: Does Whizi expose an API?
A: Whizi is primarily a product experience rather than an infrastructure layer, so for programmatic access at scale, OpenRouter or a direct provider API is the right tool. That is not a limitation being apologised for, it is the distinction between the two: one is built for software to call, the other is built for a person to work in.

Q: Is Whizi cheaper than OpenRouter for one user?
A: For light use, no, per-token billing wins and a flat plan would be poor value. For regular daily use, yes, and the crossover arrives sooner than expected because the genuinely useful operations are the expensive ones: long documents in context, image generation, running one prompt through two models, and long threads that re-send their history with every message.

Q: Can I share my Whizi workspace with teammates?
A: No. Whizi plans are individual: there is no team, seat or workspace concept, so each person subscribes on their own account. With a gateway the same is true in a different shape, since anything collaborative is something you build and maintain yourself. If shared billing is the requirement, neither product answers it out of the box today.

Q: Which has more models?
A: OpenRouter, by a wide margin, since aggregating a very large catalogue including niche and experimental models is the core of what a gateway does. Whizi carries the major first-party families plus leading open models, which covers what most professional work needs. If a specific long-tail model is essential to you, the gateway is the correct tool.

Q: Can non-developers use OpenRouter?
A: There is a basic chat interface for testing, but it is built for developers verifying that routing works, not as a daily workspace, and it does not aim to provide document workflows, templates, or team features. Pointing non-technical colleagues at an API gateway generally ends with someone building an internal chat UI that nobody has time to maintain.

### Whizi vs Perplexity Pro: research plus the writing that follows

URL: https://whizi.io/vs/whizi-vs-perplexity-pro/
Updated: August 2026

Quick answer: Whizi wins if your research usually ends in a document. Perplexity Pro is the better product for pure cited search, around $20 a month. Whizi starts below that and keeps the research, the drafting in Claude, and the coding in one thread, so the reasoning carries into the deliverable.

Compare Whizi and Perplexity Pro on research, writing, coding, and cost. Keep the citation flow, add Claude and GPT, and get a full workspace instead of a search engine.

#### Quick verdict

Pick **Perplexity Pro** if your AI use is mostly answering questions. What happened, who makes this, what does this term mean, what is the current state of X. Perplexity is the cleanest product built for that, and if research is the whole job rather than the first step of one, it wins.

Pick **Whizi** if research is usually followed by producing something. A memo, a brief, a deck, an email, a draft, a piece of code. That second half is a different job requiring different models, and doing it in the same thread as the research is meaningfully faster than doing it in a second tool.

The test is simple. Look at your last ten Perplexity searches and ask what you did immediately afterwards. If the answer was mostly "nothing, I had what I needed", stay. If it was mostly "opened another tool and started writing", you are paying for a handoff.

#### Two kinds of research

**Answering a question** means you want a specific fact with a source you can check. Perplexity is built end to end for this: query, sourced answer, follow-up refinement. The Focus modes and the follow-up flow are better for that specific loop than a general chat.

**Building an understanding** is different. You are reading fifteen sources, extracting comparable facts from each, noticing where they disagree, and then writing something that takes a position. This is less like search and more like reading a stack of documents in one place. It needs a large context window, extraction into consistent structure, and a model that writes well at the end of it.

Most professional research is the second kind wearing the clothes of the first. You start with a question and end up needing a document.

| | Perplexity Pro | Whizi |
| --- | --- | --- |
| Cited web search | Best in class for the query-answer loop | Yes, with web-connected models |
| Follow-up and refinement UX | Purpose-built | General chat |
| Reading many documents at once | Supported | Large-context model holds the whole set |
| Extraction into a consistent table | Possible | GPT with a strict schema |
| Drafting the deliverable | Not the focus | Claude in the same thread |
| Coding | Limited | GPT and Claude |
| Image generation | Limited | Included on Pro and above |
| Model choice | Selectable within the product | Switch per message across families |
| Cost | Around $20 per month | Starts below $20 per month |

#### What the handoff actually costs

The obvious cost of two tools is the second subscription. The larger cost is the context that does not travel.

When you finish researching in one product and start writing in another, you carry over the conclusions and lose the reasoning: which sources disagreed and why, the figure you decided not to use, the caveat attached to the headline number, the three things you checked and rejected. The writing model starts blind, so you either re-explain all of it or, more commonly, you do not, and the draft is thinner than your understanding.

Doing both in one thread means the drafting model has the full research history in front of it. Practically, that shows up as fewer instructions needed, fewer factual slips between research and draft, and a first draft that already contains the nuance you found rather than the summary of it.

The other thing that only works in one place is verification against your own research. `Check this draft against the sources above. Flag every claim that is stronger than what the source supports, and every number I have carried over incorrectly.` That prompt requires the research and the draft to be in the same conversation.

#### Reproducing the Perplexity flow in Whizi

The citation habit is worth keeping, and it transfers if you make the demands explicit rather than relying on the product to enforce them.

**The search prompt**

`Research [question]. Return findings with a source URL and publication date for each. Where sources disagree, show both rather than reconciling them. Mark anything you cannot source as UNVERIFIED. Do not fill gaps with plausible values.`

**The follow-up**

`For the claim that [specific finding], find the original source rather than an article citing it, and tell me what the original actually measured.`

That second prompt is one Perplexity users often want and rarely ask for. Statistics degrade as they are passed along, and the figure three citations deep is frequently unrecognisable at the origin.

**Then the part Perplexity does not do**

`Using only the sourced findings above, draft [the deliverable] for [audience]. Every claim must trace to a source in this thread. Mark anything I still need to verify. Voice: [paste sample].`

Switch to Claude for that last step. The research stays visible in the thread. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation) and the fuller treatment in [AI for market research](https://whizi.io/resources/ai-for-market-research).

#### Where Perplexity still wins

This is not a page arguing that Perplexity is bad, and pretending otherwise would be unhelpful.

It remains the best product for pure cited search. The query-to-answer loop is tighter, the follow-up interaction is designed rather than improvised, the citation presentation is cleaner, and Focus modes are a genuinely good idea for constraining a search to a source type. For quick factual lookups throughout the day it is a better experience than typing a carefully constrained prompt into a general chat.

If your AI use is largely reading and answering questions, and you already have writing covered elsewhere or do not need it, keeping Perplexity is a reasonable decision. Running both is also perfectly sensible, though at that point the arithmetic is worth doing: two subscriptions at around $20 each is more than one plan that covers research, writing, coding, and images.

#### Decide it with your own work

1. Take your last ten Perplexity searches. Write down what you did in the ten minutes after each.
2. Count how many ended in a deliverable rather than in the answer itself.
3. Take one of those and run it end to end in Whizi: the sourced research prompt, then the draft in Claude in the same thread.
4. Compare on two things: how much you had to re-explain, and whether the draft contained the nuance from the research or just the conclusion.

If most of your searches end at the answer, stay where you are. If most of them are the first half of a task, one thread beats two tabs.

Key points:

- Where Perplexity is genuinely better, stated plainly, and it is a real category
- Why research that produces a deliverable is a different job from research that answers a question
- What the handoff between two tools actually costs you, beyond the second subscription

Checklist:

- Look at your last ten searches and note what you did immediately after each
- Count how many ended in a document rather than in the answer
- Demand a URL and date on every finding, in the prompt itself
- Ask for the original source rather than an article citing it
- Draft in the same thread as the research so the nuance carries over
- Run the check-the-draft-against-the-sources prompt before you send anything
- Add up two subscriptions before deciding to keep both

FAQ:

Q: Can I get Perplexity-style citations in Whizi?
A: Yes, when you use web-connected models, though you have to ask for it explicitly rather than getting it by default. Include the requirement in the prompt: a source URL and publication date for every finding, both figures shown where sources disagree, and anything unverifiable marked as such. The presentation is less polished than Perplexity’s, and the verification discipline is identical either way, since you should be opening the sources regardless.

Q: Should I cancel Perplexity Pro?
A: If your research usually ends in something you write, yes, because the handoff between two tools costs you the reasoning and not just the second subscription. If you use Perplexity purely as a cited search engine and value the query-answer loop, keeping it is reasonable, and Whizi can run alongside. Worth doing the arithmetic first: two subscriptions at around $20 each is more than one plan covering research, writing, coding, and images.

Q: Is Whizi cheaper than Perplexity Pro plus ChatGPT Plus?
A: Yes. Those two together run around $40 a month, and Whizi Starter is below the price of either one individually while covering GPT, Claude, and Gemini. The saving is real, but the more useful difference is that research and drafting happen in one thread, so the context survives instead of being summarised across a tab switch.

Q: Which is better for academic research?
A: They serve different stages. Perplexity is good for finding what exists on a topic. Whizi is better for the stage after that, where you upload the papers themselves and work across them: an evidence table with one row per study, extraction into a consistent schema, and a drafting model that preserves academic hedging. Neither removes the requirement to open every source and verify every DOI yourself, which is the failure mode that actually matters.

Q: Does Whizi have Focus modes or Collections?
A: Not as named features. The equivalent is a reusable prompt: one you keep and paste, that constrains sources, sets the output format, and encodes the standards you want applied every time. It requires slightly more setup than clicking a mode, and in exchange it is fully editable and runs against any model rather than being fixed by the product.

### Whizi vs Poe: a focused AI workspace vs a bot marketplace

URL: https://whizi.io/vs/whizi-vs-poe/
Updated: August 2026

Quick answer: Whizi wins once you know which models you use. Poe runs on compute points, so every message spends from a wallet and expensive models drain it fast. Whizi is a flat monthly plan with first-party GPT, Claude, and Gemini. Stay on Poe for community bots, niche models, or light occasional use.

Compare Whizi and Poe on models, pricing, and professional workflow. Decide between a focused multi-model workspace and an exploratory bot marketplace.

#### Quick verdict

Pick **Poe** if the appeal is breadth: community bots, personas, obscure and experimental models, and the ability to try something you have never heard of on a whim. It is genuinely the best product for that, and there is nothing dishonest about wanting it.

Pick **Whizi** if you have already discovered which models you use and now want to use them without thinking about the meter. Flat pricing, first-party Claude, GPT, and Gemini, and a workspace built around finishing a piece of work rather than exploring.

The clearest signal is your own behaviour. If you open Poe and go straight to the same two or three bots every time, you are paying for a marketplace you have stopped shopping in.

#### The pricing models are the real difference

Poe runs on a compute-points economy. Each subscription tier includes a points allowance, and every message spends points at a rate determined by which model answers it. Expensive models drain the wallet fast; cheap ones barely register.

That design is flexible, and it has a behavioural cost that people underestimate. Once a meter is visible, you start optimising against it: choosing a weaker model for a task that deserved the strong one, hesitating before a second attempt, not running the comparison you should have run. The rationing is invisible in your usage stats and shows up as slightly worse work.

Flat pricing removes that decision. You pick the model that fits the task, and if you want to run the same prompt through two models to see which is better, you just do it. The value is not only cost, it is that a category of small daily decisions disappears.

| | Poe | Whizi |
| --- | --- | --- |
| Pricing model | Compute points per message, rate varies by model | Flat monthly plan |
| Cost predictability | Depends on which models you use | Fixed |
| Behaviour under load | Encourages rationing near the end of a cycle | No meter to manage |
| Best for | Occasional heavy use of expensive models | Consistent daily use |

#### Marketplace breadth against workspace depth

Poe's catalogue is enormous. Alongside the major models there are thousands of community bots: personas, characters, task-specific wrappers, and prompt bundles published by other users. For discovery that is a real advantage, and there is no equivalent in a focused workspace.

The trade-off is that most community bots are a system prompt in a costume. Once you understand that, the ones you use regularly are usually reproducible in a couple of minutes, and reproducing them gives you something better: a prompt you can read, edit, and run against any model rather than a black box tied to one.

Whizi is not the only workspace shaped this way, and the wider field is laid out in [apps like Poe](https://whizi.io/alternatives/apps-like-poe).

| | Poe | Whizi |
| --- | --- | --- |
| First-party models | Yes, among many bots | Yes, the core of the product |
| Community bots and personas | Extensive | Not offered |
| Reusable prompts | Per bot | Cross-model templates with variables |
| Side-by-side comparison | Not the focus | Built in on Powerhouse |
| Image generation | Via separate bots | Integrated in the chat |
| Document and file work | Supported | Supported, with a large-context model for long files |
| Interface | Bot list and chats | Threads and projects built around a piece of work |

#### Rebuilding your Poe bots as reusable prompts

If two or three bots are the reason you stay, port them. The process takes a few minutes each.

1. Open the bot and look at its description and any published instructions. For bots you built yourself, you already have the system prompt.
2. Keep that instruction text as a reusable prompt, or put it in a Whizi project's instructions so every chat in the project starts with it.
3. Mark whatever changes per use, for example the topic or the draft, so the fixed part stays obvious.
4. Add the parts most community bots leave out: an explicit output format and an escape hatch for inputs that do not fit.
5. Run it once against two models. Very often the bot was pinned to a model that is no longer the best choice for that job.

That last step is the actual upgrade. A bot is tied to whatever model its author picked; a reusable prompt is not, so your prompt library survives every model release instead of quietly aging.

#### Who should stay on Poe

Being straightforward about it: Poe is the better product for several real cases.

- **You are exploring rather than producing.** If trying unfamiliar models is the point, breadth beats depth and Poe has far more of it.
- **You want specific niche or experimental models** that a focused workspace does not carry.
- **Character and persona chat is your main use.** That is a first-class use case on Poe and not one Whizi is built for.
- **Your usage is genuinely occasional.** Light, sporadic use can be cheaper on a points allowance than on any flat plan.

And who should move: anyone who has settled into a routine with two or three models, anyone who budgets and wants a predictable line item, anyone currently paying for Poe alongside another AI subscription, and anyone who has noticed themselves picking a cheaper model to save points on work that deserved better.

#### How to decide in a week

Do not decide from a feature table. Run the test.

1. For one week, note which bot answered every message on Poe. If three names cover most of it, the marketplace is not what you are buying.
2. Check where your points go. Points spent on the major models are exactly what a flat plan covers; points spent across many exotic models are the case for staying.
3. Recreate your top two bots as reusable prompts in Whizi and run a real task through both products.
4. Compare on the thing that matters: how much editing the output needed, and whether you hesitated before running a second attempt.

That last question is the one people are surprised by. The difference a meter makes to your own behaviour is bigger than it sounds until you work without one.

Key points:

- Flat monthly pricing against a compute-points wallet, and why that changes how you use the product
- A marketplace of thousands of bots against a small set of first-party models you actually pick between
- How to rebuild the Poe bots you rely on as reusable prompts

Checklist:

- Track which bots you actually use for a week, not which you have bookmarked
- Check whether your points go to the major models or to exotic ones
- Note every time you chose a cheaper model to preserve points
- Recreate your top two bots as reusable prompts before deciding
- Run a real task end to end in both products
- Compare on editing effort, not on catalogue size

FAQ:

Q: Does Whizi have a bot marketplace?
A: No, and that is a deliberate design choice rather than a missing feature. Whizi focuses on first-party models, and the bot equivalent is a reusable prompt: an instruction you can read, edit, and run against any model, whereas a bot is fixed to whichever model its author selected, which ages badly as models are released and retired. Standing instructions can live in a project, where they apply to every chat automatically.

Q: Is Whizi cheaper than Poe?
A: For consistent daily use, generally yes, because flat pricing amortises heavy use while a points wallet charges more for exactly the expensive models you want on important work. For light or sporadic use, a points allowance can work out cheaper. The hidden cost of points is behavioural: people downgrade the model or skip a second attempt to preserve the wallet, and that shows up as slightly worse output rather than as a number.

Q: Can I rebuild my favorite Poe bots in Whizi?
A: Yes, and most take a few minutes. Copy the bot instructions into a reusable prompt, or into a project's instructions if they should apply to a whole body of work, and add an explicit output format. The result is better than the original in one specific way: it runs against any model, so when a new model turns out to be better at that task you switch instantly instead of waiting for someone to rebuild the bot.

Q: Does Whizi have the same models Poe does?
A: Whizi carries the major first-party families, meaning GPT, Claude, Gemini, and leading open models, plus image generation. Poe carries considerably more in total, including experimental and niche models alongside thousands of community bots. If a specific obscure model is essential to your work, check that it is available before switching, since breadth of catalogue is genuinely Poe’s advantage.

Q: What happens to my Poe chat history?
A: Chat history does not transfer between providers, and that is true of any switch between AI products. What is worth carrying over is the prompts, not the conversations: the instructions behind the bots you use, and any prompt you found yourself reusing. Rebuild those as reusable prompts and you have moved the part that had lasting value.

## Model comparisons

### ChatGPT vs Claude: Which AI Should a Beginner Choose?

URL: https://whizi.io/compare/chatgpt-vs-claude/
Updated: August 2026

Quick answer: Use ChatGPT when you want something organised and Claude when you want something written. ChatGPT is stronger on plans, tables, checklists, and anything with a fixed shape. Claude writes better prose and handles difficult tone, so an email needs less fixing. Both have usable free tiers and cost around $20 a month.

A simple, jargon-free comparison between ChatGPT and Claude. Discover which AI is best for your everyday needs and writing tasks.

#### The short answer

Both are excellent and you will not go wrong with either. The difference is one of personality rather than capability.

ChatGPT is quick, structured, and eager. It is very good at organising things: lists, plans, step by step instructions, tables, and anything with a clear shape. It is also the more familiar one, which matters if you are learning.

Claude writes better. That is the single most consistent difference people notice. Ask both to write an email and Claude's version usually needs less fixing before you would send it. It uses fewer filler phrases, sounds less like a press release, and handles awkward situations, like saying no or delivering bad news, with more tact.

If you only take one thing from this page: **use ChatGPT when you want something organised, and Claude when you want something written.**

#### The same prompt, two different answers

Abstract comparisons are hard to judge, so here is what the difference actually looks like.

Ask both: `Write a short message to my landlord asking them to fix the heating, which I have already reported twice.`

ChatGPT will typically produce something correct and slightly formal, often with a subject line, a numbered summary of the previous reports, and a polite closing. Well organised, and it reads a little like a template.

Claude will typically produce something shorter that sounds like a person who is mildly annoyed but staying reasonable. It will usually mention the previous reports in a sentence rather than a list, and it will ask for a specific next step.

Neither is wrong. If you wanted a formal record of the complaint, the first is better. If you wanted to sound like yourself, the second is. That pattern holds across most writing tasks.

Now ask both: `Give me a 30 day plan to learn basic Spanish, 20 minutes a day.`

Here ChatGPT tends to win. It produces cleaner structure, more consistent formatting across all thirty days, and is less likely to drift away from your constraint halfway through. Structure is what it is good at.

#### Which one to open, by task

| What you are doing | Open | Why |
| --- | --- | --- |
| An email that matters | Claude | Sounds human, needs the least editing |
| A difficult message: complaint, apology, saying no | Claude | Best at tone that is direct but not cold |
| A plan, checklist, or schedule | ChatGPT | Cleaner structure, sticks to the format |
| A quick factual question | Either | Both fine, and verify anything important |
| Summarising a long document | Claude | Keeps the nuance rather than flattening it |
| Making a table or organising messy notes | ChatGPT | More reliable at exact formats |
| Helping with code | Either | ChatGPT is quicker, Claude explains better |
| Feedback on something you wrote | Claude | Gives real criticism instead of encouragement |
| Brainstorming a long list of ideas | ChatGPT | Faster and more prolific |
| Anything that happened in the last few weeks | Neither, by default | Use a model with web access, and check the sources |

That last row matters. Both models can be confidently wrong about recent events, and both will occasionally invent a detail with total conviction. For anything current, use a version with web access and open the links yourself.

#### A trick that works better than choosing

Write your draft in one, then paste it into the other and ask: `What are the three weakest things about this? Quote each one. Do not rewrite it.`

This works surprisingly well because a model reviewing something it did not write is far more critical than one reviewing its own output. Ask ChatGPT what is wrong with ChatGPT's draft and it will mostly tell you it is good. Ask Claude, and you get actual objections.

It takes about thirty seconds and it is the closest thing to a free quality upgrade available. Use it on anything that someone else will read.

#### What they cost

Both have free versions that are genuinely usable. The free tiers give you access to capable models with limits on how much you can use them, and for occasional use they may be all you need. The limits tend to appear at the worst moment, which is the honest catch.

The paid plans are around $20 a month each. Subscribing to both is therefore around $40, which is a lot for two chat boxes, and it is what most people are actually trying to avoid when they search for this comparison.

A workspace like Whizi is the third option: both model families in one place for less than a single one of those subscriptions, which also means you can run the paste-into-the-other trick without paying twice. See [pricing](https://whizi.io/pricing) or the [savings calculator](https://whizi.io/tools/calculator) if you already pay for one of them.

#### Things worth knowing before you rely on either

**They both make things up.** Confidently, and with no change in tone when they do. This is most dangerous with specific facts: statistics, dates, quotes, legal details, and citations. If a detail matters, check it somewhere else.

**They do not know what happened recently** unless they can search the web, and even then you should open the sources.

**They agree with you too easily.** If you push back, both will often fold and change their answer even when the original was right. If you want a real opinion, ask before revealing yours.

**Be careful what you paste.** Neither is the right place for passwords, financial details, or other people's personal information. Removing names and identifiers costs you nothing in answer quality.

**Longer is not better.** Both will produce more text when uncertain. If an answer is padded, ask for it again in half the words and it usually improves.

Key points:

- The difference explained without jargon, with real examples of the same prompt in both
- A simple table telling you which one to open for which kind of task
- What each one costs, including the free versions, and how to avoid paying twice

Checklist:

- Use Claude for anything a person will read: emails, messages, writing
- Use ChatGPT for plans, lists, tables, and anything with a fixed shape
- Try the same prompt in both once, so you can see the difference yourself
- Paste a draft from one into the other and ask what is wrong with it
- Check any specific fact, date, or number somewhere else
- Start with the free versions before paying for anything
- If you want both, compare a single workspace against two subscriptions

FAQ:

Q: Which one is easier to use?
A: Both are equally easy, since you simply type in plain English. ChatGPT is more familiar, which helps when you are starting out, and there are far more beginner guides written about it. Claude gives you a slightly cleaner window with fewer features to learn. Neither requires any technical knowledge at all.

Q: Is it expensive to use both?
A: Subscribing to both premium plans costs around $40 a month, which is a lot for two chat boxes. Both have free versions that are genuinely usable if your needs are occasional, with the catch that limits arrive when you are busiest. A combined workspace like Whizi gives you both model families for less than one of those subscriptions, which is the usual answer for people who want both.

Q: Is Claude actually better at writing?
A: For text that a person reads end to end, yes, and it is the most consistently reported difference between the two. Claude uses fewer filler phrases, sustains a voice across a longer piece, and handles awkward tone better. ChatGPT is better when the writing has a fixed structure, such as a report with required sections or a formatted list, because it follows a format more reliably.

Q: Can either one search the internet?
A: Both have versions that can, but neither does so by default in every mode, and this is the most common cause of a confidently wrong answer about something recent. If you are asking about current events, prices, or anything that changes, make sure you are using a version with web access and open the sources it gives you rather than trusting the summary.

Q: Do I have to pick one?
A: No, and picking one is usually the wrong instinct, since their strengths barely overlap. The cheapest approach is to use both free tiers for different tasks. The most convenient is a single workspace containing both, which also lets you draft in one and get criticism from the other, a habit that improves anything important and costs about thirty seconds.

### ChatGPT vs Google Gemini: What’s the Difference?

URL: https://whizi.io/compare/chatgpt-vs-gemini/
Updated: August 2026

Quick answer: Choose Gemini when the question is about something recent or spans a very long document, since it is tied to Google Search and reads far more text at once. Choose ChatGPT for structure: plans, tables, and anything with a fixed format. Both have usable free tiers and paid plans around $20 a month.

A simple, jargon-free comparison of ChatGPT and Google Gemini. Find out which everyday AI assistant is right for your workflow and research.

#### The short answer

Both are excellent general assistants and for most everyday questions you would struggle to tell them apart. Two differences actually matter.

**Gemini knows what is happening now.** Google built it, and it is closely connected to Google Search, so it is the more reliable choice for recent events, current prices, and anything that changed this month. It gives you links, and you can open them.

**Gemini can read far more at once.** This is the difference people underestimate. You can give Gemini a very long document, or several documents, and ask questions across all of it. A 200 page report, a full contract, a semester of lecture slides. Other assistants handle long documents by working through them in pieces, which works less well when your question spans the whole thing.

ChatGPT's advantage is structure and polish. It organises information more cleanly, follows a format more reliably, and is better at anything with a fixed shape: plans, tables, step by step guides, and formatted output.

#### Which one to open, by task

| What you are doing | Open | Why |
| --- | --- | --- |
| Anything from the last few weeks | Gemini | Live search, with links you can check |
| Reading a very long document | Gemini | Handles far more text in one go |
| Several documents at once | Gemini | Can compare across all of them properly |
| A plan, checklist, or schedule | ChatGPT | Cleaner structure, sticks to your format |
| A table or organising messy notes | ChatGPT | More reliable at exact formats |
| Writing an email or message | Either, but see below | Neither is the strongest at this |
| Understanding a photo, chart, or screenshot | Gemini | Strong at images alongside text |
| Brainstorming a list of ideas | ChatGPT | Faster and more prolific |
| Something you will publish | Check with both | Two answers make a wrong one visible |

One honest note on writing. Neither of these is the best at prose that a person reads end to end, and if that is your main need, Claude is the model people consistently prefer. See [ChatGPT vs Claude](https://whizi.io/compare/chatgpt-vs-claude) for that comparison.

#### What "reads more at once" actually means

This sounds technical and it changes what you can do, so it is worth explaining plainly.

Every AI assistant has a limit on how much text it can hold in mind during one conversation. When a document exceeds that limit, the tool has to work through it in chunks, which is fine for a question about one section and unreliable for a question about the whole thing.

Gemini's limit is much higher than most, so questions like these actually work: `Do any two sections of this contract contradict each other?` `Across these twelve papers, where do the findings disagree?` `Is this term defined anywhere in the document?` Each of those requires everything to be present at once, and none of them work well with chunking.

If you regularly deal with long PDFs, reports, or transcripts, this single difference is likely to matter more to you than anything else on this page.

#### Using them together, which is the actual answer

The pattern that works is to let each one do what it is good at, in sequence.

Research in Gemini: `Find out what changed about [topic] in the last six months. Give me a source link and date for each finding, and say clearly if you cannot verify something.`

Then organise in ChatGPT: `Turn the findings above into a one page summary with a table of what changed, when, and why it matters.`

That takes one extra step and produces something noticeably better than either could alone, because the research half and the structuring half genuinely want different strengths.

The friction is that these are separate products, so you paste between two tabs and the second tool has no idea what the first one found beyond what you copied. A workspace that holds both in the same conversation removes that step and keeps the context, which is what Whizi is for.

#### What they cost

Both have free versions that are genuinely useful, with limits on the more capable models that tend to bite when you are busiest.

ChatGPT Plus is around $20 a month. Google's paid plan is around $20 a month too, but it arrives bundled with 2TB of cloud storage as part of Google One. That is worth checking honestly: if you would buy that storage anyway, the AI is arriving cheaply. If you are on the free tier using a fraction of it, part of your subscription is buying capacity you will never fill.

Paying for both is around $40 a month. Whizi puts both model families plus Claude in one workspace for less than a single one of those plans. See [pricing](https://whizi.io/pricing) or the [savings calculator](https://whizi.io/tools/calculator).

#### Things worth knowing before you rely on either

**Both make things up.** Confidently, with no change in tone. Search access reduces this considerably but does not remove it, and the harder error is a real source attached to a claim it does not support. Open the links.

**Web access is not always on.** Depending on which mode you are in, an assistant may answer from memory rather than searching. If your question is about something current, make sure it actually looked.

**They agree with you too readily.** Push back and both will often abandon a correct answer. Ask before revealing what you think.

**Be careful what you paste.** Neither is the place for passwords, financial details, or other people's personal information. Removing names costs nothing in answer quality.

**Check the summary against the source.** For long documents especially, ask for the exact quote supporting an answer. It takes seconds and it is the only reliable way to know the tool read your document rather than answering from general knowledge about documents like yours.

Key points:

- The two real differences: what they know about right now, and how much they can read at once
- A table telling you which to open for which task, with the reason
- What each costs, including the storage bundled into Google's plan

Checklist:

- Use Gemini for anything recent, and open the links it gives you
- Use Gemini for long documents and for questions spanning several files
- Use ChatGPT for plans, tables, and anything with a fixed format
- For writing that matters, consider Claude rather than either of these
- Research in one, organise in the other, rather than picking a winner
- Ask for the exact quote when working from a document
- Check whether you would pay for Google's bundled storage on its own

FAQ:

Q: Are ChatGPT and Gemini the same thing?
A: No. They are separate products from different companies, OpenAI and Google, built on different underlying models. They feel similar because they do similar things, and the practical differences are that Gemini has closer access to live search and can read far more text at once, while ChatGPT is better at structure and formatting.

Q: Why not just use Google instead of Gemini?
A: Regular Google gives you a list of links to read yourself, which is the right tool when you want to see the sources and judge them. Gemini reads them and answers your question directly, which is faster but adds a step where something can be misread or overstated. The reasonable habit is to use Gemini for the answer and open its links when the answer matters.

Q: Which is better for students?
A: Gemini, mainly because of the long-document handling. Lecture slides, textbook chapters, and reading lists can go in as whole files rather than being fed through in pieces, which makes questions across a whole module actually work. ChatGPT is better for structured study plans and formatted revision material, and using both free tiers costs nothing.

Q: Which is better for work?
A: It depends on your work, which is the honest answer. If you read long documents, do research, or need current information, Gemini. If you produce structured output such as plans, tables, and reports, ChatGPT. If you write things other people read, neither is the strongest and Claude is usually preferred. Most professional work touches at least two of those.

Q: Do I have to pay for either one?
A: No. Both free tiers are genuinely usable and cover a lot of everyday use. What you hit is limits on the more capable models, and they reliably arrive during the week you are busiest. Paying makes sense at the point where being interrupted mid-task costs you more than the subscription does, which is a personal calculation rather than a rule.

## Product guides

### Billing, trials and how to cancel your Whizi subscription

URL: https://whizi.io/docs/billing-and-cancellation/
Updated: August 2026

Quick answer: Cancel a Whizi subscription from your account settings if you subscribed on whizi.io, or through Apple or Google if you subscribed inside the iOS or Android app. Cancellation takes effect at the end of the current paid term, so access continues until then. The 7 day trial for $0.99 converts unless you cancel first.

How the $0.99 trial converts, where to cancel depending on how you subscribed, when changes take effect, and what happens to your account afterwards.

#### The short answer

Cancel from your account settings if you subscribed on the web. Cancel through Apple or Google if you subscribed inside the iOS or Android app. There is no way to cancel a store subscription from the Whizi website, because the store owns that billing relationship.

Cancellation takes effect at the end of the current paid term. You keep full access until then rather than losing it the moment you cancel.

#### The trial

New accounts get 7 days for $0.99. It is a paid trial, not a free one, and it converts to the plan you chose at signup unless you cancel before the 7 days are up.

Cancelling during the trial keeps your access for the remainder of the 7 days and then stops, with nothing further charged. You do not lose the days you paid for by cancelling early, so if you have decided, there is no reason to wait until the last day.

The trial gives you the full plan you selected, not a reduced version of it. A Powerhouse trial includes [Claude Opus 5](https://whizi.io/docs/using-claude-in-whizi), [GPT-5.6 Sol](https://whizi.io/docs/using-gpt-in-whizi), side-by-side comparison and the media generation, which is the only honest way to evaluate whether the top tier is worth it.

#### Where to cancel, by how you signed up

| You subscribed | You cancel | Notes |
| --- | --- | --- |
| On whizi.io | Account settings, on the web | Takes effect at the end of the current paid term |
| In the iOS app | Apple subscription settings | Must be cancelled before the renewal deadline Apple shows |
| In the Android app | Google Play subscription settings | Must be cancelled before the renewal deadline Google shows |

This split catches people out, so it is worth being explicit: deleting the app does not cancel a store subscription, and neither does deleting your Whizi account. A store subscription keeps billing until it is cancelled with the store that sold it.

Store refunds are handled under the applicable store terms, by Apple or Google rather than by Whizi.

#### When plan changes take effect

**Upgrading is immediate.** The new models and the larger credit allowance become available straight away, which is why starting a trial one tier low and moving up mid-week costs you nothing.

**Downgrading waits.** It takes effect at the end of the current paid term, so you keep the tier you already paid for until that period runs out. Your conversations are unaffected: downgrading changes which models you can send new messages to, not what you can read.

**Cancelling waits too.** Access continues to the end of the paid term and then stops.

Yearly billing charges once for the year, at $131.88 for Starter, $239.88 for Pro or $419.88 for Powerhouse. The per-month figures quoted for yearly plans are that total divided by twelve, not a separate monthly charge.

#### Deleting your account

Account deletion is separate from cancellation, and it is the stronger action. Cancelling stops billing and ends access at the end of the term; deletion removes the account and its contents.

Deletion covers your conversations and messages, uploaded and generated media, saved memory, entitlement and usage state, guest identity links, the authentication account, and local data on the device completing the deletion. There is a limited retention window described in the privacy policy for the records that have to survive briefly.

You can start deletion at [whizi.io/account-deletion](https://whizi.io/account-deletion). If you also have a store subscription, cancel that with Apple or Google as well, because deleting the account does not stop store billing. The steps and the full list of what is removed are on [delete your account](https://whizi.io/docs/delete-your-account).

Key points:

- Where to cancel, which depends entirely on where you subscribed
- When an upgrade, a downgrade and a cancellation each take effect
- What the 7 day trial for $0.99 actually commits you to

Checklist:

- Web subscriptions cancel in account settings
- App Store and Play subscriptions cancel with Apple or Google
- Deleting the app does not cancel a store subscription
- Cancellation takes effect at the end of the current paid term
- Upgrades take effect immediately, downgrades at term end
- The $0.99 trial converts after 7 days unless cancelled
- Account deletion is separate from cancellation

FAQ:

Q: How do I cancel my Whizi subscription?
A: If you subscribed on whizi.io, log in and cancel from account settings; it takes effect at the end of the current paid term and you keep access until then. If you subscribed inside the iOS or Android app, you must cancel through Apple or Google instead, because the store handles that billing and Whizi cannot cancel it for you.

Q: Will I be charged after the $0.99 trial?
A: Yes, the trial converts to the plan you selected at signup after 7 days unless you cancel first. Cancelling during the trial keeps your access for the rest of the 7 days and charges nothing further, so there is no penalty for deciding early.

Q: Does deleting the app cancel my subscription?
A: No. This is the most common billing mistake with any app-store subscription: removing the app removes your access to it, but the store keeps billing until the subscription itself is cancelled in Apple or Google subscription settings. Deleting your Whizi account does not stop store billing either.

Q: What happens to my conversations if I cancel?
A: They remain in your account, which stays in place unless you separately delete it. If you resubscribe later, your history is still there. If you want the data gone as well as the billing stopped, use account deletion rather than cancellation, since those are two different actions.

Q: Can I switch from monthly to yearly billing?
A: Yes, and it is roughly a 30 percent discount at every tier: Pro drops from $29.99 to $19.99 per month equivalent, billed once at $239.88. Since upgrades take effect immediately, the switch applies without waiting out the current month.

### Whizi chat import errors: why a ChatGPT or Claude import is rejected

URL: https://whizi.io/docs/chat-import-errors/
Updated: August 2026

Quick answer: A Whizi import fails in one of two places, and the message tells you which: the browser parser reading your export file, or the upload posting conversations to the server. The file Whizi reads is conversations.json inside a ChatGPT or Claude export. Re-running an import never duplicates chats, so retry first.

Every error message the ChatGPT and Claude import can show, the condition that triggers each one, and the fix. Plus why re-running a failed import never duplicates your chats.

#### The short answer

If Whizi rejected your import, check the file you picked first: the only file Whizi reads is `conversations.json` inside a ChatGPT or Claude export, so unzip the export and pick that file rather than the zip itself. Past that, an import fails in one of two places and the wording of the message tells you which: the browser parser that reads your export file, or the import endpoint that receives the conversations the parser produced. File errors are about the file you picked. Upload errors are about the network or the session. Nothing in between is ambiguous, because each message is a fixed string tied to one condition.

The second thing to know before you retry anything: re-running an import never duplicates chats. Every imported conversation is keyed by account, source, and source id, so importing the same export twice, resuming after a dropped connection, or clicking Import twice all land the same chats once.

| What you saw | Where it failed | What to do |
| --- | --- | --- |
| A message about the export format or conversations.json | Reading your file, before anything uploaded | Pick a different file out of the export, see the table below |
| A message about your session or your connection | Uploading, part way through | Sign in again or reconnect, then run the import again |
| A count of conversations that could not be imported | Some batches failed every retry | Run the import again, it picks up only what is missing |
| Nothing, but it is taking a long time | Rate limiting, working as designed | Wait, the importer retries on its own |

#### Errors from reading your export file

These come from the browser, before a single byte leaves your machine. The export file itself is never uploaded: the browser parses it and posts only normalized conversations, so an error here means the file never got far enough to be sent.

| Message | Trigger | Fix |
| --- | --- | --- |
| `That file is not a ChatGPT or Claude export. Look for conversations.json inside the export zip.` | The parser cannot identify either export format, or a zip yields an unrecognized payload | Open the export archive and pick `conversations.json` specifically |
| `There is no conversations.json inside that zip. Unzip it and pick the file directly.` | The zip you picked has no matching entry | Unzip it yourself, find `conversations.json`, and pick that file |
| `That file is not valid JSON, so there was nothing to read.` | The picked file fails to parse as JSON | Download the export again from ChatGPT or Claude and pick `conversations.json` from it |
| `That export has no conversations with text in them.` | The export parses but yields zero conversations | The export is genuinely empty of text transcripts. Nothing to fix on the Whizi side |
| `That file is too large for the browser to open. Pick the export .zip instead, which is read one file at a time.` | The picked raw .json file is over the browser read cap | Pick the `.zip` instead. A raw `.json` export is accepted up to 300 MB; the `.zip` path has no ceiling of its own, because only the conversations.json entry is decompressed |
| `Could not read that file. Please try again.` | File reading threw for any other reason | Genuinely a retry case. If it repeats on the same file, pick the export `.zip`, which is read one entry at a time |

The first message is the broad one, because one string covers two conditions: an export format the parser cannot identify, and a zip that yields an unrecognized payload. Both land on the same fix, which is to unzip the export and pick `conversations.json` directly.

#### Errors from the upload itself

Once the file parses, the importer posts the conversations in batches. Two things can go wrong at that stage, and both are recoverable.

| Message | Trigger | Fix |
| --- | --- | --- |
| `Your session ended. Sign in again, then start the import.` | The token fetch returned nothing part way through the import | Sign in again, then start the import over. Whatever already landed stays, and the re-run skips it |
| `The import could not be finished. Check your connection and try again.` | Any other upload failure during the import run | Check the connection and run it again |

A third outcome is not an error at all but a summary: `{count} could not be imported. Run the import again to pick them up.` That is the count of conversations whose batches failed every retry. The instruction in the message is the whole fix, and the dedupe key described in the short answer is why the second run costs you nothing.

#### Errors the import endpoint returns

These are the codes and strings the import route itself produces. The website mirrors the same caps and validates before posting, so they matter most when something is posting to the endpoint directly or when a batch is malformed.

| Code | Message | Trigger |
| --- | --- | --- |
| `invalid_import` | `A JSON object body is required.` | The import body is not an object |
| `invalid_source` | `An import source of chatgpt or claude is required.` | The source field is missing or is not one of the two accepted sources |
| `empty_import` | `At least one conversation is required.` | The conversations array is missing or empty |
| `too_many_conversations` | `This batch has {count} conversations, over the 25 per request limit.` | More than 25 conversations in one request |
| `too_many_messages` | `Conversation {n} carries {count} messages, over the 400 per request limit. Send the rest as an append.` | More than 400 messages in one conversation in one request |
| `invalid_import` | `Conversation {n} is not an object.` | A conversations array entry is not an object |
| `invalid_import` | `Conversation {n} has no messages array.` | A conversation carries no messages array |
| `invalid_import` | `Conversation {n} has an invalid append target.` | The appendTo field is neither a string nor null |
| `invalid_message` | `Conversation {n} message {m} is not an object.` | A message entry is not an object |
| `invalid_message` | `Conversation {n} message {m} needs a role of "user" or "assistant".` | A message has a role outside those two |
| `invalid_message` | `Conversation {n} message {m} has no text.` | The message content is not a string |
| `request_too_large` | `Request is too large.` | The JSON body exceeds the route byte cap, either by declared Content-Length or by bytes actually read |

The caps behind those messages: 25 conversations per request, 400 messages per conversation per request, and a request body of at most 2,000,000 bytes. The website importer splits at exactly 25 and 400 and holds each request under 1,200,000 bytes, deliberately below the 2,000,000 byte cap to leave room for the request envelope. A conversation longer than 400 messages is not rejected, it is sliced: the remainder is sent as append slices onto the chat the first slice created.

#### A large import pauses for 20 seconds at a time

Importing conversations is rate limited to 20 requests per minute per user. A large export will hit that ceiling, and that is expected rather than broken.

You will not see an error when it happens. An import that hits HTTP 429 is not surfaced: the client waits 20 seconds and retries, up to 4 attempts per batch. So an import that looks stalled for twenty seconds at a time on a big export is doing exactly what it should. Leave it running.

If batches still fail after those 4 attempts, they land in the `{count} could not be imported. Run the import again to pick them up.` summary, and running the import a second time is the fix rather than support.

#### What the import drops on purpose

Some content is dropped by design, and an import that succeeds can still look incomplete next to the original.

For a ChatGPT export, only the branch that was on screen is imported. Regenerated replies and discarded drafts are not replayed as messages. Non-transcript content is skipped too: hidden reasoning, analysis-tool code and output, browsing scratchpads, custom-instruction blocks, and function calls.

For a Claude export, thinking blocks, tool calls, and attachments are dropped. Only text blocks are imported.

Conversations in the file holding no readable text are counted and reported as skipped rather than silently dropped, so a skipped count is information, not a failure.

A conversation with no usable title becomes `Imported chat`, and a title longer than 120 characters is trimmed. An imported message longer than the content cap is truncated rather than rejected, so one oversized message does not cost you the rest of that conversation.

Finally, timestamps: an imported message dated before 2015-01-01, or dated in the future, is discarded as a timestamp and the row falls back to the import time. A chat that lands with the date you ran the import, instead of its original date, hit that rule.

#### Why running the import again is safe

The dedupe key in the short answer has a fallback. When an export gives no stable conversation id, the key becomes a hash of the transcript itself, so a re-run still matches those conversations rather than adding second copies.

One consequence to know about: deleting an imported chat and then importing the same export again brings the chat back rather than reporting a duplicate. If you deliberately deleted an imported conversation, do not re-run that export unless you want it returned.

Key points:

- Format errors fire in your browser, before any of the export is uploaded
- Which errors come from your file and which come from the upload
- What the import drops on purpose, and why a skipped count is not a failure

Checklist:

- Pick conversations.json out of the export, not the export zip itself
- A raw .json export is accepted up to 300 MB, past which use the .zip
- Session and connection errors mean sign in or reconnect, then re-run
- A count that could not be imported means run the import again
- After any import error, run the same import again before anything else
- Deleting an imported chat and re-importing brings it back
- A large import pausing for 20 seconds at a time is rate limiting, not a fault
- ChatGPT regenerated branches, and Claude thinking blocks and attachments, are dropped by design

FAQ:

Q: Whizi says my ChatGPT export is not a ChatGPT export. Why?
A: The message is `That file is not a ChatGPT or Claude export. Look for conversations.json inside the export zip.` and it fires when the parser cannot identify either format in the file you picked, or when a zip yields an unrecognized payload. The file Whizi reads is conversations.json, so unzip the export and pick that file directly. If you picked the zip and got `There is no conversations.json inside that zip. Unzip it and pick the file directly.` instead, the zip has no matching entry at all and the export needs downloading again.

Q: My import stopped part way through. Will running it again duplicate my chats?
A: No. Re-running is the intended recovery for a dropped connection, an ended session, or a failed batch: the second run lands the same chats once and picks up only what is missing. There is no cleanup step to do first.

Q: Why did only some of my conversations import?
A: Two different counts can appear, and only one of them is a failure. `{count} could not be imported. Run the import again to pick them up.` counts batches that failed every retry, and re-running the import is the fix. A skipped count is the other one, and it reports conversations in the file that hold no readable text rather than conversations that went wrong.

Q: The import has been sitting there for a while. Is it broken?
A: Watch whether it moves again rather than how long it has been going. A pause of roughly 20 seconds followed by more progress is the rate limiter, because the client waits 20 seconds and retries instead of showing you anything. An import that is actually finished says so, either with one of the error strings on this page or with the `{count} could not be imported. Run the import again to pick them up.` summary. Silence plus periodic progress means leave it alone.

Q: Does my ChatGPT or Claude export file get uploaded to Whizi?
A: No. The export file itself never reaches the server. Your browser reads it, parses it, and posts only the normalized conversations, which is also why every format error on this page happens before anything is sent. The full data handling detail is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

### How to chat with a PDF online in Whizi

URL: https://whizi.io/docs/chat-with-pdf/
Updated: August 2026

Quick answer: Attach the PDF by dragging it into the message box or clicking the paperclip icon, then ask your question in plain language. Any model on your plan can read it, and you can switch models against the same file without uploading it again. Ask the model to quote the passage it used before relying on an answer.

Upload a PDF to Whizi, ask questions across its full content, compare answers between Claude and Gemini, and export the summary in seconds.

#### The short answer

Drag the PDF into the message box or click the paperclip icon to attach it, then ask your question in plain language. Any model on your plan can read it, and you can switch models against the same file without uploading it again.

Do one thing before you rely on an answer: ask the model to quote the passage it used. The failure that costs people money is not a wrong answer, it is a confident answer drawn from training data instead of from your file, and a locating question catches it in about ten seconds. The rest of this page is how to do that well.

#### First, confirm it is reading your document

The most dangerous failure in document chat is not a wrong answer, it is a confident answer drawn from training data instead of from your file. It happens most with documents that resemble something common: a standard commercial lease, a well known regulation, a widely discussed paper. The model knows the genre well enough to produce a plausible answer without consulting your specific text, and nothing in the output signals that it did so.

One habit removes this. Before asking your real question, ask a locating question.

`Quote the passage in the attached document that discusses [topic], and give the page or section it appears in. If it is not in the document, say so.`

If it returns a real quote, it is reading your file. If it produces a generic paraphrase or cannot find something you know is there, the extraction has failed and every subsequent answer is suspect. That takes ten seconds and it is the difference between a tool and a liability.

Carry the habit into the real questions too: `Quote the exact text that supports your answer` should appear in any prompt whose answer will be relied on.

#### Choosing a model for the document

| Document | Model | Why |
| --- | --- | --- |
| 100+ page reports, filings, whole contract sets | Gemini | Largest context window, so the document is genuinely present rather than chunked |
| Contracts, policy, anything where qualification matters | Claude | Best at nuance and at telling you what a document does not say |
| Invoices, forms, short structured files | GPT | Fast, and the most reliable at strict extraction into tables or JSON |
| Papers with figures, scanned pages, diagrams | Any multimodal model | Reads the page as an image where text extraction falls short |

The [context window](https://whizi.io/docs/context-windows-compared) point is worth spelling out. When a document exceeds what a model can hold, it has to be processed in pieces, and questions that require the whole document at once (does anything here contradict clause 14, is this term defined anywhere, which of these three sections is inconsistent) become unreliable. That is the specific reason to reach for the large-context model on long files rather than a general preference.

You can switch models without re-uploading, so read with one and extract with another against the same file. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### Prompts by document type

**Contracts and agreements**

`From the attached agreement, extract into a table: term and renewal, notice period for termination, payment terms, liability cap, indemnities, governing law, assignment, and any clause that survives termination. For each, give the clause number and quote the operative sentence. Then list separately the obligations that fall on us rather than on them.`

Follow up with the question that actually matters: `What is in this agreement that is unusual compared to standard terms for this type of contract, and what is conspicuously absent?` Absence is where the risk usually lives, and it is the thing a keyword search can never find.

**Research papers**

`For the attached paper, return: the research question, the study design, the sample and population, the primary outcome, the headline result with its effect size, the stated limitations, and the funding source. Then tell me the three claims in this paper that would need independent verification before I cite them.`

**Long reports and filings**

`Summarize the attached report in 10 bullets ordered by importance rather than by document order. Then list: the three numbers a decision maker would care about with their page references, anything the report presents as fact without a source, and any place where the summary or executive section disagrees with the detail later in the document.`

That last check finds real problems surprisingly often. Executive summaries are written early and edited less than the sections they summarise.

**Comparing two documents**

`Compare contract-a.pdf and contract-b.pdf clause by clause. Return a table of every substantive difference: topic, what A says, what B says, and which favours us. Ignore formatting and numbering differences. List separately anything present in one and missing from the other.`

**Turning it into something you can use**

`Rewrite the findings for an executive audience in 150 words. Lead with the decision required. No jargon that is not defined. Flag anything I should verify before circulating this.`

#### Troubleshooting

**"I cannot see a document."** Confirm the file finished uploading, then reference it explicitly by name in your message. On very long threads, re-attaching or starting a fresh chat with just the file is faster than arguing about it.

**Scanned PDF, garbled or missing text.** The document is an image and OCR is doing the work. Whizi handles OCR automatically on most plans and you can also ask explicitly for OCR before answering. Accuracy drops with low resolution, unusual fonts, handwriting, and dense tables. Because a misread digit is invisible in a fluent answer, verify every figure from a scanned document against the original page.

**Tables come out wrong.** PDF tables are notoriously badly structured underneath. Ask for the table to be reproduced verbatim first, check it against the page, and only then ask for analysis. For heavy table work, a multimodal model reading the page as an image is often more reliable than text extraction.

**It missed something you know is in there.** Ask the locating question. If the model cannot quote a passage you can see, the problem is extraction rather than reasoning. Try another model, or upload just the relevant pages.

**The file is too large.** Use the large-context model, or split by purpose rather than by page count. Uploading the three sections you actually need beats uploading 400 pages and hoping.

**Answers get vaguer as the conversation goes on.** Long threads accumulate context that competes with the document. Start a fresh chat with the file and a concise statement of what you need.

#### Verify before you rely on it

Three checks, in order of importance.

**Require quotes on anything consequential.** A quoted passage is checkable in seconds; a paraphrase is not. This single practice eliminates most of the risk in document work.

**Cross-check with a second model.** Ask the same question of a different model against the same file. Agreement is meaningful evidence; disagreement tells you exactly where to look. Whizi's [side-by-side comparison](https://whizi.io/docs/compare-models-side-by-side) exists for this.

**Ask what it is uncertain about.** `Which of your answers above are you least confident in, and what in the document is ambiguous?` Models are imperfect at self-assessment, but the ambiguities they surface are usually genuine ambiguities in the source, which is useful information about the document itself.

And the standing rule for anything that carries consequences, meaning legal, financial, medical, or compliance material: the model finds the passage, you make the judgment. It is a reading assistant, not an adviser, and the accountability for a decision made on its output is entirely yours.

Key points:

- The verification habit that catches the most dangerous failure: an answer from memory rather than from your file
- Prompts for contracts, research papers, reports, and multi-document comparison
- Troubleshooting for scans, tables, and documents the model claims not to see

Checklist:

- Ask a locating question first to confirm the model is reading your file
- Require an exact quote for every consequential answer
- Use the large-context model for long documents so nothing is chunked away
- Ask what is conspicuously absent, not just what the document says
- Verify figures from scanned documents against the original page
- Reproduce tables verbatim and check them before analysing them
- Cross-check anything important against a second model

Steps:

1. Open a new chat: In Whizi, start a new conversation.
2. Upload the PDF: Drag the file into the message box or click the paperclip icon to attach it.
3. Pick the right model: Choose Gemini or Claude for long PDFs. Use GPT for shorter files or quick extraction.
4. Ask a locating question first: Ask it to quote the passage on your topic, so you can confirm it is reading the document rather than answering from memory.
5. Ask your real question: Ask for a summary, an extraction, or the specific thing you opened the document for.
6. Compare answers across models: Re-run anything consequential against a second model to cross-check.
7. Export the result: Copy the answer, paste it into your draft, or save the chat for reference.

FAQ:

Q: What is the maximum PDF size?
A: Limits depend on your plan and the model, and the real constraint is the model context window rather than a file size cap. [Gemini in Whizi](https://whizi.io/docs/using-gemini-in-whizi) supports the largest documents, which is why it is the recommendation for long reports, filings, and multi-contract sets. Above that, upload the sections you actually need rather than the entire document, since targeted context usually produces a sharper answer as well.

Q: Does Whizi store my PDFs?
A: Files are stored securely in your workspace, are not used to train models, and can be deleted at any time. Deleting a file does not remove what was already discussed about it in the conversation, so if the document is sensitive enough to warrant deletion, delete the chat too. Each provider’s data handling policy is available for review before you enable that model.

Q: Can I chat with multiple PDFs at once?
A: Yes, and comparison across documents is where this is most valuable. Upload several files to the same thread and reference them by name in your prompts, otherwise the model chooses which one to answer from without telling you. Clause-by-clause contract comparison, consolidating findings across reports, and finding contradictions between documents all need everything present at once, which is another reason to use the large-context model.

Q: Does it work with scanned PDFs?
A: Yes. Scanned files are processed with OCR automatically on most plans, and you can ask the model to OCR and then answer. Accuracy is good on clean scans and degrades with low resolution, unusual fonts, handwriting, and dense tables. Since a misread character produces a wrong number inside an otherwise fluent answer, check any figure taken from a scanned document against the original page.

Q: How do I know the answer actually came from my document?
A: Require a quote. Ask the model to quote the exact passage supporting its answer along with the page or section, and to say plainly if the information is not in the document. Answers drawn from training data rather than from your file are the most dangerous failure in document chat, precisely because they are plausible for common document types, and demanding a verifiable quote is the only reliable way to tell the difference.

### The cheapest AI models per answer in Whizi, and what one credit buys

URL: https://whizi.io/docs/cheapest-models-per-answer/
Updated: August 2026

Quick answer: The cheapest priced row in the Whizi cost index is Ling from inclusionAI, at $0.000053 to answer once. It charges one credit per message, as does every other row at the bottom of the ladder, and no message is ever charged less than one credit. A standard answer is 1,000 input tokens plus 500 output tokens.

The cheap end of the Whizi cost index in dollars per standard answer, cross-referenced with what each row charges in credits and which plan unlocks it.

#### The short answer

The cheapest priced row in the Whizi cost index is Ling from inclusionAI, at $0.000053 to answer once. It charges one credit per message, as does every other row at the bottom of the ladder, and the floor on a turn guarantees nothing is ever charged less than one.

One cheap row per plan, as a lead into the full tables below:
| Model | Provider | Cost per answer | Context | Credits | Plan |
| --- | --- | --- | --- | --- | --- |
| Ling-3.0-flash | inclusionAI | $0.000053 | 262K | 1 | Powerhouse |
| Llama 3.3 70B Instruct | Meta | $0.00026 | 131K | 1 | Pro |
| GPT-5.6 Luna | OpenAI | $0.0008 | 1M | 1 | Starter |

A standard answer in the index is 1,000 input tokens plus 500 output tokens, so list rates from different providers become comparable. The source prices were fetched from OpenRouter on 2026-08-20. The index prices 100 rows across 29 providers, and 89 of those rows carry a Whizi credit charge. The full dataset is published as the [AI model cost index](https://whizi.io/tools/model-cost-index).

This page is the cheap end of that index. For the catalogue by plan see the [model list](https://whizi.io/docs/models), and for the credit ladder itself see the [credits reference](https://whizi.io/docs/credits).

#### Rows under a tenth of a cent per answer

A selection of the rows that answer once for less than $0.001 at the standard answer size. All but one of the rows printed here charge a single credit.

| Model | Provider | Cost per answer | Context | Credits | Plan |
| --- | --- | --- | --- | --- | --- |
| Ling-3.0-flash | inclusionAI | $0.000053 | 262K | 1 | Powerhouse |
| Nex-N2-Mini | Nex Agi | $0.000075 | 262K | 1 | Powerhouse |
| Solar Pro 4 | Upstage | $0.00009 | 524K | 1 | Powerhouse |
| Qwen3.7 Flash | Qwen | $0.000095 | 1M | 1 | Powerhouse |
| Granite 4.1 8B | IBM | $0.0001 | 131K | 1 | Powerhouse |
| Laguna XS 2.1 | Poolside | $0.00012 | 262K | 1 | Powerhouse |
| Phi 4 | Microsoft | $0.00014 | 16K | 1 | Powerhouse |
| Nemotron 3.5 Lightning | NVIDIA | $0.00018 | 262K | 1 | Powerhouse |
| Llama 3.3 70B Instruct | Meta | $0.00026 | 131K | 1 | Pro |
| DeepSeek V4 Flash 0731 | DeepSeek | $0.00028 | 1.31M | 1 | Powerhouse |
| Gemini 2.5 Flash Lite | Google | $0.0003 | 1M | 1 | Pro |
| DeepSeek V3.2 | DeepSeek | $0.000469 | 164K | 1 | Pro |
| Qwen3 Coder Next | Qwen | $0.00052 | 262K | 1 | Powerhouse |
| Llama 4 Maverick | Meta | $0.0006 | 1M | 1 | Pro |
| DeepSeek V3.1 | DeepSeek | $0.000725 | 164K | 1 | Pro |
| Codestral 2508 | Mistral | $0.00075 | 256K | 1 | Powerhouse |
| Qwen3.6 Flash | Qwen | $0.00075 | 1M | 1 | Powerhouse |
| MiniMax M2 | MiniMax | $0.000765 | 205K | 1 | Powerhouse |
| GPT-5.6 Luna | OpenAI | $0.0008 | 1M | 1 | Starter |
| GPT-5.4 Nano | OpenAI | $0.000825 | 400K | 2 | Powerhouse |
| MiniMax M3 | MiniMax | $0.0009 | 1M | 1 | Powerhouse |
| Qwen3.7 Plus | Qwen | $0.00096 | 1M | 1 | Pro |

Several of the cheapest rows carry a million tokens or more, and [the largest window in the whole price index](https://whizi.io/docs/context-windows-compared) belongs to a row near the bottom of the cost list. Cheap and small are not the same axis here. The smallest window printed above is 16K.

Most of these rows sit on the most expensive plan, for the reason set out below.

#### The next band, up to the index median

The median priced row is a Kimi reasoning row at $0.00185 per standard answer. A sample of the rows between a tenth of a cent and that median:

| Model | Provider | Cost per answer | Context | Credits | Plan |
| --- | --- | --- | --- | --- | --- |
| Gemini 3.1 Flash Lite | Google | $0.001 | 1M | 1 | Powerhouse |
| Muse Glimmer 30B | Meta | $0.0011 | 131K | 2 | Powerhouse |
| Mistral Large 3 2512 | Mistral | $0.00125 | 262K | 2 | Pro |
| GLM 4.7 | Z.ai | $0.001275 | 205K | 2 | Powerhouse |
| Gemini 3.7 Flash | Google | $0.001313 | 1M | 2 | Pro |
| Mistral Medium 3.1 | Mistral | $0.0014 | 131K | 2 | Pro |
| GLM 4.6 | Z.ai | $0.0015 | 205K | 1 | Powerhouse |
| Gemini 2.5 Flash | Google | $0.00155 | 1M | 2 | Pro |
| Gemini 3.5 Flash Lite | Google | $0.00155 | 1M | 2 | Powerhouse |
| GLM 5 | Z.ai | $0.00156 | 205K | 2 | Pro |
| Kimi K2 0711 | Moonshot | $0.00172 | 131K | 10 | Powerhouse |
| Kimi K2 Thinking | Moonshot | $0.00185 | 262K | 10 | Powerhouse |

The credit column and the dollar column do not cover the same set of rows. Eleven of the 100 priced rows carry no credit charge at all: they are priced for comparison but are not offered as their own metered row in the product. A dollar figure in the index therefore does not by itself mean the row can be opened in the picker.

#### Why most of the cheapest rows are Powerhouse

The plan column above is not a pricing decision, it is a fallback. Model access resolves by identifier: the four Starter identifiers first, then the agent personas, then an explicit list of 37 further identifiers for Pro, and anything else returns Powerhouse. A cheap row that nobody ever promoted into the Pro list therefore sits on the top tier by default, which is why the bottom of the cost list reads as a Powerhouse column. The fallback itself is documented on the [model list](https://whizi.io/docs/models).

| Plan | What it reaches at the cheap end |
| --- | --- |
| Free | No credit allowance. Seven messages, lifetime, on any text model |
| Starter | Four identifiers in total: Auto, Whizi AI, GPT-5.6 Luna, Gemini 3 Flash. All of them cost 1 credit |
| Pro | Starter plus 37 more, including the DeepSeek, Llama, Gemini Flash, Mistral and Qwen rows marked Pro above |
| Powerhouse | Everything else in the catalogue, which is where the smaller-lab rows at the very bottom of the cost list sit |

The free tier is the odd one out: free accounts are not model-gated. The chat path returns before the model gate for free accounts, so a prospect can try any text model inside a lifetime allowance of seven messages that never resets. The cheapest row in the index and the dearest row in the catalogue are equally reachable on it.

Starter is the narrow one. It reaches four identifiers, every one of them charges a single credit, and its 400 credit allowance therefore equals 400 messages. There is no Anthropic model on Starter, and no access to the cheap smaller-lab rows in the tables above.

#### What one credit buys

A credit is defined as one message on the house model, which runs on the base OpenAI model pinned at one credit. A build test fails if either of them ever leaves that rung, because moving either one re-denominates the entire ladder.

The charge is a fixed integer per model id. At this end of the catalogue that means a 1 credit row costs one credit whatever the input or output length: paste a long document into it, let the reply run long, and the turn is still charged one.

| Row | Cost per answer | Cost per 1,000 answers | Credits |
| --- | --- | --- | --- |
| Nemotron 3.5 Lightning | $0.00018 | $0.18 | 1 |
| Llama 3.3 70B Instruct | $0.00026 | $0.26 | 1 |
| DeepSeek V4 Flash 0731 | $0.00028 | $0.28 | 1 |
| Gemini 2.5 Flash Lite | $0.0003 | $0.30 | 1 |
| DeepSeek V3.2 | $0.000469 | $0.469 | 1 |
| Llama 4 Maverick | $0.0006 | $0.60 | 1 |
| GPT-5.6 Luna | $0.0008 | $0.80 | 1 |
| Qwen3.7 Plus | $0.00096 | $0.96 | 1 |

Turned into an allowance, the bottom rung is the one place where the credit figure and the message figure are the same number: every turn costs one, so a plan allowance of N credits is N messages for as long as you stay on 1 credit rows. The allowance per plan, the reset and the rollover rule are all in the [credits reference](https://whizi.io/docs/credits).

One more mechanic that matters at the cheap end: a turn that does not fit in the remaining balance is refused whole rather than part-charged. Twenty credits on three remaining credits is rejected, not discounted. On a 1 credit model that almost never bites.

#### Where a credit rung comes from, and when owner policy overrides it

The cost index measures a standard answer of 1,000 input tokens plus 500 output tokens against provider list rates. The credit ladder is derived from a different reference turn, 3,000 input tokens and 800 output tokens, divided by what that same turn costs on the anchor model, which is $0.00156 of provider spend. The result is then rounded up to a legal rung. The ladder holds eighteen of them: 1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 20, 25, 30, 40, 50, 60, 80, 100. Fourteen carry at least one model today, and the empty four are kept deliberately, because a missing rung would round the next expensive model up to the one above it. Rounding goes up when a model falls between two rungs.

On top of that, some rows are pinned by owner policy rather than arithmetic. Three examples visible in the tables above and in the wider catalogue:

- The Kimi family is charged 10 credits against cost-true rates of 2, 3, 4 and 12 depending on the row. That is why two Moonshot rows sit within a hair of each other in dollars, at $0.00172 and $0.00185, yet both charge 10 credits.
- The Gemini row on Starter is charged 1 credit against a cost-true 2, which is what keeps Starter a flat one credit per message.
- The GPT Terra row is charged 4 credits against a cost-true 10.

Search-grounded rows are priced differently again, because the provider charges a flat fee per search that does not scale with the per-token rate. One search at $0.005 is worth 3.2 credits, so a single search costs more than an entire message on most of the catalogue. Those rungs are floors that assume one search per answer.

A model identifier with no rung at all is charged the top of the ladder, 100 credits, deliberately: an unpriced model should surface as a support ticket rather than quietly eat margin. A build test fails if any catalogue model or agent identifier reaches production without a rung.

Finally, credit pricing is enabled per platform. A platform on the allowlist sees real multipliers in the picker and is charged them against the credit allowance. A platform that is not sees a flat one credit per turn and keeps the older message allowance instead, which is 400 messages a month on Starter, 800 on Pro and 5,000 on Powerhouse. The web client is on the credit system today.

Key points:

- The cheap end of the cost index in dollars per standard answer, with the credit charge beside it
- Which plan unlocks each cheap row, including the ones that sit on Powerhouse
- Where a credit rung actually comes from, and which rows are pinned by owner policy instead

Checklist:

- The cheapest priced row in the index answers once for $0.000053
- A standard answer is 1,000 input tokens plus 500 output tokens, priced from OpenRouter list rates on 2026-08-20
- Every charge is floored at one credit, so no message costs less than one
- Most of the very cheapest rows sit on Powerhouse, because anything not on the Starter or Pro list falls through to the top tier
- Starter reaches four identifiers and all of them cost 1 credit, so 400 credits is 400 messages
- Free accounts have no credit allowance, but they are not model-gated within seven lifetime messages
- The credit rung is derived from a different reference turn than the index figure, and some rungs are pinned by policy

FAQ:

Q: What is the cheapest AI model on Whizi?
A: By cost per standard answer, the cheapest priced row in the Whizi cost index is Ling from inclusionAI at $0.000053, followed by Nex from Nex Agi, Solar Pro from Upstage and the Qwen fast tier. All of them charge one credit per message. By what you actually pay, they are tied with every other one credit row, including the Whizi house model, [the base GPT model on Starter](https://whizi.io/docs/using-gpt-in-whizi) and several DeepSeek, Llama and Gemini rows.

Q: Which plan do I need to reach the cheapest models?
A: It depends on the row. Several of the very cheapest rows require Powerhouse, because model access falls through to the top tier for any identifier that is not on the four-item Starter list or the 37-item Pro list. The cheap rows reachable on Pro include the DeepSeek, Llama, Mistral, Qwen and Gemini Flash entries marked Pro in the tables on this page.

Q: Does a cheaper model always cost fewer credits?
A: No. The two figures are measured differently. The index prices a 1,000 in, 500 out standard answer against provider list rates, while the credit rung comes from a 3,000 in, 800 out reference turn rounded up to a legal rung on a fixed ladder, with some rows pinned by policy above or below their arithmetic. That is why a row at $0.0015 per answer can charge one credit while a row at $0.000825 charges two.

Q: What happens to the credit if an answer fails partway through?
A: It comes back. A generation that fails or is aborted is refunded by deleting its usage row, so the balance returns to where it was. The refund is keyed to a server-minted request id rather than anything the client supplies, so applying the same refund twice restores exactly one turn and not two. One thing that is not refunded: a turn rejected for being out of quota still spends its rate-limit tokens, otherwise an out-of-credit client could retry without limit.

Q: Can I try the cheap models without a subscription?
A: Yes, inside the seven message lifetime allowance described above, which applies to any text model. That allowance never resets, and claiming a guest account carries its usage into the registered account, so starting as a guest and signing up afterwards does not hand back the seven.

### How to check your credit balance in Whizi

URL: https://whizi.io/docs/check-your-credit-balance/
Updated: August 2026

Quick answer: Open Settings and read the Usage card at the top of the page. Its first row shows what you have left, written as remaining of limit left, and it is the only place in the Whizi web app that shows a balance. The row is labelled Messages, but on the web the number is credits.

The remaining balance lives in one place: the Usage card at the top of Settings. What each row counts, why the row is labelled Messages, and what free accounts see instead.

#### The short answer

Open Settings. The Usage card is the first thing on the page, and its top row shows what you have left, written as `{remaining} of {limit} left`. That card is the only place in the Whizi web app that shows a balance.

Two ways to get there, both in the bottom left of the chat sidebar: click your account name and plan, or click the gear icon beside it. Both open Settings at the top of the page, where the Usage card sits above Account.

The row is labelled `Messages`, but on the web app the number in it is credits, not messages. That is the single most confusing thing about this screen, and the reason is in the section below.

#### What the Usage card shows

The card holds up to three rows, each with a coloured count and a bar underneath it. Every row counts down. The bar empties as you spend, so a fresh allowance reads full.

| Row label | What the number means | When the row appears |
| --- | --- | --- |
| `Messages` | Credits remaining out of your plan allowance for this period | Always |
| `Image generations` | Image generations remaining for this period | Only when your plan has an image allowance above zero |
| `Voice minutes` | Realtime voice minutes remaining, with part minutes of usage counted as whole minutes | Only when your plan has voice minutes above zero |

Once the real numbers arrive, the rows you have are a readout of your plan.

| Plan | Rows in the Usage card |
| --- | --- |
| Free | `Messages` only |
| Starter | `Messages` and `Voice minutes`. No Images row, because Starter has no image allowance |
| Pro | All three |
| Powerhouse | All three |

The bar and the remaining number share one colour, and that colour is a straight function of how much is left: green at a full allowance, sliding through amber, to red at zero. So the card is readable at a glance without reading the numbers.

Under the rows is a one line footer that tells you which allowance you are looking at.

| Footer line | What it means |
| --- | --- |
| `Resets monthly, shared across web and mobile` | A paid plan on a monthly billing cycle |
| `Resets weekly, shared across web and mobile` | A paid plan on a weekly billing cycle |
| `Free message allowance` | A free account, which has no credit allowance at all |

"Shared across web and mobile" is literal. The same counters back both clients, so a message sent on your phone moves the bar you are looking at in the browser.

#### Why the row says Messages when the number is credits

Whizi has two allowances in its code, a message allowance and a larger credit allowance. Your client decides which one you get, not your plan. The web app declares itself on every request to the backend, and web is on the credit list, so the entitlement the server returns to your browser is the credit allowance and the credit count spent against it.

This is the reason the number is trustworthy even though the label is not: the same predicate picks the allowance that Settings displays and the allowance the server enforces when it decides whether your next message can be sent. They cannot disagree about the size of your allowance or how much of it is gone.

The label itself is the leftover. So read the top row as credits, and use the allowance table below to check which number you should be seeing at the start of a period.

| Plan | Credits per month | Credits per week on a weekly cycle |
| --- | --- | --- |
| Starter | 400 | 100 |
| Pro | 2,000 | 500 |
| Powerhouse | 8,000 | 2,000 |

That table holds six cells but five distinct numbers: 400, 100, 2,000, 500 and 8,000. A row reading 2,000 is Pro on a monthly cycle or Powerhouse on a weekly one, and the footer line tells the two apart. Three limits belong to the message allowance alone and never appear on the credit side: 800, 5,000, and 400 under a weekly footer. Seeing one of those means the client you are on is reading the message allowance. What each credit figure buys in real messages is in [how Whizi credits work](https://whizi.io/docs/credits), and the rows that spend the least of it per answer are in [the cheapest models per answer](https://whizi.io/docs/cheapest-models-per-answer).

#### What a free account sees

A free account has no credit allowance. The credit limit for the free tier is zero, so the Usage card shows something else in the top row: the lifetime free message allowance, which is 7 messages and never resets.

The footer confirms it by reading `Free message allowance` instead of one of the reset lines. The Images and Voice rows do not render, because a free plan has no allowance in either.

Free accounts are not restricted to a subset of models, so the 7 messages can be spent on any text model. What happens when they run out is on [when you run out of credits](https://whizi.io/docs/out-of-credits).

#### Where the balance is not shown

It is worth knowing which surfaces look like they should show a balance and do not, so you stop hunting for a number that is not there.

| Surface | What it actually shows |
| --- | --- |
| The model picker | The price of the next message on that model, as a violet badge reading `2x`, `10x`, `20x` and so on. Never a balance |
| A model with no badge | Not a cheap model. A blank badge slot means this client is not on the credit system, or that model carries no credit rung. Every priced row shows a badge, and a model that costs one credit shows `1x` |
| The Auto row | On Pro and Powerhouse, a range such as `1-10x`, because the cost depends on which model Auto picks for that message. Free and Starter reach one rung only, so their Auto badge collapses to a plain `1x` |
| The Deep Research panel | A flat per-run price of 200 credits before you start, and the credits that run spent once it finishes. Not a balance |
| The composer and the sidebar | Nothing. The credit card icon at the bottom of the sidebar opens pricing, not a balance |
| Settings, anywhere below Usage | Your plan name, as `Current Plan: {plan}` under Account. No invoice, receipt or billing history screen exists. See [invoices and receipts](https://whizi.io/docs/invoices-and-receipts) |

There is also no history view. Only the current period is reachable from the app, and no screen shows what last month cost you.

#### If the card shows no numbers

Two states are not a bug.

1. **Grey placeholder bars where the counts should be.** The rows deliberately render before the numbers arrive so the card keeps its height and the page does not jump under your cursor. The counts fill in a moment later.
2. **A single line reading `Usage is unavailable right now. Please try again later.`** The usage request failed or you are not signed in. The card blanks its rows instead of printing a zero, because a wrong balance is worse than no balance.

The card loads once when you open Settings and does not poll. After sending a message, reopen Settings to see the bar move.

A number that looks wrong instead of missing is a support question, and the route to a human is in [when you run out of credits](https://whizi.io/docs/out-of-credits).

Key points:

- One surface shows the balance: the Usage card, first thing in Settings
- The row reads Messages, but on the web app the number is credits
- What the Images and Voice rows are, and when each one appears at all

Checklist:

- Settings, top of the page, Usage card: the only balance in the web app
- Reach it from the account name or the gear icon at the bottom of the sidebar
- The `Messages` row is credits on web, written as `{remaining} of {limit} left`
- Green to amber to red tracks how much of the allowance is left
- Images and Voice rows appear only when your plan has those allowances
- Free accounts see a 7 message lifetime allowance instead of credits
- The model picker shows price per message, never a balance
- No history screen: only the current period is reachable

FAQ:

Q: Where do I see how many credits I have left in Whizi?
A: In Settings, in the Usage card at the top of the page, on the row labelled `Messages`. It is written as `{remaining} of {limit} left` with a coloured bar underneath, and it is the only place in the web app that shows a balance. Open Settings from the bottom left of the chat sidebar, either by clicking your account name or the gear icon next to it.

Q: Why does my usage row say Messages instead of Credits?
A: Because the label predates the credit system. Whizi keeps both a message allowance and a larger credit allowance in its code, and the client decides which one applies, not the plan. The web app is on the credit list, so the number under that label is credits remaining, taken from the same entitlement the server uses to decide whether your next message can be sent. Starter is 400, Pro is 2,000 and Powerhouse is 8,000 per month.

Q: Does the model picker show my remaining credits?
A: No, it shows the price of the next message. Every priced row carries a badge such as `10x` or `20x` meaning that model costs that many credits per message, and [a model that costs one credit](https://whizi.io/docs/cheapest-models-per-answer) is badged `1x` too. Auto is the exception: on Pro and Powerhouse it shows a range, because what it costs depends on which model it picks, while Free and Starter reach a single rung and see a plain `1x`. For the balance itself you still have to open Settings.

Q: Can I see my credit usage for last month?
A: No. Whizi does keep the underlying usage rows, so the record exists, but nothing in the product ever displays them back to you. Why the previous period stops being reachable once it rolls over is explained in [when you run out of credits](https://whizi.io/docs/out-of-credits).

Q: Why is my Images row missing?
A: Because your plan has no image allowance. Image generations are 100 a month on Pro and 500 on Powerhouse, and zero on Starter and free, and the row renders only when the limit is above zero. Starter therefore shows `Messages` and `Voice minutes` and no Images row, and a free account shows a single row. On a paid plan you may see an Images row flash up and vanish: before the request settles, the card reserves both media rows from the cached plan name, and the Starter one disappears when the real limit of 0 lands. The boundaries are in [plans and limits](https://whizi.io/docs/plans-and-limits).

Q: The Usage card says usage is unavailable. What now?
A: The card is display only, so nothing about your account changes while it is blank: messages keep going through, and the server keeps charging and counting them exactly as it would with the numbers on screen. Reload the page while signed in to retry the request.

### How to compare AI models side by side in Whizi

URL: https://whizi.io/docs/compare-models-side-by-side/
Updated: August 2026

Quick answer: To compare AI models side by side in Whizi, send the same prompt to two models in parallel and read the answers next to each other. Neither model sees the answer from the other, so neither is anchored by the other. Side-by-side comparison is a Powerhouse feature and runs two models at a time, each counting as one message.

Run the same prompt against Claude, GPT, and Gemini in parallel, rank the outputs, and keep only the answer that best fits your work.

#### The short answer

Send the same prompt to two models in parallel and read the answers next to each other. Neither model sees the other's output, so neither answer is anchored by the other, which is the whole reason the comparison is worth anything.

Use it to decide which model should be your default for a kind of work. If you want one answer improved rather than two answers compared, do the other thing instead: switch models inside the same thread and ask the second to critique the first.

#### Why benchmarks do not answer your question

Published benchmarks measure performance on standardised tasks. Your question is narrower and more useful: which model is better at the thing you personally do twenty times a week. Those are different questions, and the second one has no published answer because nobody has your workload.

Running one prompt against two models takes about thirty seconds and answers it directly. The important part is not that you get two answers; it is that you find out how large the gap is. Sometimes the outputs are near identical, which tells you to stop thinking about model choice for that task. Sometimes one is unusable, which is worth knowing before you build a workflow on it.

The other thing comparison catches is confident error. When two models give substantively different answers to a factual question, at least one is wrong, and you would not have known from reading either alone. That signal is not available in a single-model workflow at any price.

#### Decide what better means before you read

The trap in side-by-side comparison is preferring whichever output is longer, more confident, or more polished. Those are not quality. Pick your criterion first, then read.

| Task | What better means | What to ignore |
| --- | --- | --- |
| Writing | Needs less editing to be sendable | Length, vocabulary, enthusiasm |
| Factual research | Sources that resolve and support the claim | Fluency and confidence |
| Extraction | Correct schema, no invented fields, consistent labels | Prose quality around the table |
| Reasoning | The steps hold and the edge case is addressed | Whether the conclusion matches your prior |
| Code | Handles the failure path, is reviewable | Cleverness, brevity |
| Summarizing | Keeps what matters and drops what does not | Comprehensiveness |

A practical trick: before reading either output, write down one sentence describing what a good answer would contain. Then read. It takes ten seconds and it prevents the polish bias, which is strong and mostly unconscious.

For factual questions, check the disagreement rather than the winner. Where two models agree on a specific number and a source, confidence is reasonable. Where they diverge, that is the thing to go verify, and it is the single most valuable output of the whole exercise.

#### Prompts that expose real differences

Some tasks separate models sharply and some do not. If you want to learn something from a comparison, use prompts that put pressure on a specific capability.

- **Tone under difficulty.** `Write a note to a client explaining that we missed the deadline, taking responsibility without over-apologising and without excuses. Under 120 words.` Differences in register show up immediately here.
- **Strict extraction.** `Extract every date, amount, and party from this text into a JSON array with exactly these keys. If a field is absent, use null. Do not infer.` Tests format discipline and the tendency to invent.
- **Reasoning with a trap.** Give a problem with a plausible wrong answer, such as a rate or proportion question where the intuitive route is incorrect. Models differ in whether they take the shortcut.
- **Long-context recall.** Upload a long document and ask about something in the middle. Reveals real usable context, not advertised context.
- **Admitting ignorance.** `What was announced about [something genuinely obscure or very recent]?` The best answer is a clear "I do not know" or a sourced retrieval. Fabrication here is disqualifying.
- **Following a negative constraint.** `Explain X without using any analogy or metaphor.` Compliance with negative instructions varies more than you would expect.

Run comparisons on your own real work rather than on puzzles. A model that is better at your quarterly report is more useful than one that is better at a logic riddle.

#### Turning a week of comparisons into a routing map

A single comparison is interesting. A week of them is actionable. The routine:

1. For five days, whenever a task matters, run it against two models instead of one.
2. Note the task type, the winner, and how large the gap was. Three words is enough.
3. At the end of the week, look at where one model won repeatedly and where the outputs were interchangeable.
4. Set your defaults from that, and stop comparing on the tasks where the gap was consistently nil.

What people typically find is that comparison matters on a minority of their work and is a waste of time on the rest. Quick factual lookups, simple rewrites, and routine formatting rarely separate the models. Writing that will be read by a customer, research that will inform a decision, and reasoning about something unfamiliar separate them a lot.

That result is the actual payoff. You end up comparing less, not more, but on the tasks where it changes the outcome.

#### Side by side versus sequential

Two different techniques, worth distinguishing.

**Side by side** runs the same prompt in parallel against two models that cannot see each other's output. Use it when you want an unbiased comparison, because neither answer is anchored by the other. This is the right tool for choosing a default model and for catching factual disagreement.

**Sequential** means getting an answer, then switching models in the same thread and asking the new one to critique it. Use it when you want the flaw found rather than a comparison made, because the second model can engage directly with the specific argument. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Rough rule: parallel to decide which model, sequential to improve an answer.

Key points:

- How to judge two outputs without just picking the one that sounds more confident
- Six prompts that reliably expose the differences between models
- A one-week routine that produces a routing map specific to your own work

Checklist:

- Write and refine the prompt in a normal chat before comparing
- Decide what better means for this task before you read either output
- Write one sentence describing a good answer, then read, to avoid the polish bias
- On factual questions, treat disagreement as the finding and go verify it
- Compare on your own real work, not on puzzles
- Log the winner and the gap size for a week, then set defaults from the pattern
- Stop comparing on the tasks where the gap is consistently nil

Steps:

1. Write your prompt once: Draft the prompt in a regular Whizi chat first, so you know it is clear.
2. Enable side-by-side: Click the compare icon in the composer to open a parallel view.
3. Pick the models: Choose the two models you actually want to decide between.
4. Send the prompt: Whizi runs the prompt in parallel and streams each output.
5. Judge against a criterion: Decide what "better" means for this task before you read, then rank the outputs.

FAQ:

Q: How many models can I compare at once?
A: Side-by-side comparison is a Powerhouse feature, and it runs two models at a time, which is deliberate: two columns is the layout you can actually read carefully. Three columns tends to become skimming, and skimming defeats the purpose, since the value of the exercise is in noticing where the answers genuinely differ.

Q: Does side-by-side use more of my monthly messages?
A: Yes, each model counts as one message, so a comparison costs two. Most people find it pays for itself on the tasks where it matters, because catching a wrong answer or an unusable draft before it ships is worth more than a message allowance. The efficient approach is to compare on decisions and deliverables, and to use a single model for routine lookups and quick rewrites.

Q: What if both answers are equally good?
A: That is a real and useful result: it tells you this task does not depend on model choice, so stop spending attention on it and use whichever is your default. Most workloads split this way, with a majority of tasks where the models are interchangeable and a minority where the gap is large. Finding out which is which is the point of running comparisons for a week.

Q: How do I avoid just picking the answer that sounds better?
A: Decide your criterion before reading. Longer, more confident, and more polished outputs are systematically preferred even when they are worse, and that bias is largely unconscious. Writing one sentence about what a good answer would contain, before you look, is enough to counteract most of it. For factual work, judge on whether the sources resolve and support the claim rather than on how the answer reads.

Q: Which two models should I compare?
A: Compare the two you are actually deciding between for that specific task, which for most people is Claude against GPT for writing and reasoning, or a large-context model against your default when documents are involved. Comparing a model you would never use against your favourite tells you nothing you will act on.

### How to contact Whizi support

URL: https://whizi.io/docs/contact-whizi-support/
Updated: August 2026

Quick answer: Whizi support is one email address, whizihelp@gmail.com. The support page at whizi.io/support carries the same address, and inside the app the route is Settings, then Support & legal, then Contact support. The inbox is monitored in English, with a typical reply within two business days. There is no phone number and no live chat.

The contact routes published on the website, what to put in the first message so it can be answered without a reply, and what support has and has not promised.

#### The short answer

Whizi support is one email address, whizihelp@gmail.com. The support page at [whizi.io/support](https://whizi.io/support) puts that address next to written guidance, and inside the app the route is Settings, then Support & legal, then Contact support, which opens the same page. The support page states that the inbox is monitored in English and that Whizi typically responds within two business days.

On the website there are four routes, and they are not interchangeable:

| Route | Where it is | What it is for |
| --- | --- | --- |
| Email | whizihelp@gmail.com | Anything. This is the address the Terms and the Privacy Policy both name |
| Support page | whizi.io/support, linked from the site footer and from Settings under Support & legal | Written guidance on sign-in, subscriptions, deletion and AI output, with the same address at the top |
| Send feedback | Settings, under Support & legal, above Contact support | A bug report or a feature request, sent from inside the product with your setup attached |
| Report this chat | The Report control on a public /share link | Content on a conversation someone shared with you, reportable without an account |

For problematic model output the support page asks for one specific thing: email whizihelp@gmail.com with the subject line "AI output report", and include the model used, the approximate date and time, what was problematic, and a screenshot or the relevant excerpt if it is safe to share.

#### What to put in the first message

Nothing about a chat is attached to an email you write, so the reply is only as good as what you paste in. Include these five things.

**The exact error text, character for character.** Paste it exactly. A paraphrase loses the string that identifies it.

**Any number you can see.** Some strings render a live value into a placeholder. `Account deletion failed ({status}).` carries a real HTTP status where `{status}` sits, but it reaches only the browser console and network tab: the Settings screen shows `Your account could not be deleted. Please try again, or contact support.` instead. Quote whichever one you have, with the number intact if there is one.

**The error code, if you can see one.** Most errors from the current backend come back as JSON with a code beside the message, so the code sits in a browser network trace even when the screen shows only the sentence. Some older routes return the sentence alone, and a few messages are generated in the browser and have no code at all.

**The model you had selected, and the approximate time.** Both narrow the search on the server side, and the support page already asks for them on an AI output report.

**Write from the address on the account.** The support page asks for this twice, once for sign-in problems and again for deletion requests, and adds the device type, operating system and a short description of what happened for the sign-in case. The Privacy Policy is blunter about why: Whizi may need to verify your identity before completing a rights request.

#### Which error strings are worth pasting

These are the strings the docs already route to support, because nothing on your side clears them. Each row links the page that owns the full explanation.

| String | What it is | Owned by |
| --- | --- | --- |
| `Something went wrong. Please retry.` | HTTP 500 `internal_error`, the catch-all for any unhandled backend error. It identifies no cause, so retry once and then send it in | [message not sending](https://whizi.io/docs/message-not-sending) |
| `User not found.` | The account record behind your session is gone. There is no self-serve repair | [login and account access](https://whizi.io/docs/login-and-account-access) |
| `Account deletion failed ({status}).` and `Export request failed ({status}).` | The raw errors behind the Settings toasts, with the HTTP status filled in | [login and account access](https://whizi.io/docs/login-and-account-access) |
| `Generation is paused on this account after repeated requests that Whizi cannot fulfil. It reopens automatically. Contact support from Settings if you think this is wrong.` | HTTP 429 `content_policy_locked`, a lockout earned by repeatedly tripping the generation screener | [rate limits](https://whizi.io/docs/rate-limits) |
| `Failed to create Stripe checkout session.` | The checkout call failed and the backend supplied no message of its own | [payment declined](https://whizi.io/docs/payment-declined) |
| `Media could not be stored.` and `Stored media metadata was invalid.` | The storage write returned no object, or the stored row could not be read back | [file upload failed](https://whizi.io/docs/file-upload-failed) |
| `Realtime voice is not configured yet. Please try again later.` | HTTP 503, returned when the voice provider key is not set on the server | [voice mode not working](https://whizi.io/docs/voice-mode-not-working) |

The image content policy refusals carry the instruction inside the string, each ending "If you think this was a mistake, contact support from Settings." They are listed with their codes on [image generation failed](https://whizi.io/docs/image-generation-failed). A share link blocked at creation time says the same thing, in the message `This conversation cannot be shared publicly. Contact support from Settings if you think this is wrong.` returned as HTTP 400 `content_policy_share_blocked`.

#### The in-app feedback form, and how it differs

Settings has a Send feedback row above Contact support, with the subtitle "Tell us what broke, or what you wish Whizi did." It opens a form, not an email client, and it is a different queue from support: product feedback on one side, account and billing problems on the other.

What the form does: three kinds to choose from, labelled "Something broke", "I have an idea" and "Something else"; a message box that accepts up to 4,000 characters and truncates anything past that; and an optional email field for a reply, separate from the account address on purpose.

What rides along without you typing it, quoted from the line under the box: "Sent with your browser, window size, language, and selected model so we can reproduce it. Nothing from your chats is attached."

The route accepts 6 submissions per minute per user, and going past that returns the message "That is a few in a row. Give it a minute, then send this again." If the feedback route is unavailable the form says so and offers the other door: "Feedback is not available right now. The support page still works."

#### What support is the only route to

Some requests have no self-serve equivalent anywhere in the product.

| Request | Why support | Detail |
| --- | --- | --- |
| Changing the email address on an account | There is no setting for it | [switch account email](https://whizi.io/docs/switch-account-email) |
| A credit balance that looks wrong rather than spent | The usage rows behind a balance are kept for 400 days, but no screen in the product shows them | [out of credits](https://whizi.io/docs/out-of-credits) |
| A charge whose amount looks wrong | The Terms give one address for it | [invoices and receipts](https://whizi.io/docs/invoices-and-receipts) |
| Anything resembling a team or multi-seat arrangement | No team plan exists to self-serve | [team accounts](https://whizi.io/docs/team-accounts) |
| A content policy lockout you believe is wrong | It reopens automatically, but no duration is published | [rate limits](https://whizi.io/docs/rate-limits) |
| `User not found.` when you did not delete anything | There is no self-serve repair | [login and account access](https://whizi.io/docs/login-and-account-access) |

#### What support has not promised

**No refund is promised anywhere in the Terms.** The Cancellation clause names whizihelp@gmail.com for anyone unsatisfied, and cancellation takes effect at the end of the current paid term. What that route does and does not promise is set out on [refunds](https://whizi.io/docs/refunds).

**No turnaround is published for a rights request.** The Privacy Policy says you may have rights to access, correct, delete, or receive a copy of personal information, and that Whizi may need to verify your identity first. It attaches no number of days to any of it.

**No phone number and no live chat appear on the support page, in the Terms, or in the Privacy Policy.** Those three between them publish an email address, the support page URL, the account deletion page, and a postal address: Whizi, 238 Featherstone Crescent, Kitchener, Ontario N2R 1Z3, Canada. There is no ticket portal to log into and no chat widget.

**Support cannot undo an account deletion.** Deletion is permanent, and what it removes is listed on [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

#### What never to send

The support page carries this as a red panel: "Never email us your password, full payment-card details, authentication codes, or sensitive prompts and files. Whizi support will never ask for your password."

The same page adds one rule specific to AI output reports: remove personal, confidential, or regulated information before sending the report, and if the output suggests an immediate danger, contact the appropriate local emergency service rather than support.

Key points:

- Every contact route published on the website, and which one fits which problem
- What to include so the first reply can solve it
- The requests support is the only route to, and the ones it has not promised

Checklist:

- The support address is whizihelp@gmail.com, named in the Terms and the Privacy Policy
- The support page is whizi.io/support, reachable from the footer and from Settings
- The inbox is monitored in English, with a stated typical reply within two business days
- Paste the error string verbatim, including any number rendered into a placeholder
- Write from the email address on the account, since identity may need verifying
- Send feedback in Settings is a separate queue, capped at 6 submissions per minute
- Never send a password, full card details, authentication codes, or sensitive files

FAQ:

Q: What is the Whizi support email address?
A: whizihelp@gmail.com. It is the address on the support page, the address the Terms name in the Cancellation clause for anyone unsatisfied, and the address the Privacy Policy names for privacy questions, rights requests and support. There is no second address.

Q: How long does Whizi support take to reply?
A: Two business days is the only turnaround published anywhere. Requests governed by the Privacy Policy, such as access, correction or deletion of personal information, carry no published timeframe at all, and Whizi may need to verify your identity before completing one.

Q: Does Whizi have live chat or a phone number?
A: No. The routes published on the website are the email address, the support page at whizi.io/support, the Send feedback form in Settings, and the Report control on a conversation someone shared with you. Section 28 of the Terms adds a postal address in Kitchener, Ontario.

Q: Should I use Send feedback or email support?
A: Send feedback for "this broke" and "please add this", because it attaches your browser, window size, language and selected model so the report can be reproduced without a reply. Email support for anything tied to your account, your subscription or a charge, because those need the address on the account and often need identity verification. If the feedback form fails it says so and links the support page.

Q: Can support change the email address on my account?
A: Yes, and support is the only published route to it, because no setting in Whizi changes the sign-in address. The Privacy Policy frames it as a correction request, which means identity verification may be required and no turnaround is published. The full answer is on [switch account email](https://whizi.io/docs/switch-account-email).

Q: Can I report a shared conversation without a Whizi account?
A: Yes. A public share link carries a "Report this chat" control that needs no sign-in, because whoever was sent the link is the person who needs to be able to flag it. It asks for one of seven reasons, from offensive or hateful through to something else, plus an optional note of up to 2,000 characters.

### The conversation or document is too long for the model

URL: https://whizi.io/docs/context-too-long/
Updated: August 2026

Quick answer: Whizi shows no context limit error. A conversation that outgrows the model is trimmed to fit before the request is sent, silently. Four caps do refuse a message outright: 100,000 characters in one message, 100 messages in one request, a request body over the route cap, and a 32,000 character system prompt.

Four length caps refuse a Whizi message outright, and each one names itself. Here they are quoted exactly, along with the token budget every model in the catalogue gets for a single turn.

#### The short answer

Whizi does not show a context limit error. No string in the product mentions a context window or a token limit, because a conversation that outgrows the model is trimmed to fit before the request is sent rather than refused. Nothing tells you it happened. If the model seems to have forgotten the middle of a long chat, that is what you are looking at.

Four things do refuse you outright, and each names itself:

| Message | HTTP code | Trigger |
| --- | --- | --- |
| `This message is {count} characters, over the 100,000 character limit. Attach a shorter file, or ask about one section at a time.` | 400 `message_too_long` | One message, including text extracted from your files, is over 100,000 characters |
| `This conversation has {count} messages, over the 100 message limit.` | 400 `too_many_messages` | A single request carried a transcript of more than 100 messages |
| `Request is too large.` | 413 `request_too_large` | The whole JSON request body went over the route byte cap |
| `That system prompt is missing or over the 32,000 character limit.` | 400 `invalid_system_prompt` | A system prompt over 32,000 characters |

The `{count}` in the character limit string is filled in with your real number, formatted with thousands separators. The code beside each string is what shows in a network trace, even when the interface only shows you the sentence.

If you saw a message that names a context length or a token count, it did not come from Whizi. Skip to the last section.

#### The budget every model gets, and what happens past it

Every model in the catalogue gets the same budget for one turn: 40,000 input tokens and 20,000 output tokens. The Canadian CPA agent is the single exception, at 55,000 input tokens and 40,000 output tokens.

When a conversation runs past that budget, the backend trims history to the model token budget silently rather than refusing. There is no toast, no banner and no error code for it.

The token count Whizi trims against is an estimate rather than a real tokenizer. It can undercount a real tokenizer by up to 1.66 times on JSON, so an allowance of 1.8 times is applied to input to cover the worst measured case.

#### Switching to a bigger context model does not send more

This is the most common wrong fix. The 40,000 token input budget is flat: it is the same on a model with a one million token window as on a model with a 200,000 token window. Moving a long conversation from one large-window model to another changes nothing about how much of it is sent.

[Context window size](https://whizi.io/docs/context-windows-compared) only changes the budget in one direction, downward. Any model whose window sits below 93,000 tokens gets a smaller, proportional budget instead of the flat one, with output capped at 40 percent of the window and a 1,000 token safety margin held back. 31 models in the catalogue are clamped this way, and the smallest window among them is 6,144 tokens.

So the switch that does help is the opposite of the one people try: moving off a small, clamped model onto an ordinary one. Moving between two large-window models is a change with no effect on length. How switching mid-conversation behaves is covered in [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

One detail behind the 93,000 figure, worth knowing if you are reasoning about why a model refused something small: OpenRouter counts input plus the requested maximum output against a model window, not input alone.

#### A long upload gets cut, and it tells you

Uploads are the usual reason a single message runs past 100,000 characters, because PDF, Word and spreadsheet files are extracted to text in your browser and that text counts against the same cap as anything you typed.

When the extracted text does not fit, it is trimmed rather than refused, and two strings appear. The toast reads `Your upload was too large, so only the first part of it was sent. Ask about a smaller section for full coverage.` The message itself carries the marker `[Attachment truncated: the upload was larger than one message can carry, so the content past this point was not included.]` at the cut, so the model can see where its copy stops.

If the message is refused instead of trimmed, you get the `message_too_long` string from the first table. The fix is the one the error itself names: attach a shorter file, or ask about one section at a time.

One more behavior that reads as forgetting: only the most recent user message carrying attachments has its attachments forwarded to the model. An image or PDF you attached ten turns ago is not re-sent on every subsequent turn. If you need the model to look at it again, attach it again. What Whizi accepts and how extraction works is in [supported file types](https://whizi.io/docs/supported-file-types).

One cost note, since length changes what a turn is worth. On Auto, a message over about 6,000 characters routes to the long-form rung at 4 credits, on the reasoning that a message that long is a pasted document rather than a question. A short question routes to the quick rung at 1 credit. The credit scale is in [how credits work](https://whizi.io/docs/credits).

#### Pinned project files are charged to the prompt on every turn

A pinned project file rides along as prompt text on every single turn in that project rather than once, which is why images are deliberately excluded from the pinned file types.

The caps are their own: each pinned file contributes at most 32,000 characters of extracted text, and the whole project block is capped at 120,000 characters, across at most 10 pinned files. Project custom instructions can be up to 32,000 characters.

Pinned files and instructions are rebuilt into the prompt on every turn in that project, inside the same input budget as the conversation. If a project chat seems to lose the thread faster than an ordinary one, unpin the files you are not asking about.

#### If you did see a context length error

Then it came from the model provider, not from Whizi, and it reached you through the passthrough. Any upstream failure that is not a rate limit is returned as HTTP 502 with the code `provider_error`, carrying the provider message, truncated to 300 characters. When the provider error body cannot be parsed at all, you get `The model provider rejected the request.` instead.

A provider refusal on length will name a context length and a token count. Those numbers are the provider counting your request against a window smaller than the budget Whizi sent, and they are exactly what support needs to look at it.

No interface setting changes this. Switch to another model to get your answer, and send support the exact message including the numbers.

Two more strings people mistake for a length problem. `Generation failed mid-stream.` and `The model stream was interrupted.` are connection failures partway through a reply, not length refusals. Retry those. `Too many requests. Please wait and try again.` is a rate limit at 10 messages per minute, which also has nothing to do with length.

Key points:

- A long upload is trimmed with a toast and an in-message marker, both quoted here
- The four length caps that do refuse a message, quoted exactly
- Why a bigger context window does not send more, and the one switch that does help

Checklist:

- Whizi has no context limit error: it trims the conversation silently
- Every model gets a flat 40,000 input and 20,000 output token budget per turn
- The Canadian CPA agent is the one exception, at 55,000 in and 40,000 out
- A window below 93,000 tokens lowers that budget, it never raises it
- One message is capped at 100,000 characters, extracted file text included
- One request carries at most 100 messages of transcript
- Only the most recent attachment carrying message forwards its attachments
- A pinned project file rides along as prompt text on every turn in that project
- On Auto, a message over about 6,000 characters routes to the long-form rung at 4 credits
- A message naming a context length came from the provider: switch models and tell support

FAQ:

Q: Does Whizi have a context limit error?
A: Not a context one. The length messages Whizi does show are counts, not windows: 100,000 characters in one message, 100 messages in one request, 32,000 characters in a system prompt, and a 413 `Request is too large.` on the whole request body. If the message in front of you names a token count or a context length, it reached you through the 502 provider passthrough and belongs to the model provider.

Q: Why did the model forget something I said earlier in the chat?
A: Because it was trimmed out of the request before the request was sent. The backend trims history to the model token budget silently rather than refusing, so there is no error to read and no setting that turns it off. Starting a new conversation when the subject changes is the practical answer.

Q: Will switching to a model with a bigger context window let me send more?
A: No. The input budget is a flat 40,000 tokens on every model in the catalogue, so a one million token window and a 200,000 token window get the same amount of your conversation.

Q: What does "over the 100,000 character limit" mean?
A: One message went past the per message cap of 100,000 characters and was refused with HTTP 400 and the code message_too_long. Text extracted from an attached PDF, Word file or spreadsheet counts toward the same cap, so an upload is the usual cause rather than typing. The fix is in the message itself: attach a shorter file, or ask about one section at a time in the same conversation.

Q: What does "over the 100 message limit" mean?
A: A single request carried a transcript of more than 100 messages, and it was refused with HTTP 400 and the code too_many_messages. It is a cap on what one request may carry, not a cap on how long a chat may be. Starting a new conversation for the next subject is the practical answer, and it also gives the model a cleaner view of what you are asking about.

Q: How do I summarize a document that is longer than the limit?
A: Split it and work section by section in one conversation, which is what the message_too_long string itself recommends. Ask for a summary of each section, then a summary of those summaries. If a single section still runs past 100,000 characters once its file text is extracted, split that section again.

Q: Can I pay for a bigger context budget?
A: No. The budget is a property of the model in the catalogue rather than of your plan, and there is no setting anywhere that raises it. What a higher plan buys is access to more models and a larger monthly allowance, not more room in a single turn. The plan mapping is in [the model reference](https://whizi.io/docs/models).

### Context windows compared: what the priced models in Whizi can hold

URL: https://whizi.io/docs/context-windows-compared/
Updated: August 2026

Quick answer: Context windows in the Whizi priced index run from 8,192 tokens to 1,310,720 tokens, and 47 of those 100 rows carry 1,000,000 tokens or more. The window is not what decides your result. Every model gets a flat 40,000 input tokens and 20,000 output tokens per turn, and only a window below 93,000 tokens lowers that.

Context window, cost per answer, credit cost and plan gate side by side for the rows in the Whizi Cost Index, plus the flat per turn token budget that decides how much of a real document actually reaches the model.

#### The short answer

Windows in the priced index run from 8,192 tokens to 1,310,720 tokens, and 47 of those 100 priced rows carry a window of 1,000,000 tokens or more. The window is the maximum a model can consider in one request, counted in tokens, but it is not the number that decides your result in Whizi.

The number that decides it is the per turn budget. Every model in the catalogue gets the same flat allowance for one turn: 40,000 input tokens and 20,000 output tokens. A larger window does not raise that budget. A window smaller than 93,000 tokens lowers it.

| Figure | Value |
| --- | --- |
| Largest window in the priced index | DeepSeek V4 Flash 0731, 1,310,720 tokens |
| Smallest windows in the priced index | The two Tencent Hy-MT2 rows, 8,192 tokens |
| Rows at 1,000,000 tokens or more | 47 of the 100 priced rows |
| Input sent per turn, every catalogue model | 40,000 tokens |
| Output allowed per turn, every catalogue model | 20,000 tokens |
| Window below which that budget shrinks | 93,000 tokens |

So for a real document, the two caps that actually bite are the 100,000 character limit on a single chat message and that 40,000 token input budget. Neither of them is the window column.

#### Context windows by model, with cost and plan

Every row below comes from the Whizi Cost Index, which prices 100 OpenRouter rows across 29 providers on a standard answer of 1,000 input tokens plus 500 output tokens. The provider rates behind these figures were fetched on 2026-08-20. On the web, the credits column is what one message on that model costs against your credit allowance, and the plan column is the lowest plan that can open it. A platform that is not on the credit allowlist sees a flat 1x instead, is charged one per turn, and keeps the message allowance.

**OpenAI.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| GPT-5.6 Luna | 1M | $0.0008 | 1 | Starter |
| GPT-5.4 Nano | 400K | $0.000825 | 2 | Powerhouse |
| GPT-5.4 Mini | 400K | $0.003 | 4 | Powerhouse |
| GPT-5.6 Terra | 1M | $0.008 | 4 | Pro |
| GPT-5.4 | 1M | $0.01 | 15 | Powerhouse |
| GPT-5.6 Sol | 1M | $0.01 | 20 | Powerhouse |
| GPT-5.5 | 1M | $0.02 | 20 | Pro |
| GPT Chat Latest | 400K | $0.02 | 25 | Powerhouse |

**Anthropic.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| Claude Haiku 4.5 | 200K | $0.0035 | 4 | Pro |
| Claude Sonnet 5 | 1M | $0.007 | 10 | Pro |
| Claude Sonnet 4.6 | 1M | $0.0105 | 10 | Pro |
| Claude Sonnet 4.5 | 1M | $0.0105 | 10 | Pro |
| Claude Opus 5 | 1M | $0.0175 | 20 | Powerhouse |
| Claude Opus 4.8 | 1M | $0.0175 | 20 | Powerhouse |
| Claude Fable 5 | 1M | $0.035 | 50 | Powerhouse |

**Google.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| Gemini 2.5 Flash Lite | 1M | $0.0003 | 1 | Pro |
| Gemini 3.1 Flash Lite | 1M | $0.001 | 1 | Powerhouse |
| Gemini 3.7 Flash | 1M | $0.001313 | 2 | Pro |
| Gemini 2.5 Flash | 1M | $0.00155 | 2 | Pro |
| Gemini 3.5 Flash Lite | 1M | $0.00155 | 2 | Powerhouse |
| Gemini 3.6 Flash | 1M | $0.002625 | 3 | Pro |
| Gemini 3.5 Flash | 1M | $0.006 | 8 | Pro |
| Gemini 3.1 Pro Preview | 1M | $0.008 | 10 | Pro |

**DeepSeek.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 | 1.31M | $0.00028 | 1 | Powerhouse |
| DeepSeek V3.2 | 164K | $0.000469 | 1 | Pro |
| DeepSeek V3.1 | 164K | $0.000725 | 1 | Pro |
| DeepSeek V4 Pro 0813 | 1M | $0.00297 | 2 | Powerhouse |

**Qwen.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| Qwen3.7 Flash | 1M | $0.000095 | 1 | Powerhouse |
| Qwen3 Coder Next | 262K | $0.00052 | 1 | Powerhouse |
| Qwen3.6 Flash | 1M | $0.00075 | 1 | Powerhouse |
| Qwen3.7 Plus | 1M | $0.00096 | 1 | Pro |
| Qwen3.8 27B | 1M | $0.00205 | 3 | Powerhouse |
| Qwen3.7 Max | 1M | $0.003688 | 5 | Powerhouse |
| Qwen3.8 Max | 1M | $0.005 | 6 | Powerhouse |

**xAI.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| Grok Build 0.1 | 256K | $0.002 | 3 | Powerhouse |
| Grok 4.3 | 1M | $0.0025 | 4 | Powerhouse |
| Grok 4.5 | 500K | $0.005 | 6 | Powerhouse |
| Grok 4.6 | 500K | $0.005 | 6 | Pro |

**Meta.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| Llama 3.3 70B Instruct | 131K | $0.00026 | 1 | Pro |
| Llama 4 Maverick | 1M | $0.0006 | 1 | Pro |
| Muse Glimmer 30B | 131K | $0.0011 | 2 | Powerhouse |
| Muse Spark 1.2 | 1M | $0.003375 | 5 | Powerhouse |

**Mistral.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| Codestral 2508 | 256K | $0.00075 | 1 | Powerhouse |
| Mistral Large 3 2512 | 262K | $0.00125 | 2 | Pro |
| Mistral Medium 3.1 | 131K | $0.0014 | 2 | Pro |
| Mistral Medium 3.5 | 262K | $0.00525 | 6 | Powerhouse |

**Z.ai and Moonshot.**

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| GLM 4.6 | 205K | $0.0015 | 1 | Powerhouse |
| GLM 4.7 | 205K | $0.001275 | 2 | Powerhouse |
| GLM 5 | 205K | $0.00156 | 2 | Pro |
| GLM 5.2 | 1M | $0.002484 | 3 | Powerhouse |
| GLM 5.3 | 1M | $0.0036 | 5 | Powerhouse |
| Kimi K2 0711 | 131K | $0.00172 | 10 | Powerhouse |
| Kimi K2 Thinking | 262K | $0.00185 | 10 | Powerhouse |
| Kimi K2.7 Code | 262K | $0.00246 | 10 | Powerhouse |
| Kimi K3 | 1M | $0.0105 | 10 | Pro |

**A further fourteen rows.**

These are a selection of the priced rows from the remaining labs, not all of them.

| Model | Context window | Cost per answer | Credits | Lowest plan |
| --- | --- | --- | --- | --- |
| MiniMax M2 | 205K | $0.000765 | 1 | Powerhouse |
| MiniMax M3 | 1M | $0.0009 | 1 | Powerhouse |
| Nemotron 3.5 Lightning, NVIDIA | 262K | $0.00018 | 1 | Powerhouse |
| Nemotron 3 Ultra, NVIDIA | 512K | $0.0024 | 3 | Powerhouse |
| Nova Pro 1.0, Amazon | 300K | $0.0024 | 3 | Powerhouse |
| Command A, Cohere | 256K | $0.0075 | 10 | Powerhouse |
| Solar Pro 4, Upstage | 524K | $0.00009 | 1 | Powerhouse |
| Ling-3.0-flash, inclusionAI | 262K | $0.000053 | 1 | Powerhouse |
| Nex-N2-Mini, Nex Agi | 262K | $0.000075 | 1 | Powerhouse |
| Laguna XS 2.1, Poolside | 262K | $0.00012 | 1 | Powerhouse |
| Granite 4.1 8B, IBM | 131K | $0.0001 | 1 | Powerhouse |
| Phi 4, Microsoft | 16K | $0.00014 | 1 | Powerhouse |
| Inkling, Thinkingmachines | 1M | $0.002975 | 4 | Powerhouse |
| Fugu Ultra, Sakana | 1M | $0.02 | 25 | Powerhouse |

The catalogue holds 280+ models, and the rows above are drawn from the Cost Index, which prices 100 OpenRouter rows for comparison, and neither the tables nor the index are the whole catalogue. The plan column is derived rather than listed, and the [model list](https://whizi.io/docs/models) covers how that derivation works.

Starter is the one plan the table barely covers. Its four entries are Auto, the house model Whizi AI, [the base GPT model](https://whizi.io/docs/using-gpt-in-whizi) that Whizi AI runs on, and [one fast Gemini model](https://whizi.io/docs/using-gemini-in-whizi). Whizi AI is a house persona on that base GPT model rather than a model of its own, and it costs 1 credit a message. Free accounts are not model gated at all: a free account can try any text model inside a lifetime allowance of 7 messages.

#### The windows small enough to change the budget

The window column is not the deciding column for length, because the per turn budget is flat across the catalogue. How that budget is built, and what happens when a conversation runs past it, is covered in [when the conversation is too long](https://whizi.io/docs/context-too-long).

What a window does decide is whether that flat budget applies at all, and the line sits at 93,000 tokens. Four index rows named on this page fall under it, and two of them are the smallest windows in the whole priced index.

| Row | Context window | Printed in the tables above |
| --- | --- | --- |
| Tencent Hy-MT2, two rows | 8,192 | No |
| Phi 4, Microsoft | 16K | Yes |
| DeepSeek R1 | 64K | No |

#### What a window means for a real document

The first cap a document hits is not the context window. It is the 100,000 character limit on a single message, and text extracted from an attached file counts against it exactly like typed text.

On the web, PDF, Word and spreadsheet files are extracted to text in your browser before anything is sent, so what the model receives is that extracted text. Each file is capped at 10 MB. The full format list is in [supported file types](https://whizi.io/docs/supported-file-types).

Projects are the one place a document really does eat the budget, because a pinned project file is rebuilt into the prompt on every single turn in that project, inside the same flat input budget no matter which window the model carries. The pinned file and project block caps are in [when the conversation is too long](https://whizi.io/docs/context-too-long).

#### Window against price and credits, in the rows above

Sort the tables by window and nothing else sorts with them. The largest window in the priced index, at 1,310,720 tokens, costs $0.00028 per standard answer, charges 1 credit a message and sits on Powerhouse. [The cheapest row of all](https://whizi.io/docs/cheapest-models-per-answer) costs $0.000053 per answer on a 262K window. The median row sits at $0.00185 per answer, also on a 262K window. The dearest row costs $0.105 per standard answer and carries no credit rung at all, because it is not one of the rows Whizi offers.

The scatter is easiest to read on rows that share a window.

| Row | Context window | Cost per answer | Credits |
| --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 | 1.31M | $0.00028 | 1 |
| Qwen3.7 Flash | 1M | $0.000095 | 1 |
| Inkling, Thinkingmachines | 1M | $0.002975 | 4 |
| Claude Fable 5 | 1M | $0.035 | 50 |
| Granite 4.1 8B, IBM | 131K | $0.0001 | 1 |
| Kimi K2 0711 | 131K | $0.00172 | 10 |

Three rows there carry the same 1,000,000 token window and charge 1, 4 and 50 credits. Two rows carry the same 131K window and charge 1 and 10. The priced index spans roughly 2000x from cheapest row to dearest, and the rows above show that spread does not follow the window column.

Key points:

- Context window, cost per answer, credit cost and plan gate for the rows Whizi prices
- The flat per turn token budget that decides what is actually sent
- Which windows are small enough to change your result, and which are not

Checklist:

- Windows in the priced index run from 8,192 tokens to 1,310,720 tokens
- 47 of the 100 priced rows carry a window of 1,000,000 tokens or more
- The largest window in the index answers once for $0.00028 and charges 1 credit
- The cheapest row in the index carries a 262K window, not the largest one
- The median priced row sits at $0.00185 per standard answer, also on a 262K window
- On the web, the credits column is what one message spends against the credit allowance
- A credit rung is derived from provider per token rates, so window size is not an input to it
- Whizi AI is a house persona on a base GPT model, at 1 credit a message
- Free accounts are not model gated: any text model, inside 7 lifetime messages
- The Cost Index prices 100 OpenRouter rows, which is not the whole 280+ catalogue

FAQ:

Q: Do I get a bigger context window on a higher plan?
A: No. The window belongs to the model, not to your subscription. What a higher plan buys is access to more models and a larger monthly credit allowance: 400 credits on Starter, 2,000 on Pro and 8,000 on Powerhouse.

Q: Does a model with a bigger context window cost more credits?
A: Not in the tables above. The largest window in the priced index, at 1,310,720 tokens, charges 1 credit a message, while a 131K window row in the same tables charges 10. A credit rung is derived from a reference turn of 3,000 input tokens and 800 output tokens against live provider per token rates, then rounded up to a legal rung, and window size is not one of the inputs.

Q: Can I try a large context window model on a free account?
A: Yes. The free tier is deliberately not model gated, so a free account can open any text model in the catalogue, including the 1,000,000 token rows. The limit is the allowance rather than the model list: 7 messages, lifetime, which does not reset.

Q: Does Whizi publish a context window for every model in the catalogue?
A: No. The windows on this page come from the Cost Index, which prices 100 OpenRouter rows for comparison, and the catalogue holds 280+ models. The picker in the app is a separate surface served live by the shared worker, so it lists the current catalogue rather than this index.

Q: Where do the context window figures on this page come from?
A: From the Whizi Cost Index, which prices 100 OpenRouter rows on a standard answer of 1,000 input tokens plus 500 output tokens, with source prices fetched from OpenRouter on 2026-08-20. The Cost Index records that single fetch date, so treat the dollar figures as of that day rather than as live rates.

### How Whizi credits work

URL: https://whizi.io/docs/credits/
Updated: August 2026

Quick answer: Whizi gives you one credit allowance per month and spends it per message, and how much a message costs depends only on which model answered it. Starter gets 400 credits, Pro gets 2,000, and Powerhouse gets 8,000. Roughly half the catalogue costs 1 credit per message, and the frontier tier costs 20.

One monthly allowance, spent per message at a rate set by the model you chose. The full credit scale, what each plan gets, and how to make an allowance last.

#### The short answer

Whizi gives you one credit allowance per month and spends it per message. How much a message costs depends only on which model answered it. Starter gets 400 credits, Pro gets 2,000, and Powerhouse gets 8,000.

Roughly half the catalogue costs 1 credit per message. The expensive end is the frontier tier: Claude Opus 5 and GPT-5.6 Sol are 20 credits each. Nothing else about a message changes the price, so a long conversation with a 1 credit model still costs 1 credit per turn.

#### The full credit scale

There are fourteen rungs across roughly 297 priced models. The distribution is heavily weighted to the cheap end on purpose.

| Credits per message | Models on this rung | Examples |
| --- | --- | --- |
| 1 | 143 | Whizi AI, GPT-5.6 Luna, Gemini 3 Flash, DeepSeek V3.2, Qwen3.7 Plus, the study and writing agents |
| 2 | 51 | GLM-5, Amazon Nova Lite, Arcee Virtuoso Large |
| 3 | 20 | Canadian CPA agent, Amazon Nova Pro, Seed 2.0 Code |
| 4 | 12 | GPT-5.6 Terra, Claude Haiku 4.5, o3-mini |
| 5 | 8 | o3-mini-high, o4-mini-high, Muse Spark |
| 6 | 8 | Grok 4.6, Mistral Large 2407, Mixtral 8x22B |
| 8 | 12 | Gemini 2.5 Pro, Magnum v4 72B |
| 10 | 20 | Claude Sonnet 5, Claude Sonnet 4.6, GPT-5.2, Gemini 3.1 Pro, Kimi K3 |
| 12 | 4 | GPT-5.3 Codex, Sonar, Sonar Reasoning Pro, Nova Premier |
| 15 | 1 | GPT-5.4 |
| 20 | 12 | Claude Opus 5, Claude Opus 4.5 through 4.8, GPT-5.6 Sol, Sonar Pro |
| 25 | 3 | GPT Chat Latest, Sonar Pro Search, Fugu Ultra |
| 50 | 1 | Claude Fable 5 |
| 80 | 2 | Claude Opus 4 and Opus 4.1, the original generation |

The two 80 credit models are the first-generation Opus releases, which remain the most expensive things in the catalogue to run. The current Opus 5 is 20, a quarter of the price of its own ancestor.

#### What each allowance buys

The useful way to read the allowance is in messages of the model you actually intend to use.

| Plan | Credits | At 1 credit | At 10 credits (Claude Sonnet 5) | At 20 credits (Claude Opus 5) |
| --- | --- | --- | --- | --- |
| Starter | 400 | 400 messages | not included | not included |
| Pro | 2,000 | 2,000 messages | 200 messages | not included |
| Powerhouse | 8,000 | 8,000 messages | 800 messages | 400 messages |

Nobody spends an allowance on a single model, so the real arithmetic is a mix. A Pro subscriber who sends 40 Claude Sonnet 5 messages (400 credits) still has 1,600 credits left, which is 1,600 messages on the fast tier. In practice the frontier models are the constraint and the 1 credit tier is effectively unmetered.

The allowance resets at the start of each billing month. Credits do not roll over. Where the remaining figure is shown, and what it does not include, is in [check your credit balance](https://whizi.io/docs/check-your-credit-balance).

#### What is not charged in credits

Media generation is counted separately, in its own units, and never touches the credit balance.

| Feature | Starter | Pro | Powerhouse |
| --- | --- | --- | --- |
| Image generation | not included | 100 per month | 500 per month |
| AI voice | 10 minutes per month | 80 minutes per month | 500 minutes per month |
| AI video generation | not included | not included | 100 per month |
| AI music generation | not included | not included | included |

This is why the credit table above lists no image or video models. They skip the per model credit gate entirely and are tiered by section instead.

File uploads, web search, side-by-side comparison and sharing a conversation are also not charged in credits. Side-by-side does cost two messages rather than one, because two models each answer, and each answer is billed at its own rate.

#### Making an allowance last

The credit system rewards routing rather than rationing. A few habits that matter more than the rest:

1. **Draft on the cheap tier, finish on the expensive one.** Getting the prompt right takes several turns, and none of them need a frontier model. Draft on a 1 credit model, then switch the same conversation to Claude Sonnet or Opus for the answer you will actually use. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).
2. **Do not use side-by-side for routine work.** It costs two messages, and most tasks show no meaningful gap between models. Save it for decisions and deliverables.
3. **Watch the 20 credit rung.** Claude Opus 5, GPT-5.6 Sol and Sonar Pro are each worth twenty fast-tier messages. They are worth it when the task is genuinely hard, and pure waste on a lookup.
4. **Check the badge in the model picker.** Every row shows its credit cost before you send, so the price is never a surprise after the fact.

The pattern most subscribers land on is a 1 credit default for everything routine and one expensive model reserved for the handful of messages per week that decide something.

Key points:

- The complete credit scale, all fourteen rungs, with how many models sit on each
- What 400, 2,000 and 8,000 credits buy in actual messages
- Why image, video and music generation are not charged in credits at all

Checklist:

- One allowance per month, spent per message, reset at the billing date
- Cost is set by the model alone, not by message length or conversation length
- 143 of the priced models cost 1 credit
- Claude Opus 5 and GPT-5.6 Sol are 20 credits each
- Image, voice, video and music have their own monthly counts
- Side-by-side costs two messages, one per model
- Credits do not roll over to the next month

FAQ:

Q: Does a longer message cost more credits?
A: No. The credit cost is fixed per message by the model you selected, so a one line question and a ten page document analysis cost the same on the same model. This is deliberately simpler than token billing: you can predict what a conversation will cost before you start it, which you cannot do when the price depends on how long the answer turns out to be.

Q: How many Claude messages do I get on Whizi?
A: On Pro, 2,000 credits buys 200 Claude Sonnet 5 messages if you spend the whole allowance there, or 500 Claude Haiku 4.5 messages at 4 credits each. On Powerhouse, 8,000 credits buys 400 Claude Opus 5 messages or 800 Sonnet 5 messages. Most people mix, using a 1 credit model for routine turns so the Claude allowance goes further.

Q: Do unused credits roll over?
A: No, the allowance resets at the start of each billing month and unused credits are not carried forward. If you are consistently finishing the month with a large unspent balance, the cheaper plan is likely the better fit, and if you are running out in the third week the next tier up costs less than the gap in models suggests.

Q: Why do some models cost twenty times more than others?
A: Because they cost roughly that much more to run. The multiplier tracks provider pricing rather than a product decision, which is why the spread is so wide: the gap between the cheapest and the most expensive model in the wider market is close to three orders of magnitude per answer. Charging a flat rate would mean either overcharging everyone who uses fast models or losing money on everyone who uses frontier ones.

Q: Can I see how many credits I have left?
A: Yes, the remaining balance is shown in the app, and each row in the model picker displays its credit cost before you send. When the balance runs out, sending pauses until the reset date and the app offers the next plan up rather than silently failing.

### Does Whizi train on my data? Privacy, retention, and deletion

URL: https://whizi.io/docs/data-and-privacy/
Updated: August 2026

Quick answer: Whizi does not use your prompts, files, conversations, voice transcripts or generated content to train Whizi-owned AI models, and does not sell that content as training data. Uploaded attachments and generated media are configured to expire after up to 30 days. Your messages reach the provider of whichever model you selected.

Whether your conversations train models, how long uploads are kept, what deletion actually removes, and what leaves Whizi when a model answers.

#### The short answer

Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models. Whizi does not sell that content as training data.

Uploaded attachments and generated media are configured to expire after up to 30 days. You can delete them sooner, and account deletion removes them along with the rest of your data.

#### What happens when you send a message

Whizi is a workspace in front of models built by other companies, so a message you send to [Claude](https://whizi.io/docs/using-claude-in-whizi) reaches Anthropic, and a message you send to [GPT](https://whizi.io/docs/using-gpt-in-whizi) reaches OpenAI. That is what selecting a model means, and no aggregator can change it.

The practical implication is that the provider whose model you selected handles that request under their own terms. If a particular provider is unacceptable for a class of work, the answer is not to trust a wrapper, it is to not route that work to that model. A multi-model workspace makes that easy in a way a single-vendor subscription does not, because you can move sensitive work to a different provider without changing tools.

Your messages go to the model you selected. Switching models mid-conversation sends the existing thread to the newly selected model, which is what makes continuity work and is worth knowing before you switch on a sensitive thread. The one background exception is [memory](https://whizi.io/docs/whizi-memory): the passes that extract and condense remembered facts run on a fixed house model rather than the one you picked for the chat.

#### Retention

| Data | How long |
| --- | --- |
| Uploaded attachments | Configured to expire after up to 30 days |
| Generated media | Configured to expire after up to 30 days |
| Conversations and messages | Kept until you delete them or delete the account |
| Saved memory | Kept until deleted |
| Entitlement and usage state | Kept while the account exists |

You can delete particular content without deleting the account, so removing one conversation or one uploaded file does not require the nuclear option.

The local chat cache on a device is separate from the server-side record. On mobile it can also be removed by clearing app storage or uninstalling the app, which is worth knowing if you are handing a device on.

#### Account deletion

Deletion is available at [whizi.io/account-deletion](https://whizi.io/account-deletion) and can be started from the app as well. It removes your conversations and messages, uploaded and generated media, saved memory, entitlement and usage state, guest identity links, push tokens, the Clerk authentication account, and the local data on the device completing the deletion.

A limited retention window applies to the records that have to survive briefly for legal and operational reasons. That is described in the privacy policy rather than summarised here, because the specifics are what matter.

One thing deletion does not do: stop an App Store or Google Play subscription. Those bill until cancelled with the store that sold them, independently of whether the Whizi account still exists. See [billing and cancellation](https://whizi.io/docs/billing-and-cancellation), and [delete your account](https://whizi.io/docs/delete-your-account) for the steps.

#### What to think about before uploading work documents

Two habits worth having, neither specific to Whizi.

First, remember that extracted document text becomes part of the conversation. If you later share that conversation as a link, the extracted text goes with it. Check a thread before sharing it, particularly a long one that changed subject partway through. See [sharing a conversation](https://whizi.io/docs/share-a-conversation).

Second, know your own organisation policy on which providers are approved. The useful property of a multi-model workspace here is that you are choosing per message rather than per subscription, so an approved-provider rule is something you can actually follow rather than something you have to work around.

Key points:

- The training question answered directly, with no hedging
- How long uploads and generated media are kept
- Exactly what account deletion removes

Checklist:

- Whizi does not train Whizi-owned models on your content
- Whizi does not sell your content as training data
- Uploads and generated media expire after up to 30 days
- Messages reach the provider of whichever model you selected
- Switching models sends the existing thread to the new model
- Individual conversations and files can be deleted without deleting the account
- Account deletion does not cancel an App Store or Play subscription

FAQ:

Q: Does Whizi train AI models on my conversations?
A: No. Whizi does not use your prompts, files, conversations, voice transcripts or generated content to train Whizi-owned AI models, and does not sell that content as training data. Uploaded attachments and generated media are additionally configured to expire after up to 30 days rather than being retained indefinitely.

Q: Where does my message actually go?
A: To the provider of whichever model you selected, because Whizi is a workspace in front of models built by OpenAI, Anthropic, Google and others rather than a model of its own. That request is handled under that provider terms. The advantage of the multi-model shape is that you can route sensitive work to a different provider without switching tools.

Q: How do I delete my chat history?
A: Individual conversations and uploaded files can be deleted directly, without touching the rest of the account. For everything at once, account deletion at whizi.io/account-deletion removes conversations, messages, uploaded and generated media, saved memory, usage state and the authentication account, subject to a limited retention window described in the privacy policy.

Q: Is my data deleted if I cancel?
A: No. Cancelling stops billing and ends access at the end of the paid term, but the account and its contents remain, which is what allows you to resubscribe later and find your history intact. If you want the data removed as well, use account deletion, which is a separate action.

Q: Can my colleagues see my conversations?
A: Not unless you deliberately share one. Conversations are private to your account, and sharing creates an explicit read-only public link for a single conversation that you can revoke at any time. There is no implicit team visibility.

### How to delete your Whizi account, and exactly what is removed

URL: https://whizi.io/docs/delete-your-account/
Updated: August 2026

Quick answer: Delete your Whizi account under Settings, in the mobile app or on the website, using the two-step confirmation flow. If you cannot sign in, follow the instructions at whizi.io/account-deletion or contact support. Export your data first, because deletion is not reversible, and cancel an Apple or Google subscription separately.

The two-step deletion flow, the complete list of what it removes, the records that survive it, how to export your data first, and the guest-data version for people with no account.

#### The short answer

Delete your Whizi account under Settings, in the mobile app or on the website, using the two-step confirmation flow. If you cannot get into the account, follow the instructions at [whizi.io/account-deletion](https://whizi.io/account-deletion) or contact support instead.

Two things are worth doing before you confirm, because neither can be done afterwards:

| Do this first | Why |
| --- | --- |
| Export your data | Settings, under Account, has Export Data. It returns your account as one downloadable JSON file. |
| Cancel a store subscription | Deleting the account does not cancel Apple App Store or Google Play billing. That has to be cancelled with the store. |

The terms of service put the second point plainly: "Deleting the App or your Whizi account does not cancel store billing. You must use the applicable store's subscription-management settings to cancel."

#### Exactly what deletion removes

The privacy policy lists the current deletion flow item by item, and this is the whole list rather than a summary of it:

- The live Whizi account record
- Synchronized chats and messages
- Uploaded and generated media stored under the user account
- Saved mobile memory
- Referral records
- Mobile entitlement and usage state
- Rate-limit state
- Push tokens
- The Clerk authentication account, which is the sign-in itself
- The local chat cache on the device completing deletion

All of it is subject to the limited retention described in the next section. The local cache matters more than it sounds: on mobile, Whizi may keep a cache of up to 40 recent conversations on the device, and that cache is not a separate cloud backup. It is cleared as part of a successful in-app deletion, and it can also be removed by clearing app storage or uninstalling the app. If you completed deletion on one device and a second device still holds a cache, clearing storage there is the direct fix.

Deleting the whole account is not the only lever. Individual conversations and uploaded files can be deleted on their own, and uploaded attachments and generated media are configured to expire after up to 30 days regardless. The retention table for everything else is on [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

#### What survives deletion, and why

An honest answer here is more useful than a clean one, so here is every carve-out the policy names.

| What can remain | For how long | Why |
| --- | --- | --- |
| Provider caches at OpenRouter and routed model providers | Varies by provider and endpoint | Whizi cannot instantly remove a provider cache that has already been created |
| Cloudflare D1 point-in-time recovery history | Up to 30 days before aging out | Protected recovery history covering deleted database records |
| AI-output reports | As long as reasonably needed | Stored separately from chat history for safety and security review |
| Limited account, transaction, security, tax, fraud-prevention or dispute records | As legally or operationally required | Legal and operational obligations |
| Payment processor records at Apple, Google, RevenueCat and Stripe | Periods required by law or their compliance obligations | Transaction, receipt, tax, fraud, chargeback and dispute records |

One clarification that answers the question people usually mean by this: "Whizi does not keep deleted chat content or uploaded or generated media merely because a payment record must be retained." A retained invoice is not a retained conversation.

An AI-output report is the exception worth understanding, because it is the only category the policy says may outlive the account by design. A report may contain a bounded excerpt of the reported output plus related account or technical identifiers, and reports may remain after account deletion only for as long as reasonably needed for safety, fraud prevention, dispute resolution, or legal obligations.

#### Export your data before you delete

Export Data sits in Settings under Account, and the export endpoint is described in the code as the access counterpart to account deletion. It returns the account as a single downloadable JSON file.

| In the export | Not in the export |
| --- | --- |
| Chats and messages | Quota counters |
| Memory | Rate windows |
| Projects and their file metadata | Push tokens |
| Media, as metadata plus a signed URL | |
| Share links | |
| Subscription | |
| Referral profile | |

The three omissions are deliberate: they are transient bookkeeping rather than content, and they are the same categories deletion wipes anyway. Media is exported as metadata plus a signed URL rather than inlined bytes, which is what keeps the file readable JSON instead of an unopenable blob, so download anything you actually want to keep while those links still resolve.

Export is rate limited to 5 requests per minute. If a first attempt fails, the toast is "Failed to export your data. Please try again." and the underlying error is "Export request failed ({status})." with the HTTP status filled in.

#### Deleting guest data when you never made an account

Whizi on mobile can be used without an account, which creates an anonymous device-bound guest profile that needs no name or email address. That profile holds real chats, so it has its own deletion path.

Select Delete guest data in Settings and confirm the deletion twice. It removes the anonymous guest profile, Whizi chats and messages, uploaded and generated media, saved memory, entitlement and usage state, guest identity links, and the local guest data on the device completing deletion, subject to the same limited retention.

Export works for guest sessions too, which is the point worth knowing before a device is wiped or handed on: a guest holds chats and memory that exist nowhere else, so exporting first is the only way to keep them.

#### If deletion fails

The failure surface is short. In Settings, a failed deletion shows "Your account could not be deleted. Please try again, or contact support." The raw error behind it is "Account deletion failed ({status})." with the HTTP status filled in.

Retrying is the first move. If it fails again, the contact route is the same one the policy names for people locked out of their account: support at [whizi.io/support](https://whizi.io/support), or the instructions at [whizi.io/account-deletion](https://whizi.io/account-deletion).

Whizi does not publish a completion time for deletion, and there is no stated "within X days" figure to quote. The only related number the policy commits to is the Cloudflare D1 recovery history of up to 30 days before deleted records age out.

#### Deleting is not cancelling

These are two separate actions and doing one does not do the other. The terms of service state it directly: "Account deletion and subscription cancellation are separate."

Cancelling stops billing and ends access at the end of the paid term while the account and its contents stay put. Deletion removes the account and its contents. If you subscribed through Apple or Google, deletion also leaves that subscription renewing until you cancel it with the store. The order that avoids both mistakes is cancel first, confirm the cancellation, then delete. See [billing and cancellation](https://whizi.io/docs/billing-and-cancellation) for where to cancel by how you signed up.

Key points:

- The complete removal list, item by item, from the privacy policy
- The records that survive deletion, and the reason each one does
- How to export everything as one JSON file before you delete

Checklist:

- Deletion is under Settings, in the app or on the website, with a two-step confirmation
- Export Data in Settings returns the account as one JSON file
- Cancel an Apple or Google subscription before deleting, because deletion does not stop store billing
- Deletion removes the account record, chats, messages, media, memory, referral records, usage and rate-limit state, push tokens, the Clerk sign-in and the local cache
- Deleted database records may sit in Cloudflare D1 recovery history for up to 30 days
- AI-output reports may remain after deletion for safety and legal reasons
- No account needed: Delete guest data in Settings, confirmed twice
- If you cannot sign in, use whizi.io/account-deletion or contact support

Steps:

1. Export your data first: Settings, under Account, has Export Data. It returns your account as one downloadable JSON file. Deletion is not reversible, so do this before anything else if you want a copy.
2. Cancel a store subscription separately: If you subscribed inside the iOS or Android app, cancel with Apple or Google. Deleting the account does not stop store billing.
3. Open Settings: Account deletion sits under Settings, in the mobile app or on the website.
4. Confirm twice: The flow uses a two-step confirmation. Both confirmations are required before anything is removed.
5. If you cannot sign in, use the instructions page: Follow the instructions at whizi.io/account-deletion, or contact support if you cannot access the account at all.

FAQ:

Q: How do I delete my Whizi account?
A: Under Settings, in the mobile app or on the website, using the two-step confirmation flow. If you cannot get into the account, follow the instructions at whizi.io/account-deletion or contact support. Export your data first if you want a copy, because deletion is not reversible.

Q: What exactly does account deletion remove?
A: The live Whizi account record, synchronized chats and messages, uploaded and generated media stored under the account, saved mobile memory, referral records, mobile entitlement and usage state, rate-limit state, push tokens, the Clerk authentication account, and the local chat cache on the device completing deletion. All of that is subject to a limited retention window for provider caches, recovery history, safety reports and records kept for legal, tax, fraud-prevention or dispute reasons.

Q: Does deleting my Whizi account cancel my subscription?
A: Not if you subscribed inside the iOS or Android app. Apple and Google own that billing relationship, and the terms say deleting the app or the Whizi account does not cancel store billing. Cancel through the store first, confirm it, and delete afterwards.

Q: Can I get my data out before deleting?
A: Yes. Export Data in Settings under Account returns one downloadable JSON file containing chats and messages, memory, projects and their file metadata, media, share links, subscription and referral profile. Quota counters, rate windows and push tokens are left out as transient bookkeeping. Media arrives as metadata plus a signed URL rather than inlined bytes, so download the files you want while those links resolve. Export is rate limited to 5 requests per minute, and it works for guest sessions too.

Q: How long does deletion take to complete?
A: Whizi does not publish a completion time, and no "within X days" figure exists to quote. The one related number the privacy policy commits to is Cloudflare D1 point-in-time recovery history, where deleted database records may remain for up to 30 days before aging out. Provider caches already created at a model provider are governed by that provider policy rather than by Whizi.

Q: How do I delete my data if I never created an account?
A: Select Delete guest data in Settings and confirm the deletion twice. That removes the anonymous guest profile, chats and messages, uploaded and generated media, saved memory, entitlement and usage state, guest identity links, and the local guest data on the device completing deletion. Export works for guests as well, so take a copy first if the chats matter.

### How to generate Excel, Word, and other real files with AI in Whizi

URL: https://whizi.io/docs/export-chats-and-files/
Updated: August 2026

Quick answer: Ask a model in Whizi for a downloadable spreadsheet or document and the answer arrives as a file card rather than a table inside a chat message. The file is generated in your browser as a genuine .xlsx or .docx that opens in Excel, Numbers, Word or Google Docs. One request can produce a multi-tab workbook.

Ask a model for a spreadsheet or a document and download a genuine .xlsx or .docx, plus charts as images and HTML you can preview before you use it.

#### The short answer

Ask a model in Whizi for a spreadsheet or a document and you get a downloadable file card rather than a formatted table stuck inside a chat message. The file is generated in your browser as a genuine `.xlsx` or `.docx` that opens in Excel, Numbers, Word or Google Docs.

This is the difference that matters against copying a markdown table out of a chat window: the columns are real columns, numbers are real numbers rather than text, and a multi-sheet workbook arrives as a multi-sheet workbook.

#### Asking for a spreadsheet

Be explicit that you want a file. "Put this in a table" gets you a table in the message; "give me this as a downloadable spreadsheet" gets you a file card.

Three things worth specifying in the same request, because they are far easier to state up front than to fix afterwards:

- **The columns you want, by name.** Otherwise you get the model idea of a sensible schema, which is usually close and never exactly right.
- **What to do with missing values.** Say "leave the cell empty rather than guessing" if the source is incomplete, because inventing plausible values is the failure mode here.
- **Separate tabs, if you want them.** A single request can produce a multi-tab workbook, so "one tab per region" or "raw data on one tab and the summary on another" both work.

The strongest input for this is an uploaded `.xlsx` or `.csv` rather than a PDF, because the model receives real cell structure rather than extracted prose. Spreadsheet in, spreadsheet out is by some distance the most reliable version of this workflow.

#### Documents, charts and HTML

**Documents.** The same approach produces a real `.docx` with headings, paragraphs and tables. Useful when the output is going to someone who expects a document rather than a chat log: a summary of a report, a first draft of a policy, meeting notes structured into sections.

**Charts.** When a model produces a chart, it renders as an actual chart in the conversation, in bar, horizontal bar, line, area or pie form, and you can download it as an image. That makes it usable in a deck or a document without rebuilding it.

**HTML.** Generated HTML renders in a preview pane alongside the code, so you can look at the thing before deciding whether to use it, and switch between the rendered view and the source. It can also be downloaded.

| What you ask for | What arrives | What it carries |
| --- | --- | --- |
| A downloadable spreadsheet | A file card that generates a genuine `.xlsx` in your browser | Real columns and real numeric cells, and separate tabs from a single request |
| A downloadable document | A real `.docx` | Headings, paragraphs and tables, opening in Word or Google Docs |
| A chart | A chart rendered in the conversation, downloadable as an image | Bar, horizontal bar, line, area or pie |
| Generated HTML | A preview pane alongside the code, also downloadable | The rendered view and the source, switchable |
| A table in the message | A markdown table you copy out by hand | Formatting only: the columns are not real columns |

#### Check the numbers

A generated spreadsheet looks authoritative in a way a chat message does not, and that is exactly the risk. The formatting being correct tells you nothing about whether the figures are.

Two checks that catch most problems in under a minute. First, spot-check three or four values against the source, chosen from different parts of the data rather than the top. Errors cluster where the source was ambiguous, which is rarely row one. Second, check the row count. A silently truncated extraction is the most common failure, and it is invisible unless you count.

If the data matters, the cheap verification is to ask a second model to extract the same thing from the same source and compare the two files. Where they agree, confidence is reasonable; where they diverge is exactly where you should look. See [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side).

Key points:

- Real .xlsx and .docx files, not a table you have to copy out of a message
- Multi-tab spreadsheets from a single request
- Charts you can download as images, and HTML you can preview before trusting

Checklist:

- Say "downloadable spreadsheet" rather than "table" to get a file
- Name the columns you want in the request
- Tell the model to leave gaps empty rather than guessing
- One request can produce a multi-tab workbook
- Upload .xlsx or .csv rather than PDF when the source is tabular
- Spot-check three or four values, not from the top of the data
- Check the row count for silent truncation

Steps:

1. Give the model the data: Upload a file or paste the source material into the conversation.
2. Ask for a spreadsheet: Say explicitly that you want it as a downloadable spreadsheet, and name the columns you want.
3. Wait for the file card: The answer renders as a file card rather than a table in the message.
4. Download it: The card generates a real .xlsx in your browser.
5. Check the numbers: Open it and verify the figures against the source before you use it.

FAQ:

Q: Can Whizi create a real Excel file?
A: Yes. Ask for a downloadable spreadsheet and the answer arrives as a file card that generates a genuine .xlsx in your browser, with real columns and real numeric cells rather than a markdown table you have to copy out. Multi-tab workbooks work from a single request if you say which tabs you want.

Q: Can I get a Word document instead?
A: Yes, the same way. Ask for a downloadable document and you get a .docx with headings, paragraphs and tables that opens in Word or Google Docs. This is the right format when the output is going to someone who expects a document rather than a chat transcript.

Q: Can I download a chart the AI made?
A: Yes. Charts render as real charts in the conversation, in bar, horizontal bar, line, area or pie form, and can be downloaded as an image for use in a deck or a document. You do not have to rebuild the chart yourself from the underlying numbers.

Q: How do I export an entire conversation?
A: For sending a conversation to someone else, sharing it as a link is usually what you want: the recipient reads it in a browser with no account, and can see which model produced each answer. For getting the content of a conversation into a document, ask the model to write the summary as a downloadable .docx, which produces something more useful than a raw transcript.

### A file upload was rejected or failed in Whizi

URL: https://whizi.io/docs/file-upload-failed/
Updated: August 2026

Quick answer: Whizi caps a single attachment at 10 MB and all the attachments on one prompt at 20 MB together, and the error string names which layer refused the file. The backend accepts 4 attachments per prompt, on the final user message only, while the website composer stages up to 6. Attachments cost no credits.

Every upload error message Whizi can show, the exact condition that triggers it, and what to do about each one.

#### The short answer

Whizi caps a single attachment at 10 MB and all the attachments on one prompt at 20 MB together, and the error string names which layer refused the file.

The caps, in one place:

| Cap | Value | Where it applies |
| --- | --- | --- |
| Single attachment | 10 MB | Any file attached to a chat message |
| All attachments on one prompt | 20 MB | The prompt as a whole |
| Attachments per prompt | 4 | Enforced by the backend, on the final user message only |
| Files staged in the composer | 6 | The website composer |
| Direct image upload | 8 MB | The image upload route |
| Image uploads per minute | 10 | Per account, per UTC minute |
| Image uploads per day | 60 | Per account, per UTC day |
| Project pinned file | 10 MB, 10 files | Per project |
| Text in one chat message | 100,000 characters | Includes text extracted from your files |

Attachments cost no credits, so a refused upload costs you nothing and neither does a successful one.

#### The file is too large

Four different strings can appear on a size refusal, depending on which layer caught the file.

| Message | HTTP code | Trigger |
| --- | --- | --- |
| `File {name} exceeds 10MB limit. Please upload a smaller file.` | none, a website toast | You picked a single file over 10 MB through the attach button |
| `One or more files exceed the 10MB size limit. Please upload smaller files.` | none, a website toast | You dropped files and at least one is over 10 MB |
| `Attachments must be 10 MB or smaller.` | 413 `attachment_too_large` | The backend refused the file, checked first against the declared length and again against the real bytes |
| `Attachments are limited to 20 MB per prompt.` | 413 `attachments_too_large` | Each file is under 10 MB but they add up to more than 20 MB |

The `{name}` in the first two is filled in with your file name.

The declared length is checked before the real bytes are, so re-declaring a smaller size does not get a large file past the 10 MB check.

The fix for all four is the same: send less at once. Split a long PDF into the sections you actually want read and attach the section you are asking about. Only the most recent user message carrying attachments has its attachments forwarded to the provider, so attach the file to the message that asks the question rather than relying on one sent earlier in the conversation.

Directly uploaded images have their own, lower cap: `Image is too large. Please upload an image under 8 MB.` is HTTP 413 `upload_too_large`. It fires twice over, once on a declared `Content-Length` at more than twice the cap and again on the real file size at 8 MiB, so a mis-declared length is caught either way. Re-export the image at a smaller size or a lower resolution.

For project pinned files the cap is also 10 MB, with its own strings: `{name} is over the 10 MB limit.` as the website toast and `Project files must be 10 MB or smaller.` as HTTP 413 `project_file_too_large`.

#### The file type is not accepted

Three different allowed-type lists exist, because chat attachments, direct image uploads and project pinned files are three different paths.

| Message | HTTP code | What it accepts |
| --- | --- | --- |
| `Choose an image, PDF, text, Markdown, CSV, or JSON file.` | 415 `unsupported_attachment_type` | Chat attachments: JPEG, PNG, WebP, GIF, PDF, TXT, Markdown, CSV, JSON |
| `Unsupported image type. Please upload a JPG, PNG, WebP, or GIF.` | 400 `unsupported_image_type` | The image upload route: `image/jpeg`, `image/png`, `image/webp`, `image/gif` |
| `Choose a PDF, Word, Excel, text, Markdown, CSV, or JSON file.` | 415 `unsupported_project_file_type` | Project pinned files, which deliberately exclude images |
| `Unsupported file type. Please upload an image, document (PDF, Word), spreadsheet (Excel), or text-based file.` | none, a website toast | The composer classifying a picked file |
| `{name} is not a supported file type.` | none, a website toast | A picked project file that is an image, or otherwise unsupported |

The composer shows that unsupported-type toast when you pick a file through the attach button. Drag and drop stays silent on the same case, so a dropped file that never appears in the composer and produces no message at all is most likely an unsupported type. Pick it through the attach button instead and you will get the toast that names the problem.

GIFs are the one place the two paths disagree. The website composer rejects them outright with `GIF files are not supported. Please upload a PNG or JPEG image.`, while the backend image upload route accepts `image/gif`. On the website, convert the GIF to a PNG or JPEG frame and upload that.

For anything else off the list, the fastest route is a format conversion: save it as PDF, CSV or plain text and upload that instead. The full accepted list is in [supported file types](https://whizi.io/docs/supported-file-types).

#### The file does not match what it claims to be

Two checks run past the type list, and both produce their own message.

`The file extension does not match its content type.` is HTTP 415, code `attachment_extension_mismatch` on a chat attachment and `project_file_extension_mismatch` on a project file. It fires when the extension on the file name and the content type sent with it disagree.

`The selected file does not match its declared format.` is HTTP 415, code `invalid_attachment` on a chat attachment and `invalid_project_file` on a project file. This one is a signature check: the first bytes of the file are read and compared against the declared MIME type, so a file that has been renamed rather than converted is caught here.

The usual cause is a rename. Changing `report.txt` to `report.pdf` does not make it a PDF, and this check is what notices. Open the file in its own application and export or save-as the format you want, then upload the result.

#### Too many files, or too many uploads

Count limits produce their own messages, and they are not all the same number, because the composer and the backend count different things.

| Message | HTTP code | Trigger |
| --- | --- | --- |
| `You can only upload up to 6 files per message.` | none, a website toast | A pick or a drop would take the composer past 6 staged files |
| `Message {n} may carry up to 4 valid attachments, and only on the final user message.` | 400 `invalid_attachments` | More than 4 attachments on a prompt, or attachments on a message that is not the last user turn |
| `Message {n} repeats an attachment.` | 400 `invalid_attachments` | The same attachment id appears twice on one message |
| `Projects hold up to 10 files. Remove one first.` | 409 `project_file_limit_reached` | An 11th pinned file on a project |
| `This project already holds 10 files. Remove one first.` | none, a website toast | Shown before the upload even starts, when the list is already full |

Rate limits are separate from counts. `Upload rate limit reached. Please wait a moment before uploading another image.` is HTTP 429 `upload_rate_limited`, and it fires after more than 10 image uploads in a single UTC minute. The wait is a minute.

`Daily upload limit reached. Please try again tomorrow.` is HTTP 429 `upload_daily_limit_reached`, after 60 image uploads in one UTC day. The reset is on UTC midnight rather than midnight where you are. Project file uploads have their own ceiling of 10 per minute.

#### The upload worked but the file could not be read

When extraction of a PDF, Word or spreadsheet file throws, the upload itself succeeded and the failure is in reading the contents. Each format has its own message.

| Message | What it usually means |
| --- | --- |
| `Could not read {name}. The PDF may be corrupted or password-protected.` | PDF text extraction threw. A password-protected PDF is the common case |
| `Could not read {name}. Please try re-saving the document.` | Word extraction threw |
| `Could not read {name}. The spreadsheet may be corrupted.` | Spreadsheet extraction threw |
| `Could not read {name}. Older .xls files are not supported. Please re-save the file as .xlsx and upload it again.` | A legacy binary Excel file. The message is the fix |
| `Could not read {name}.` | Plain text extraction threw |
| `No readable text found in this file, so the model cannot see it.` | A project pinned file whose extraction produced nothing at all |

For a password-protected PDF, remove the password in your PDF reader and upload the unlocked copy. For an older `.xls`, open it and save as `.xlsx`. A scanned PDF that extracts to nothing is covered in [chat with a PDF](https://whizi.io/docs/chat-with-pdf).

A file can also be read successfully and then trimmed. A single chat message is capped at 100,000 characters, and extracted file text counts toward it. When the extracted text does not fit, you see `Your upload was too large, so only the first part of it was sent. Ask about a smaller section for full coverage.` and the message itself carries the marker `[Attachment truncated: the upload was larger than one message can carry, so the content past this point was not included.]` at the cut. The model receives the text up to that marker and nothing after it, so an answer that reads as complete may be based on a partial document.

If the whole message is refused rather than trimmed, the backend returns HTTP 400 `message_too_long` with `This message is {count} characters, over the 100,000 character limit. Attach a shorter file, or ask about one section at a time.`, where `{count}` is your actual character count.

#### The upload failed for no obvious reason

A handful of strings mean the upload never completed rather than being refused on its merits.

| Message | What it means | What to do |
| --- | --- | --- |
| `Failed to retrieve session token.` | Your session token could not be fetched before the upload call | Refresh the page and sign in again, then retry |
| `Failed to upload image: {statusText}` | The image upload response came back not OK, with the server status text appended | Retry. If it repeats, the status text is what support needs |
| `Image URL not returned by the server.` | The upload succeeded but the response carried no image URL | Retry the upload |
| `Could not upload your attachments.` | The upload step threw in the side-by-side compare view | Retry, or send the same prompt in a normal chat |
| `Could not upload {name}.` | A project file upload threw with no message of its own | Retry |
| `Invalid request` and `No file uploaded` | HTTP 400 `invalid_upload` and `no_file_uploaded`. The request was not `multipart/form-data`, the form data failed to parse, or the body carried no file field | Retry the upload |
| `Media could not be stored.` | The storage write returned no object | Retry once. If it persists, contact support |
| `Stored media metadata was invalid.` | The stored media row could not be read back | Retry once. If it persists, contact support |
| `Something went wrong. Please retry.` | HTTP 500. Any unhandled backend error | Retry once, then contact support |

A different family of messages means the file is gone rather than failing to arrive. `One or more attachments could not be found.` is HTTP 404 `attachment_not_found`, returned when an attachment id does not resolve to stored media owned by you. `Media not found.` is HTTP 404 `media_not_found`, and `Media has expired.` is HTTP 404 `media_expired`, returned once the stored object is past its retention window, which is up to 30 days. An old conversation can still hold a reference to a file that is gone, so re-upload it. The retention wording is on [data and privacy](https://whizi.io/docs/data-and-privacy).

Key points:

- Every upload error string Whizi shows, quoted exactly, with its trigger
- The size, type and count caps that actually decide whether an upload lands
- Why a file can upload cleanly and still leave the model with nothing to read

Checklist:

- One attachment is capped at 10 MB, one prompt at 20 MB total
- The composer stages up to 6 files; the backend accepts 4 attachments per prompt
- Direct image uploads are capped at 8 MB, 10 per minute and 60 per UTC day
- Renaming a file does not convert it: the signature check catches it
- Drag and drop stays silent on an unsupported type, the attach button tells you
- GIFs are rejected by the website composer, so convert to PNG or JPEG
- Older .xls files must be re-saved as .xlsx
- A password-protected PDF must be unlocked before it can be read
- Extracted file text counts against the 100,000 character message cap
- Attachments cost no credits, so a failed upload costs nothing

FAQ:

Q: What does "Attachments must be 10 MB or smaller." mean?
A: The file you attached is over the 10 MB per-file cap, and the backend refused it with HTTP 413 and the code attachment_too_large. That string is the per-file cap only. Four files that are each under 10 MB but add up to more than 20 MB return a different string, "Attachments are limited to 20 MB per prompt.", with the code attachments_too_large. A fifth attachment on the same prompt returns "Message {n} may carry up to 4 valid attachments, and only on the final user message." with the code invalid_attachments.

Q: Why does "The file extension does not match its content type." appear on a file that opens fine?
A: The extension check runs before the signature check, so a file whose extension disagrees with its declared content type is refused with this string and never reaches the magic-byte comparison that returns "The selected file does not match its declared format." The codes differ by upload path: a chat attachment returns attachment_extension_mismatch and invalid_attachment, a project file returns project_file_extension_mismatch and invalid_project_file, so the code in a network trace tells you which upload path refused the file.

Q: I dropped a file into the chat and nothing happened, with no error. Why?
A: The most likely reason is an unsupported file type, and the drop path deliberately says nothing about it. The dropped file also never appears in the staged list, so there is nothing to remove and nothing on screen to explain the silence. Not every drop is silent: a drop that would take the composer past 6 staged files does raise a toast, "You can only upload up to 6 files per message.", on the same path where an unsupported type raises none.

Q: How do I get past "Upload rate limit reached. Please wait a moment before uploading another image."?
A: You wait. It is HTTP 429, triggered by more than 10 image uploads in one UTC minute. The counter behind it belongs to the image upload route rather than to attachments in general, and the project file route keeps a separate one, so pinning files to a project and uploading images are not drawing down the same allowance. Whichever route you are on, the account also pays a shared ceiling of 30 requests per minute across every route, which a fast burst can reach before the upload limit does.

Q: My PDF uploaded but the model says it cannot see the contents. What went wrong?
A: The upload landed and the extraction step is what failed. A password-protected or corrupted PDF produces the toast "Could not read {name}. The PDF may be corrupted or password-protected." A PDF pinned to a project whose extraction produced nothing shows "No readable text found in this file, so the model cannot see it." beside the file in the project instead. Remove the password and re-upload for the first case. The second case, a scan with no text layer, is covered on the chat with a PDF page.

Q: Does upgrading my plan change anything about uploads?
A: It changes which models you can send a file to. Attachments cost no credits on any plan, so the upload itself is never the thing your plan is metering. Starter reaches four model identifiers, Pro adds 37 more on top of those, and every other model in the catalogue requires Powerhouse.

### Can you fork a conversation in Whizi?

URL: https://whizi.io/docs/fork-a-conversation/
Updated: August 2026

Quick answer: Whizi has no fork control, no branch control, and no way to duplicate a chat, because a conversation stores exactly one active branch. The chat options menu offers rename, pin, star and delete. The nearest equivalent is editing a message you already sent, which replaces the tail and files the old one as a version.

Whizi has no fork, branch, or duplicate control. A conversation holds exactly one active branch, and this is what to use instead depending on which kind of fork you wanted.

#### The short answer

Whizi has no fork control, no branch control, and no way to duplicate a chat. A conversation cannot be split into two chats that both stay live, and there is no "continue from here as a new conversation". The chat options menu in the sidebar offers rename, pin, star and delete, and nothing that copies a conversation.

What Whizi does have is message versions. Editing a message you already sent replaces the tail of the conversation and files the tail it replaced as a version you can flip back to. That is a branch you can visit, not a branch you can run in parallel: exactly one version is live at any moment.

"Fork" means four different things in practice, so start from which one you wanted:

| What you wanted | What to use | What actually happens |
| --- | --- | --- |
| A different answer from a midpoint | Edit that message and send it again | The tail is replaced, and the old tail is archived as a version |
| Both tails kept so you can compare | The version arrows on the edited turn | Both are stored, one is live, you flip between them |
| Two conversations that both stay live from the same history | Not available | There is no duplicate, copy or fork anywhere in the product |
| A side question that does not pollute the thread | A new chat | Nothing carries over except what you paste in |
| The same standing context across several chats | A project | Instructions and pinned files apply to every chat in it |
| A copy someone else can continue | Not available | A share link is a read-only view |

#### Why there is no fork

This is a structural answer rather than a roadmap answer. A Whizi conversation stores exactly one active branch and nothing else. Every reader of a conversation sees that one branch: the website, the mobile app, share links, and the transcript the model itself receives. Inactive versions are kept off to one side, as plain copies of the rows they would be restored as.

That invariant is what makes version switching cheap, and it is also what rules forking out. A fork needs two branches that are both live, both writable, and both visible. One active branch means the second one has nowhere to exist.

The route surface matches. Beyond the conversation list itself, the only things you can address on a conversation are its messages, its versions, and the project it belongs to. There is no endpoint behind a fork or a duplicate, so no client can offer one.

The consequence worth planning around: editing a turn discards the stored versions of every turn beneath it. Those versions belong to a conversation that no longer exists, so keeping them would let you walk into a branch whose parent is gone. If several branches in one thread matter to you, work backwards and edit the earliest message last. The mechanics of editing, and what a version switch costs and can fail with, are on [regenerate and edit messages](https://whizi.io/docs/regenerate-and-edit-messages).

#### What a new chat carries over

Starting a new chat is the honest answer to a side question, but it is worth being precise about what a new chat starts with, because "start a new chat" is only a real answer if you know what you are giving up.

A new chat starts empty. The previous conversation is not summarized into it, attachments are not re-attached, and the model has no view of the other thread. Whatever you want carried across, you carry across: every message and every code block has a copy control, so lifting the two or three answers that matter into the new chat takes a few seconds.

Two things do carry without pasting. Project instructions and pinned files apply to every conversation inside that project, so a new chat created in a project starts with that standing context already folded into the prompt. And a chat can be moved into a project or back out of it after the fact, which is the closest thing Whizi has to reparenting a conversation. The caps and the behavior are on [how to use Projects](https://whizi.io/docs/use-projects).

The other route people reach for is model switching, and it is worth saying what it is not. Switching models inside a conversation carries the whole thread to the new model rather than copying it into a second chat. There is one thread throughout, and the replies above the switch stay credited to whichever model wrote them. The patterns are on [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### The two cases a fork would have solved

**Comparing two answers to the same prompt.** This is the most common reason to want a fork, and Whizi answers it a different way: send the prompt to two models in parallel and read the outputs next to each other. Neither model sees the other output, which is the entire value of the comparison. That is on [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side).

**Keeping part of a thread away from a model.** A fork is the usual workaround for a conversation that has accumulated context you would rather not resend. Without one, the move is a new chat seeded with a short summary of what actually matters. That is not purely a workaround: a long thread makes older context compete for attention with the real question, so a clean start with three sentences of setup often reads better than forty messages of history.

If what you want is a copy for someone else rather than for yourself, sharing produces a public link to one conversation and the recipient reads a sanitized, read-only view of it. They cannot continue it, and it is not a fork in their account either. The details, including revoking a link, are on [sharing a conversation](https://whizi.io/docs/share-a-conversation).

For a copy of everything you have, Settings has an Export data action that produces your chats and messages, projects and their file metadata, media, share links, subscription and referral profile. It is an archive rather than a re-importable branch: Whizi imports ChatGPT and Claude exports, not its own.

Key points:

- Why a Whizi conversation can only ever hold one live branch
- The four different things people mean by "fork", and what covers each one
- What a new chat carries over from an old one, which is nothing you do not paste

Checklist:

- There is no fork, branch, or duplicate control in Whizi
- A conversation stores exactly one active branch, which is why forking has nowhere to live
- The chat options menu is rename, pin, star and delete, with no copy
- Edit and resend is the closest equivalent: it replaces the tail and archives the old one as a version
- Only one version is live at a time, so versions are visited rather than run in parallel
- Editing a turn discards the stored versions of every turn beneath it, so edit the earliest message last
- A new chat starts empty; only project instructions and pinned files carry over without pasting
- A share link is a read-only view and cannot be continued by the recipient

FAQ:

Q: How do I fork a conversation in Whizi?
A: You cannot. There is no fork control, no branch control, and no duplicate action on a chat. The underlying reason is that a Whizi conversation holds exactly one active branch, and the only things addressable on a conversation are its messages, its versions, and the project it sits in. The nearest equivalent is editing one of your own messages and sending it again, which replaces the tail from that point and files the replaced tail as a version.

Q: Can I duplicate a chat so the original stays untouched?
A: No. The chat options menu offers rename, pin, star and delete, and there is no copy anywhere else in the product either. If the original matters, the practical protection is to stop editing it: an edit is what destroys a tail. Start a new chat, paste in the few messages that matter, and continue there, which leaves the original exactly as it was.

Q: What is the difference between a version and a fork?
A: A version is one of several stored tails for the same turn, and exactly one of them is live at a time. You move between them with the arrows on the edited turn, and switching swaps the whole tail rather than one bubble. A fork would be two tails that are both live and both writable at once, and that is the part Whizi does not do. Versions are covered on [regenerate and edit messages](https://whizi.io/docs/regenerate-and-edit-messages).

Q: If I start a new chat, does the model remember the old one?
A: Not from the old conversation itself. A new chat starts empty and the model receives only what is in that chat. The exception is a project: instructions and pinned files apply to every conversation inside it, so a chat created in a project already carries that context. Anything else has to be pasted in, and every message carries a copy control for exactly that.

Q: Can I branch a conversation to try a different model?
A: You do not need a branch for that. Switching the model inside a conversation carries the existing thread to the new model, and each reply stays credited to whichever model produced it, so one thread holds the whole comparison. If you want two answers to the same prompt rather than a handover, send it to two models in parallel instead.

Q: Will editing an early message lose my later branches?
A: Yes, and this is the one surprise worth knowing in advance. Editing a turn discards the stored versions of every turn after it, because those versions hang off a conversation that no longer exists. Copy anything from a lower branch that you want to keep before you touch a message above it, and if you are reworking several turns, edit the earliest one last.

### How to generate images inside a Whizi chat

URL: https://whizi.io/docs/generate-images-in-chat/
Updated: August 2026

Quick answer: Ask for an image inside any Whizi chat, either by typing a prompt such as Generate an illustration of, or by clicking the image icon. Pick Flux 2 Pro for photorealism, Stable Diffusion for open-style images, or DALL·E for general use. Iterate in the same thread, then download the result.

Turn a conversation into a generated image without leaving the thread. Use top image models (Flux, Stable Diffusion, DALL·E) side-by-side in one workspace.

#### Why generating in the chat is different

The usual image workflow is a separate product, a separate subscription, and a separate blank box that knows nothing about what you are working on. Generating inside the conversation changes two practical things.

First, the model already has the context. If you have just spent twenty messages developing a campaign concept, "generate the hero image for this" carries all of it. You are not re-describing the brief to a different tool.

Second, you can use a text model to write the image prompt. This is the trick most people miss: ask [Claude](https://whizi.io/docs/using-claude-in-whizi) or [GPT](https://whizi.io/docs/using-gpt-in-whizi) to turn your rough idea into a properly structured image prompt, then run it. Language models are considerably better at writing image prompts than most people are, because the format is a known one and they have seen a great many examples of it.

**Prompt: let the text model write the image prompt**

`Write an image generation prompt for: [rough idea]. It will be used for [purpose and placement]. Structure it as: subject, action or state, setting, lighting, composition and framing, medium or lens, colour treatment, mood. Add a negative list of what must not appear. Give me three variations that differ in composition rather than in adjectives.`

#### The prompt structure that works

Nearly every good image prompt has the same seven parts in roughly this order. Missing parts get filled in by the model with its defaults, and the defaults are why generic prompts produce generic stock imagery. If this is your first time writing one, [how to create AI images](https://whizi.io/resources/image-generation) covers the same ground from scratch.

| Part | What to say | Effect if you omit it |
| --- | --- | --- |
| Subject | The specific thing, with the detail that matters | You get the most generic version of the noun |
| Action or state | What it is doing, or how it sits | Static, posed, lifeless |
| Setting | Where, and how much of it is visible | A blank or cluttered default background |
| Lighting | Direction, quality, time of day | Flat, evenly lit, characterless |
| Composition | Angle, distance, where the subject sits in frame | Centred medium shot every time |
| Medium or lens | Photograph and focal length, or illustration style | Defaults toward glossy digital art |
| Colour and mood | Palette and feeling | Oversaturated, high contrast |

A worked example. Weak: `a person working in an office`. Strong: `A woman in her fifties reviewing printed drawings at a standing desk, late afternoon light from a window to her left, shot from slightly behind her shoulder at eye level, 50mm, muted greens and warm neutrals, calm and unhurried. Negative: stock-photo smiling at camera, cluttered desk, visible brand logos, text.`

The negative list matters more than people expect. Most of what makes generated images look generated is a small set of recurring artefacts, and naming them removes them: `stock-photo smiling`, `over-saturated`, `text`, `watermark`, `extra fingers`, `symmetrical corporate composition`.

#### Choosing the model

- **Flux 2 Pro** for photorealism and anything that has to sit alongside real photography. It is the strongest choice for marketing imagery, product-adjacent scenes, and anything where a viewer would expect a camera to have been involved.
- **Stable Diffusion** for stylised and illustrative work, and where you want the widest range of aesthetic control. It rewards more specific prompting and gives more variety across seeds.
- **DALL·E** for quick illustrations, diagrammatic concepts, and casual generations where speed matters more than fine control.

You can switch between them inside the same chat, which is the fastest way to answer the only question that matters: run the same prompt through two models and look at them next to each other. Model quality varies enormously by subject, and the winner on portraits is often not the winner on interiors or on flat illustration.

#### Iterating without going in circles

The common failure is regenerating the same prompt over and over hoping for a better roll. Some of that is unavoidable, but most wasted attempts come from changing several things at once and losing track of what helped.

- **Change one variable per attempt.** Lighting, or framing, or palette. Not all three.
- **Keep what worked in words.** When an image is close, describe what is right about it in your next prompt rather than assuming the model remembers.
- **Fix composition before detail.** Getting the framing and the light right first is far more efficient than perfecting a subject that is in the wrong part of the frame.
- **Generate at the final aspect ratio.** Cropping a square into a wide banner throws away the composition you asked for. Ask for the ratio you actually need up front.
- **Save the prompts that worked.** Eight reliable house-style prompts you can paste again are worth more than any single image, and reusing them is what makes a set of assets look like a set.

For editing an image you already have, upload it and describe the change. This works well for adjustments to background, lighting, and palette, and less well for precise structural edits, which are still faster in a real image editor.

#### What image models still get wrong

**Text.** Rendered words remain the weakest area across every model. Short words sometimes come out; a headline, a logo, or a label usually does not. Generate the image clean and add text in your design tool.

**Hands, counts, and small repeated details.** Fingers, teeth, chair legs, and windows in a building are all places where models lose count. Zoom in before you use anything.

**Precise spatial instructions.** "Exactly three objects, the red one on the left" is unreliable. Compose in your editor if the arrangement genuinely matters.

**Consistency across images.** Getting the same character or product to appear identically in a series is difficult. Detailed, repeated descriptions help; identical results are not guaranteed.

On usage: check your rights before anything generated goes into a paid placement, and take particular care with anything resembling a real person, a recognisable location, or a distinctive style associated with a living artist. Provider terms differ, and the commercial exposure sits with whoever publishes the image.

Key points:

- A seven-part prompt structure that turns vague requests into usable images
- Which model to pick for photographic, stylised, and illustrative work
- The iteration discipline that gets you to a final image in four attempts instead of twenty

Checklist:

- Ask a text model to write the image prompt from your rough idea
- Include all seven parts: subject, action, setting, lighting, composition, medium, mood
- Add a negative list naming the artefacts you never want
- Generate at the aspect ratio you will actually use
- Change one variable per iteration
- Run the same prompt through two models and compare
- Add text in your design tool, not in the image model
- Save the prompts that worked as reusable templates

Steps:

1. Open a chat: Start a fresh conversation or continue an existing one.
2. Ask for an image: Type a prompt like "Generate an illustration of..." or click the image icon.
3. Pick the image model: Choose Flux 2 Pro for photorealism, Stable Diffusion for open-style images, or DALL·E for general use.
4. Iterate in-place: Ask for variations, edits, or different aspect ratios without leaving the thread.
5. Download or copy: Right-click the image to save, or copy the public share link.

FAQ:

Q: How many images can I generate per month?
A: It depends on the plan. Pro includes 100 generations per month and Powerhouse includes 500, which for most individuals and small teams covers regular campaign and content work. Since iteration is where generations are spent, writing a well structured prompt first (ideally with a text model) noticeably reduces how many attempts a usable image takes.

Q: Can I edit an existing image?
A: Yes. Upload the image and describe the change you want using an image-aware model. This works well for background, lighting, palette, and style adjustments. It works less well for precise structural edits such as moving an object a specific distance or changing text, which are still faster and more predictable in a normal image editor.

Q: Which image model should I use?
A: Flux 2 Pro for photorealism and anything that sits next to real photography. Stable Diffusion for stylised and illustrative work where you want more aesthetic control. DALL·E for quick illustrations and concepts. Because model quality varies a lot by subject, the fastest way to decide is to run the same prompt through two of them in the same chat and look at the results side by side.

Q: Why does the text in my image come out wrong?
A: Rendered text is still the weakest area across every image model, and no prompt reliably fixes it. Short single words sometimes work; headlines, logos, and labels usually do not. Generate the image without text and add typography in your design tool, which also gives you the correct font and proper control over placement.

Q: Do I need a separate image subscription alongside Whizi?
A: For most campaign, content, and concept work, no. Image generation is included on Pro and above, which covers the same ground a standalone subscription would for typical marketing and social output. A dedicated tool still makes sense for teams with very specific stylistic requirements or a deep existing prompt library they do not want to rebuild.

### How to use Whizi: getting started

URL: https://whizi.io/docs/getting-started/
Updated: August 2026

Quick answer: Setting up Whizi is four steps. Pick a plan by naming the one model you would be annoyed to lose. Set a cheap 1 credit default model, because that setting decides whether your monthly allowance lasts. Upload something you are genuinely working on. Then run one honest comparison and write down which model you will use for what.

The first twenty minutes: pick a plan, choose a default model, upload something real, and set up the two habits that make a multi-model workspace worth having.

#### The short answer

Four things, in order. Pick a plan by naming the one model you would be annoyed to lose. Set a cheap default model, because that single setting decides whether your monthly allowance lasts. Upload something you are genuinely working on rather than a test file. Then run one honest comparison and write down which model you will use for what.

The last step is the one people skip, and it is the one that stops you re-deciding on every prompt.

#### Pick the plan by the model you cannot do without

The plans differ mainly in model access, so the fastest way to choose is to name the one model you would be annoyed to lose.

If that model is Claude, you need Pro at minimum, because Starter includes no Anthropic model at all. If it is Claude Opus specifically, you need Powerhouse. If you have no strong opinion and mostly want a fast assistant that can read your documents, Starter is genuinely enough and costs $10.99 per month billed yearly.

Every plan starts with a 7 day trial for $0.99, and upgrading takes effect immediately, so starting one tier low and moving up mid-trial costs nothing. Full numbers are in [plans and limits](https://whizi.io/docs/plans-and-limits).

#### Set a cheap default model

This is the single setting that decides whether your allowance lasts the month. Roughly half the catalogue costs 1 credit per message, and those models are genuinely good at the majority of what anyone sends: lookups, rewrites, summaries, formatting, quick explanations.

Make one of them your default. Whizi AI, GPT-5.6 Luna and Gemini 3 Flash are all 1 credit. Then reach for the expensive tier deliberately, on the handful of messages per week that decide something. A Pro subscriber who defaults to Claude Sonnet 5 at 10 credits will burn 2,000 credits in 200 messages; one who defaults to a 1 credit model and switches up when it matters will not run out.

Every row in the model picker shows its credit cost before you send, so the price is never a surprise afterwards.

#### Upload something you are actually working on

The fastest way to find out whether this is useful is to give it real work rather than a test question. Drag a PDF, a Word document or a spreadsheet straight into the chat window. Upload is on every plan and costs no credits.

Then ask something you genuinely want to know, not something you already know the answer to. Testing an AI on questions you can already answer measures the wrong thing, and it is why so many trials end with a shrug.

Spreadsheets are worth trying early because the difference from a general chatbot is largest there: cell contents are extracted as structured data rather than flattened into prose. The full format list is in [supported file types](https://whizi.io/docs/supported-file-types).

#### Run one honest comparison

Take a task that matters and send it to two models. On Powerhouse you can do this in parallel with side-by-side; on any plan you can do it sequentially by switching models in the same conversation.

Decide what better means before you read either answer, because longer and more confident outputs are systematically preferred even when they are worse. Write one sentence describing what a good answer would contain, then read.

What most people find is that on the majority of their work the models are interchangeable, and on a minority the gap is large. Finding out which is which is the entire point of the exercise, and it takes about a week. See [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side).

#### Write down your routing and stop deciding

The failure mode of a workspace with 280 models is spending your attention on model choice instead of on work. The fix is to decide once and then stop.

A routing map that works for most people looks like three lines: a 1 credit default for everything routine, one strong model for writing and reasoning that matters, and one large-context model for documents. That is it. Three decisions, made once, and then model selection stops being a thing you think about.

Revisit it when a provider ships something new, which is roughly monthly, not daily.

Key points:

- Which plan to start on, based on whether you need Claude
- Why your default model should be a 1 credit one
- The routing habit that separates people who get value from this from people who do not

Checklist:

- Choose the plan by naming the one model you cannot do without
- Starter includes no Anthropic model, so Claude means Pro at minimum
- Set a 1 credit model as your default
- Upload real work, not a test file
- Ask questions you do not already know the answer to
- Run one comparison with the criterion decided in advance
- Write down a three line routing map and stop re-deciding

Steps:

1. Start the trial: Register and take the 7 day trial for $0.99 on the plan whose model list you actually need.
2. Set a default model: Pick a 1 credit model as your default so routine work costs nothing meaningful.
3. Upload something real: Drag in a PDF or spreadsheet you are actually working on rather than a test file.
4. Run one comparison: Send the same prompt to two models and see whether the gap is large enough to care about.
5. Decide your routing: Write down which two models you will use and for what, then stop thinking about model choice.

FAQ:

Q: Do I need to cancel ChatGPT Plus before trying Whizi?
A: No, and it is better not to for the first week. Run both in parallel during the $0.99 trial and send the same real tasks to each, because that comparison is the only reliable way to find out whether a metered multi-model allowance suits your work better than unlimited use of one model. Cancel whichever one loses at the end of the week.

Q: Can I import my ChatGPT conversation history?
A: Yes. Open Settings, choose Import Chats under Account, and pick the conversations.json file from your ChatGPT data export; the same dialog imports Claude exports too. The export file is parsed in your browser and only the normalized conversations are uploaded. The full walkthrough is in [import chats from ChatGPT](https://whizi.io/docs/import-chats-from-chatgpt).

Q: How long does it take to set up?
A: Registering and sending a first message takes under two minutes, since there is nothing to install on the web and no configuration. The twenty minutes in this guide is the part that actually determines whether you keep using it: setting a cheap default, uploading real work, and deciding your routing.

Q: What is Whizi AI, and should I use it as my default?
A: Whizi AI is the house assistant. It runs on GPT-5.6 Luna with a Whizi system prompt on top, costs 1 credit per message, and is available on every plan including Starter. It is a reasonable default for exactly that reason. If you would rather have no persona layer, the raw GPT-5.6 Luna is offered separately in the full model list at the same 1 credit price.

### How Whizi Auto picks a model for each message

URL: https://whizi.io/docs/how-auto-routing-works/
Updated: August 2026

Quick answer: Whizi Auto sorts each message onto one of six fixed rungs and answers on the model pinned to that rung, never above your plan. Auto carries no credit price of its own: a turn costs whatever the picked model costs, which is 1 credit on Starter, 1 to 10 on Pro, and 1 to 20 on Powerhouse.

The six fixed rungs Auto routes to, which of them each plan can reach, and what a turn on Auto costs in credits.

#### The short answer

Auto reads your message, sorts it into one of six fixed rungs, and answers on the model pinned to that rung, never above what your plan can open. Auto carries no credit price of its own: a turn costs whatever the model Auto picked costs, which is why the clients publish a range for your tier instead of a single number. Those ranges are 1 to 1 credits on Starter, 1 to 10 on Pro, and 1 to 20 on Powerhouse.

Auto is a row in the model picker rather than a model, it is available on every paid plan, and it sits in the Recommended section beside the house model and one current GPT, Claude and Gemini row. That section is constrained to models reachable at Pro or below, so the top of the picker is never Powerhouse-locked.

The whole ladder, cheapest rung first. The model column is reference data, quoted as the catalogue writes it:

| Rung | What sends a message there | Model | Credits | Lowest plan that reaches it |
| --- | --- | --- | --- | --- |
| Quick | Short questions and lookups | `whizi-ai` | 1 | Starter |
| Rewrite | Reformatting work such as rewrites and translation | `google gemini-3.7-flash` | 2 | Pro |
| Attachment | A turn carrying an image or an upload | `google gemini-3.7-flash` | 2 | Pro |
| Long-form | Large pasted text, over about 6,000 characters | `openai gpt-5.6-terra` | 4 | Pro |
| Code | Code | `anthropic claude-sonnet-5` | 10 | Pro |
| Reasoning | Multi-step reasoning | `openai gpt-5.6-sol` | 20 | Powerhouse |

Six rungs across five model rows, out of a catalogue of 280+. Every other family in the catalogue is a hand-pick in the model picker.

#### What Auto can pick on each plan

Auto never sends a turn to a model your plan cannot open. It walks back down the ladder until it finds a row your subscription can reach, so a message you already sent is never interrupted by a paywall.

| Plan | Rungs it reaches | Published credit range |
| --- | --- | --- |
| Starter, $15.99/month | Quick | 1 to 1 |
| Pro, $29.99/month | Quick through Code | 1 to 10 |
| Powerhouse, $49.99/month | All six | 1 to 20 |

Starter reaches exactly four picker entries: Auto, the house model Whizi AI, one base GPT row and one fast Gemini row. None of the five other rungs is in that set, so a Starter account stays on the house model at 1 credit no matter what it asks.

Pro tops out at 10 because the reasoning rung is a Powerhouse-only row. The fallback is deliberately not the credit order: a Pro subscriber who asks a reasoning question drops to the code model, which reasons well, rather than to the cheapest row on the ladder.

Auto is listed as available on every paid plan, and its lowest rung, the house model, is a Starter row. Free accounts are not model-gated on the chat path, so nothing blocks the Auto row there either: a free account can try any text model inside a lifetime allowance of 7 messages, which never resets. What a free account does not have is a credit balance. The credit limit on free is 0, so the 1 to 1 range published for free is what the unauthenticated picker endpoint ships, not a charge a free account ever incurs.

Routing to your own tier also removes one class of failure outright. Because Auto only ever picks from rows your plan can already open, a turn on Auto does not return "Upgrade your plan to use this model." That error is HTTP 403 with the code `tier_upgrade_required`, and it comes from choosing a model above your tier by hand. The refusals that do still apply on Auto are the ordinary allowance and burst refusals, and they read the same as on any model: see [rate limits](https://whizi.io/docs/rate-limits) and [model unavailable](https://whizi.io/docs/model-unavailable).

#### What a turn on Auto costs

Auto resolves to a concrete model before anything is charged, so the credit charge, the tier gate and the stored history all record the model that actually ran, not the word "auto".

The reason a range is published rather than a single figure is mechanical: the endpoint that serves the picker is unauthenticated and does not know who is asking, so it ships a per-tier map instead of one number.

A model's credit cost is a fixed integer per message. Nothing about message length, conversation length or answer length enters the calculation. Length changes what a turn costs only because it changes which rung Auto picks: a message over about 6,000 characters routes to the long-form rung at 4 credits, on the reasoning that a message that long is a pasted document rather than a question.

Because the amount varies per turn, Auto is the row where a whole-turn refusal is easiest to meet: a 20 credit reasoning turn needs 20 credits free, not whatever happens to be left. Running out mid work is covered in [out of credits](https://whizi.io/docs/out-of-credits).

Credit metering is live on the web client. A platform that is not on the credit allowlist sees a flat 1x instead and is charged one per turn, so no Auto range applies there.

The full credit scale, with how many models sit on each rung, is in [how credits work](https://whizi.io/docs/credits).

#### What is sticky inside a thread

**Difficulty only moves up.** The previous turn's model acts as a floor, so a follow-up to a reasoning answer cannot drop back to the cheapest model. A thread that has earned an expensive rung stays there for its follow-ups.

**Attachments always route the same way.** Any turn carrying an image or an upload goes to the Gemini Flash rung at 2 credits, on the reasoning that an image needs vision rather than a large context window. On Starter, which cannot reach that row, the same turn is answered on the house model at 1 credit.

**Length is the one hard threshold.** About 6,000 characters is where a message stops being treated as a question and starts being treated as a pasted document, which sends it to the long-form rung.

If you want a specific model rather than a rung, switching is a per-message decision. That is covered in [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### What Auto is not

**Auto is not configurable.** The ladder is six fixed rungs pinned to five model rows. Settings offers Subscription, Cancel, Reset password, Export data, Import chats, Default model, Dark mode, Language, and support and legal links, and nothing on that list edits the Auto ladder, weights a rung, or excludes a model from it.

**Auto is not the same thing as the router models in the catalogue.** The catalogue also carries provider router identifiers, and those are ordinary picker rows: they are charged a flat 20 credits by owner policy rather than by arithmetic, because a router picks a downstream model per request and has no fixed token price. A router that lands on the dearest Claude Opus row is still sold at 20. Those rows sit on Powerhouse, and Whizi Auto is not one of them.

**Auto does not reach every family.** Its five model rows are the house model, a Gemini Flash row, a mid GPT row, Claude Sonnet and a top GPT row. A DeepSeek, Grok, Qwen, Kimi, GLM, Llama or Mistral answer is always a hand-pick in the picker, at any tier.

The catalogue by plan is in [the model list](https://whizi.io/docs/models), and what each plan includes is in [plans and limits](https://whizi.io/docs/plans-and-limits).

Key points:

- The six rungs Auto routes to, with the model and the credit cost on each
- Which rungs each plan can actually reach, and why Pro tops out at 10 credits
- What a turn on Auto costs, and why Auto never returns an upgrade paywall

Checklist:

- Auto has six fixed rungs across five model rows, and no credit price of its own
- A turn costs whatever the model Auto picked costs: 1 on Starter, 1 to 10 on Pro, 1 to 20 on Powerhouse
- Auto never picks a model above your plan, so it never returns "Upgrade your plan to use this model."
- A message over about 6,000 characters routes to the long-form rung at 4 credits
- Any turn carrying an attachment routes to the Gemini Flash rung at 2 credits
- Difficulty is sticky: a follow-up cannot drop back below the previous turn's model
- Because the cost varies per turn, a 20 credit turn needs 20 credits free before it will send

FAQ:

Q: How do I see which model answered on Auto?
A: Reload the chat. Auto resolves to a concrete model before the turn is charged, and the stored history keeps that model, so the turn shows the row that actually answered rather than the word "auto".

Q: Is Auto cheaper than picking an expensive model myself?
A: Not at the top of the ladder. The reasoning rung and the top Claude Opus row both cost 20 credits, so choosing between them is a product call rather than a cost saving. Auto's ceiling is whatever rung your plan reaches: 10 credits on Pro, because the reasoning rung is Powerhouse-only, and 20 on Powerhouse.

Q: Can I change what Auto picks, or pin it to one model?
A: No. The ladder is fixed at six rungs, and no setting edits it, weights a rung, or excludes a model from it. If you want a specific model, pick it in the model picker for that message.

Q: Is Auto the same as the router models in the catalogue?
A: No. The catalogue's provider router identifiers are ordinary Powerhouse picker rows, charged a flat 20 credits by owner policy because a router picks a downstream model per request and has no fixed token price. Whizi Auto is not one of them: it has no rung of its own and charges you for the model it landed on.

Q: Does Auto work on the free tier?
A: A free account is not model-gated on the chat path, so the picker is not what stops it. The allowance is: 7 messages for the life of the account, which never reset, and a credit limit of 0. Whether Auto or a model you picked yourself answered makes no difference to that count.

### Image generation failed in Whizi: every error and its fix

URL: https://whizi.io/docs/image-generation-failed/
Updated: August 2026

Quick answer: When image generation fails in Whizi, the wording of the error names the cause: a plan gate, a spent allowance, a screener refusal, a rejected request, or a provider failure after the job started. Image generation is included on Pro and Powerhouse only. Provider failures refund the reservation, so send the prompt again.

The exact error strings Whizi returns when an image will not generate, the condition that triggers each one, and whether the attempt cost you anything.

#### The short answer

When image generation fails in Whizi, the wording of the error tells you which kind of failure it was: a plan gate, a spent allowance, a screener refusal, a rejected request, or a provider failure after the job had already started. Only the first two are fixed by changing plan. Provider failures are usually fixed by sending the same prompt again, and they refund what they reserved.

Image generation is included on Pro and Powerhouse. Free and Starter accounts have an image allowance of zero, so on those plans a refusal is the designed behavior rather than a fault.

| Error you see | What happened | What to do |
| --- | --- | --- |
| `Choose a subscription to generate images.` | A free account asked for an image (HTTP 402) | Subscribe to Pro or Powerhouse |
| `Upgrade to Pro or Powerhouse to generate images.` | A paid plan with no image allowance, which is Starter (HTTP 403) | Move up to Pro or Powerhouse |
| `Your monthly image generation limit has been reached.` | The period image counter is exhausted (HTTP 429) | Wait for the next period, or move up a tier |
| `Your weekly image generation limit has been reached.` | The same, on a weekly billing cycle (HTTP 429) | Wait for the next period, or move up a tier |
| `Image generation could not be started. Please try again.` | The provider refused the job at submit (HTTP 429 or 502) | Send it again. The reservation was refunded |
| `The image provider could not complete this generation.` | The provider accepted the job, then reported it failed | Send it again, or try a different image model. The reservation was refunded |
| `Image prompts must be between 1 and 4000 characters.` | The prompt was empty or over the cap (HTTP 400) | Shorten the prompt |

#### Plan and allowance: the two refusals that are not bugs

Image generation is metered on its own counter, separate from text messages. A standalone image job spends only the image counter, not the credit balance.

| Plan | Image generations per monthly period |
| --- | --- |
| Free | 0 |
| Starter | 0 |
| Pro | 100 |
| Powerhouse | 500 |

On a weekly billing period the figure is the monthly allowance divided by four, rounded up.

A free account gets `Choose a subscription to generate images.` as HTTP 402 with code `subscription_required`. A paid account whose tier carries no image allowance gets `Upgrade to Pro or Powerhouse to generate images.` as HTTP 403 with code `tier_upgrade_required`. Today that second case means Starter, because Starter and free are the two plans with an allowance of zero.

When the counter for the period runs out, the message is `Your monthly image generation limit has been reached.` on a monthly cycle and `Your weekly image generation limit has been reached.` on a weekly one, both HTTP 429 with code `image_limit_reached`. The counter is per billing period, so the options are to wait for the period to turn over or to move up a tier. The image counter is a separate meter from the credit balance.

#### The job started and then failed

These are provider failures. They are worth retrying, because the same prompt often succeeds on a second attempt, and because both of the common ones give your allowance back.

`Image generation could not be started. Please try again.` is returned when the submit call fails: HTTP 429 with code `image_provider_rate_limited` when the provider answers with a rate limit, and HTTP 502 with code `image_submit_failed` otherwise. The reservation is refunded before the error is returned.

`The image provider could not complete this generation.` is written into the job record when the provider accepts the job and later reports it as failed. The reservation is refunded at that point too.

`Image generation status is temporarily unavailable.` means the status poll failed rather than the generation: HTTP 429 when the upstream poll is rate limited and HTTP 502 otherwise, both with code `image_status_failed`. The job may still be running, so reload before assuming it died.

`Image job not found.` is HTTP 404 with code `image_job_not_found`, returned when the job id does not resolve.

Three rarer strings mean the image itself was produced but could not be retrieved or stored: `The completed image could not be saved.` when the finished image cannot be copied into storage, `Generated image could not be downloaded.` when the fetch of the provider image is not OK or declares an over-cap length, and `Generated image exceeded the size limit.` when the downloaded image is empty or over the byte cap. Two more sit earlier in that chain: `Image provider returned an invalid URL.` is raised when the provider result URL is missing or not on the allowed host list, and `Image result was unavailable.` when the result fetch is not successful. Retry is the only user-side action for all of them.

`Something went wrong. Please retry.` is the HTTP 500 body with code `internal_error`, used for any unhandled backend error, so it carries no diagnostic of its own. If it keeps coming back on the same prompt, a different prompt will not fix it, and support is reachable from Settings.

#### Errors that reject the request before generation

`Choose Whizi Image, Nano Banana, FLUX, or Stable Diffusion.` is HTTP 400 with code `invalid_image_model`, returned when the model on the request is not an image generation model. Those four names are the image models the catalogue offers.

`Image prompts must be between 1 and 4000 characters.` is HTTP 400 with code `invalid_image_prompt`. An empty prompt and a 5,000 character prompt fail the same check.

If you are generating inside a chat rather than as a standalone job, the strings come from the chat path instead and read differently, and every one of them reports that your credits were not charged. A chat image is retried on the model you picked and then re-run on a second and third image model before anything is reported to you, so these strings mean every provider failed, not that one did. `Every image provider is rate-limited right now.` is HTTP 429 with code `provider_rate_limited`. `No image provider would run this prompt.` is HTTP 400 with code `provider_rejected_prompt`, and rewording usually clears it. `This image took too long to generate.` is HTTP 504 with code `provider_timeout`. `We could not generate this image. Every provider we tried failed.` is HTTP 502 with code `provider_error`, the catch-all for anything else.

An in-chat image prompt can also be stopped by the ordinary chat submission errors, which are not image errors and are documented on their own page: a duplicate submission, a generation already running on the account, and an edited prompt whose original message is gone from the conversation. See [message not sending](https://whizi.io/docs/message-not-sending).

#### The prompt was refused, not failed

Content policy refusals return HTTP 400 rather than 403, deliberately, because the clients already render the error message on a 400. So a refusal reads as an ordinary error in the interface even though nothing broke.

Screening runs before any quota is reserved and before the provider is called, which means a refused prompt costs you no allowance at all.

The three refusals most likely to hit an image prompt:

- `Whizi does not create sexual or suggestive content involving minors. This request was not sent to any model. If you think this was a mistake, contact support from Settings.` (code `content_policy_minor_safety`)
- `Whizi does not create sexual content featuring real, identifiable people. If you think this was a mistake, contact support from Settings.` (code `content_policy_real_person`). This rule applies to generation requests only, not to text chat, which is why the same wording can pass in a conversation and fail on an image.
- `Whizi cannot help with building weapons, explosives, or other things designed to hurt people. If you think this was a mistake, contact support from Settings.` (code `content_policy_weapons`)

If the screener has been tripped repeatedly on generation prompts, the account earns a lockout and gets `Generation is paused on this account after repeated requests that Whizi cannot fulfil. It reopens automatically. Contact support from Settings if you think this is wrong.` as HTTP 429 with code `content_policy_locked`. The string carries its own remedy: it reopens automatically, and support is reachable from Settings.

#### An image that generated fine and will not load now

Generated images are stored on the server and served back through signed URLs, so the link in an old chat is not a permanent address. Uploaded attachments and generated media are configured to expire after up to 30 days.

`Media has expired.` is HTTP 404 with code `media_expired`, returned when the stored object is past its retention window.

`Media not found.` is HTTP 404 with code `media_not_found`, returned when the media id does not resolve for your account at all.

Downloading is the only durable copy. A chat is not storage.

Key points:

- Every image generation error string Whizi can return, with the condition behind it
- Which failures refund your allowance automatically and which do not
- The plans that carry an image allowance at all, and what happens on the ones that do not

Checklist:

- Image generation is on Pro and Powerhouse only, at 100 and 500 per monthly period
- Free and Starter have an image allowance of zero, so refusals there are expected
- Weekly billing gets the monthly allowance divided by four, rounded up
- A standalone image job spends only the image counter, not your credit balance
- Submit failures and provider failures refund the reservation automatically
- Content policy refusals are screened before any quota is reserved, so they cost nothing
- A content policy lockout reopens automatically, and support is reachable from Settings
- Generated media is configured to expire after up to 30 days, so download what you want to keep

FAQ:

Q: What does "Choose a subscription to generate images." mean?
A: It means the account is on the free tier, which has an image allowance of zero. The error is HTTP 402 with code subscription_required. The paid equivalent is "Upgrade to Pro or Powerhouse to generate images.", HTTP 403, which appears on a paid plan that carries no image allowance. Image generation is included on Pro at 100 per monthly period and Powerhouse at 500.

Q: I am on Starter and every image request is refused. Is that a bug?
A: No. Starter has an image allowance of zero, the same as free, so the refusal is the designed behavior and no amount of retrying will change it. Starter reaches four ids for text chat, including Auto, which is a router rather than a model. Image generation is not part of the tier, and Pro is the lowest plan that includes it.

Q: Does a failed image generation still cost me a generation?
A: Usually not. An image job reserves against the image counter before the provider is contacted, and that reservation comes back on both of the provider failures named above. A screener refusal never reaches the reservation step at all, so there is nothing to give back in the first place.

Q: Why was my image prompt refused when the same wording works in a chat?
A: One rule is deliberately generation-only. The refusal about sexual content featuring real, identifiable people applies to generation requests and not to text chat, so a phrase that passes in a conversation can be refused as an image prompt. Every content policy refusal returns HTTP 400, which is why one arrives looking like an ordinary error rather than a policy notice.

Q: My generated image will not load any more. Where did it go?
A: Two different 404s produce that symptom, and the code tells them apart. "Media has expired." (code media_expired) means the stored object is past its retention window; generated media is configured to expire after up to 30 days. "Media not found." (code media_not_found) means the id does not resolve for your account at all. Neither state can be undone from the interface.

### How to import your ChatGPT chat history into Whizi

URL: https://whizi.io/docs/import-chats-from-chatgpt/
Updated: August 2026

Quick answer: Yes, Whizi imports ChatGPT conversation history. Export your data from ChatGPT, open Settings, choose Import Chats under Account, and pick the conversations.json file out of the export. Your browser parses that file and posts only normalized conversations, so the export itself is never uploaded. Running the same import twice lands the same chats once.

Bring your ChatGPT conversations across in one pass: which file Whizi reads, where the Import Chats control lives, what carries over, and why a large export needs no splitting.

#### The short answer

Yes, Whizi imports ChatGPT conversation history. Export your data from ChatGPT, then open Settings, choose Import Chats under Account, and pick the `conversations.json` file out of the export. Whizi accepts exactly two import sources, ChatGPT and Claude, and the same dialog handles both.

The export file itself is never uploaded: your browser reads `conversations.json` or the raw export `.zip`, parses it, and posts only plain normalized conversations. The import is also safe to run more than once, because every imported conversation is keyed by account plus source plus source id, so importing the same export twice or clicking Import twice lands the same chats once.

Import failures surface as a small set of fixed messages, each listed with its trigger and its fix on [chat import errors](https://whizi.io/docs/chat-import-errors). Two of them are catch-alls: one for a file the browser could not read, one for an upload that did not finish.

| What you need | Where it comes from |
| --- | --- |
| The `conversations.json` file | Inside the data export ChatGPT produces for you |
| The Import Chats control | Settings, under Account, described as a one-time setup action |
| A signed-in Whizi account | The import writes chats into the account you are signed in as |

#### Picking the right file out of the export

The file Whizi reads is `conversations.json`. Everything else in a ChatGPT export archive is ignored, so the fastest route is to unzip the export and pick that file directly.

The `.zip` works too, and for a big history it is the better choice. A raw `.json` export is accepted up to 300 MB, past which you are told to use the `.zip` instead. The `.zip` path has no size ceiling of its own, because only the `conversations.json` entry is decompressed and the images in the archive never have to fit in memory.

You do not have to tell Whizi which product the export came from. The parser identifies the format itself, which is also how it knows to apply the Claude rules to a Claude export and the ChatGPT rules to this one.

#### What carries over

An import brings the transcript across, not the scaffolding around it. That is deliberate, and it is why a successful import can still look shorter than the original.

Two rules are specific to a ChatGPT export. Only the branch that was on screen is replayed, so regenerated replies and discarded drafts do not come across as messages. And non-transcript blocks are skipped: hidden reasoning, analysis tool code and output, browsing scratchpads, custom instruction blocks, and function calls.

The rest of the drop rules, and what they do to titles, long messages and odd timestamps, are listed on [chat import errors](https://whizi.io/docs/chat-import-errors).

#### A large export slices itself

The import endpoint takes conversations in batches rather than in one lump, and the website importer slices your export to fit, so a big history needs no splitting by hand and no setting changed. The per-request numbers, and the headroom the website leaves under them, are on [chat import errors](https://whizi.io/docs/chat-import-errors).

Whizi publishes no cap on the total number of conversations an account may import, and no per-day import cap. Alongside the 20 imports per minute route limit, the account-wide ceiling of 30 requests per minute and the per-IP ceiling of 60 per minute apply to every route, import included.

#### After the import

Imported conversations land in your chat list like any other Whizi chat, so a thread that started in ChatGPT can be carried on with [GPT in Whizi](https://whizi.io/docs/using-gpt-in-whizi) or with any other model your plan reaches. The sidebar search box finds them by title, and it matches the chat title only rather than message bodies.

On privacy, the import path is easy to reason about, because the export file never reaches a server in the first place. Once the conversations are stored they are treated like any other chat, and what Whizi does and does not do with them, training and retention included, is set out in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

Key points:

- The export file never leaves your browser, only the parsed conversations are posted
- A large export is sliced and rate limited for you, with nothing to configure
- Running the same import twice lands the same chats once

Checklist:

- Export your data from ChatGPT and keep the archive
- Whizi reads conversations.json, nothing else in the export
- A raw .json is accepted up to 300 MB, past that use the .zip
- Import Chats sits in Settings under Account
- The export file is parsed in your browser and never uploaded
- Only the branch that was on screen comes across, not regenerated replies
- A large export is sliced into batches for you, with nothing to configure
- A twenty second pause on a large export is rate limiting, not a fault
- Running the same import again never duplicates chats

Steps:

1. Export your data from ChatGPT: Request your data export from ChatGPT and download the archive it produces.
2. Find conversations.json: That is the only file Whizi reads. A raw .json file is accepted up to 300 MB; past that, pick the export .zip instead.
3. Open Settings and choose Import Chats: Import Chats sits in Settings under Account, alongside Export Data.
4. Pick the file: Your browser parses the export in place and identifies which vendor it came from. The file itself is never uploaded.
5. Leave it running: A large export is split into batches automatically, and a pause of about twenty seconds is the rate limiter rather than a fault.

FAQ:

Q: Does importing keep my Whizi chats in sync with ChatGPT?
A: No. It is a one-time copy rather than a running sync: an import brings across the conversations that exist in the export at the moment you ran it, so anything you carry on in ChatGPT afterwards only arrives if you export and import again.

Q: Can I import from Claude too?
A: Yes. ChatGPT and Claude are exactly the two accepted import sources, they run through the same pipeline and the same caps, and the same dialog parses either vendor export. Claude exports drop thinking blocks, tool calls and attachments, and import the text blocks. The Claude side has its own page: [import chats from Claude](https://whizi.io/docs/import-chats-from-claude).

Q: What happens to imported chats if I cancel my subscription?
A: They stay. Cancelling stops billing and ends access at the end of the paid term, while the account and its contents remain. Conversations and messages are kept until you delete them or delete the account, and removing the data is the separate account-deletion action.

### How to import your Claude chat history into Whizi

URL: https://whizi.io/docs/import-chats-from-claude/
Updated: August 2026

Quick answer: Whizi imports a Claude export from Settings, under Account, where the item is called Import Chats. Pick conversations.json out of your Claude export, or the export .zip it sits in. Your browser parses the file and posts only the transcripts. Thinking blocks, tool calls and attachments are dropped. Running the same import twice lands the same chats once.

Bring a Claude export into Whizi from Settings, and know exactly what carries over: text transcripts yes, thinking blocks, tool calls and attachments no.

#### The short answer

Whizi imports a Claude export from Settings, under Account, where the item is called Import Chats. Pick `conversations.json` out of your Claude export, or pick the export `.zip` and Whizi decompresses only that one entry from it. Your browser parses the file and posts the conversations, which land as ordinary chats in your list. Claude is one of exactly two accepted import sources, alongside ChatGPT.

What arrives is the text of the conversation. For a Claude export, thinking blocks, tool calls and attachments are dropped, and only text blocks are imported. That is a deliberate rule, not a failure, so an imported chat can legitimately be shorter than the original looked.

| Question | Answer |
| --- | --- |
| Which file do I pick | `conversations.json`, or the export `.zip` it sits in |
| Does my export get uploaded | No. It is parsed in the browser, and only normalized conversations are posted |
| What carries over | User and assistant text, conversation titles, original timestamps |
| What does not | Thinking blocks, tool calls, attachments |
| How big can it be | A raw `.json` is accepted up to 300 MB. The `.zip` path has no ceiling of its own |
| What if I run it twice | The same chats land once. Re-running is the intended recovery |

#### What a Claude export turns into

Each conversation in the export becomes one chat. The rules Whizi applies while it converts them are worth knowing, because every one of them can make an imported chat differ from what you remember.

| Rule | What you see |
| --- | --- |
| Only text blocks are imported | Thinking blocks, tool calls and attachments from the Claude export do not appear |
| Titles are trimmed to 120 characters | A long title arrives cut, not rejected |
| A conversation with no usable title becomes `Imported chat` | Several such chats share that name |
| A message over the content cap is truncated | One oversized message does not cost you the rest of that conversation |
| A timestamp before 2015-01-01, or in the future, is discarded | That row falls back to the time you ran the import |
| A conversation with no readable text is counted as skipped | It is reported to you rather than silently dropped |

The last row is the one people misread. A skipped count is information about your export file, not a sign that something went wrong in Whizi. It counts conversations that hold no readable text at all.

#### How a large export is handled

You do not have to split a big export yourself. The importer slices it for you, against caps that are the same for a Claude export and a ChatGPT one.

| Cap | Value |
| --- | --- |
| Conversations per request | 25 |
| Messages per conversation per request | 400, with the remainder sent as append slices onto the chat the first slice created |
| Request body accepted by the import endpoint | 2,000,000 bytes |
| Request budget the website actually uses | 1,200,000 bytes, deliberately under the cap to leave room for the request envelope |
| Import requests per minute | 20 |

The rate limit is the part that looks like a bug and is not. An import that hits it is not surfaced as an error: the client waits 20 seconds and retries, up to 4 attempts per batch. So a big Claude export that seems to stall for twenty seconds at a time, then moves again, is behaving exactly as designed. Leave it running. The rest of the rate limiting picture is in [rate limits](https://whizi.io/docs/rate-limits).

#### Running the import twice is safe

Every imported conversation is keyed by account plus source plus source id. Re-importing the same export, resuming after a dropped connection, and clicking Import twice all land the same chats once.

Two consequences follow from that key. The first is that the source is part of it, so a Claude import and a ChatGPT import on the same account are tracked separately and cannot collide with each other. The second is that when an export gives no stable conversation id, the key falls back to a hash of the transcript itself, so those conversations are still matched on a re-run rather than duplicated.

There is one behavior to watch for. Deleting an imported chat and then importing the same export again brings that chat back rather than reporting a duplicate. If you deleted an imported conversation on purpose, do not re-run that export unless you want it returned.

#### Your export file does not leave your browser

The export file itself is never uploaded. Your browser reads it, parses it, and posts only the normalized conversations. That is also why every format error appears instantly, before anything has been sent.

Once the conversations are in, they are treated like any other Whizi chat. Conversations and messages are kept until you delete them or delete the account. And on training: "Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models. Whizi does not sell that content as training data." The full picture is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

#### If the import is rejected

Almost every failed Claude import is the wrong file rather than a broken import. Two messages cover the common cases:

`That file is not a ChatGPT or Claude export. Look for conversations.json inside the export zip.` means the parser could not identify either export format in what you picked.

`There is no conversations.json inside that zip. Unzip it and pick the file directly.` means the zip you picked has no matching entry.

Every other import message, the condition behind it, and the fix are on [chat import errors](https://whizi.io/docs/chat-import-errors), which is the page to open if the wording you saw is not one of these two.

Key points:

- Claude is one of exactly two import sources Whizi accepts, alongside ChatGPT
- The export file never reaches the server: your browser parses it and posts only the transcripts
- Running the same import twice lands the same chats once, so a stopped import is fixed by re-running it

Checklist:

- The file to pick is conversations.json, or the export .zip holding it
- Import Chats sits in Settings under Account
- A raw .json export is accepted up to 300 MB, past that use the .zip
- Thinking blocks, tool calls and attachments are dropped from a Claude export
- A conversation with no usable title arrives as Imported chat
- A pause of twenty seconds at a time on a large export is the rate limiter, not a fault
- Re-running the same import lands the same chats once
- Deleting an imported chat and re-importing brings it back

Steps:

1. Get your Claude export: Download your Claude data export. The file Whizi reads is conversations.json, either on its own or inside the export .zip.
2. Open Import Chats in Settings: Import Chats sits in Settings under Account, grouped there as a one time setup action.
3. Pick the file: Pick conversations.json, or pick the export .zip and Whizi decompresses only that one entry out of it.
4. Let the browser parse it: The export is read and identified in your browser. Nothing is uploaded at this point.
5. Run the import: Conversations are posted in batches of at most 25. A large export pauses while the rate limiter catches up, then carries on by itself.

FAQ:

Q: Can I import my Claude chat history into Whizi?
A: Yes. Claude is one of exactly two accepted import sources, alongside ChatGPT, and both run through the same pipeline and the same caps. Import Chats sits in Settings under Account. You pick conversations.json out of your Claude export, or the export .zip it sits inside, and your browser parses it before anything is posted.

Q: What does not carry over from a Claude export?
A: Thinking blocks, tool calls and attachments. Only text blocks are imported. Beyond that, titles are trimmed to 120 characters, a conversation with no usable title becomes "Imported chat", a message over the content cap is truncated rather than rejected, and a timestamp dated before 2015-01-01 or in the future is discarded so the row falls back to the time you ran the import.

Q: Does my Claude export get uploaded to Whizi?
A: No. The export file never reaches the server. Your browser reads it, parses it, and posts only the normalized conversations, which is why a file that is the wrong format is rejected instantly rather than after an upload. Conversations and messages are then kept until you delete them or delete the account.

Q: My import seems stuck. Is it broken?
A: Watch whether it moves again rather than how long it has been sitting. Importing conversations is limited to 20 requests per minute, and an import that hits that limit is not surfaced as an error: the client waits 20 seconds and retries, up to 4 attempts per batch. Repeated pauses followed by progress are the rate limiter doing its job on a large export.

Q: Will importing the same export twice duplicate my chats?
A: No. Every imported conversation is keyed by account plus source plus source id, so a second run lands the same chats once and picks up only what is missing. When an export gives no stable conversation id, the key falls back to a hash of the transcript itself. The one thing to know is that deleting an imported chat and importing again brings it back.

Q: Where do the Claude models themselves live in Whizi?
A: Importing your history and using the models are separate things. The Claude family in the catalogue, with the credit cost and plan gate for each model, is on [using Claude in Whizi](https://whizi.io/docs/using-claude-in-whizi).

### Whizi invoices and receipts: where the record of your payment lives

URL: https://whizi.io/docs/invoices-and-receipts/
Updated: August 2026

Quick answer: Whizi does not issue invoices or receipts, and there is no billing history screen in the product. For a subscription bought on the web, the Terms of Service say payment is processed by Stripe and that Stripe holds the transaction information. A purchase made inside the mobile app sits with Apple or Google.

Whizi has no invoice or billing history screen. Where the transaction record for a web subscription actually sits, what the Terms say about Stripe and about store billing, and what a Whizi charge should match.

#### The short answer

Whizi does not issue invoices or receipts, and there is no billing history screen in the product. For a subscription bought on the web, the Terms of Service say that payment is processed by Stripe and that Stripe holds the transaction information. Where a subscription was bought inside the mobile app, the Terms say Apple or Google controls the payment and the store terms apply.

What Whizi itself holds is smaller than people expect: limited customer, subscription, transaction, and status information, and never a full payment card number.

| Where the payment was made | What the Terms say about it |
| --- | --- |
| On whizi.io | Website subscriptions may be processed by Stripe, which may retain transaction information under its terms and legal obligations |
| Inside the mobile app | Subscriptions purchased in the mobile App are processed by the Apple App Store or Google Play, and Apple or Google controls payment processing, renewal, cancellation, and store refunds |

If you need a document for expenses or for accounting, Whizi has no button that produces one. The routes that do exist are below.

Paying for several people is a separate question. There is no self-serve multi seat billing, and an invoice covering several seats is a manual arrangement rather than a product route: [team accounts](https://whizi.io/docs/team-accounts) has the detail.

#### What Whizi itself can show you

The Settings screen offers Subscription, Cancel, Reset password, Export data, Import chats, Default model, Dark mode, Language, and support and legal links. There is no invoices item, no billing history list, and no receipt download in it.

The closest thing you can pull yourself is the account data export. Export Data sits in Settings under Account and returns your account data as one downloadable JSON file, described in the code as the access counterpart to account deletion.

What that file covers is enumerated in the code: chats and messages, memory, projects and their file metadata, media, share links, subscription, and referral profile. Read the list precisely. The **subscription** entry is the plan state on your account, and invoices, charge amounts, card details and payment history are not among the listed contents.

#### What the Terms say about who holds the record

Three passages settle this, and they are worth quoting rather than paraphrasing.

On web subscriptions, the Terms of Service say: "Website subscriptions may be processed by Stripe. Stripe processes payment credentials and may retain transaction information under its terms and legal obligations. Whizi may receive limited customer, subscription, transaction, and status information but does not store full payment-card numbers."

On subscriptions the Terms describe as purchased in the mobile app: "Subscriptions purchased in the mobile App are processed by the Apple App Store or Google Play. RevenueCat verifies receipts and entitlement status so Whizi can provide and restore the purchased access. Apple or Google controls payment processing, renewal, cancellation, and store refunds under the applicable store terms. Whizi does not receive or store your full mobile payment-card number."

Receipts are verified in that flow so that Whizi can grant and restore access, which is an entitlement check rather than a document handed to you. It does not make Whizi the holder of a store receipt.

On the processors generally, the Privacy Policy says: "Apple, Google, RevenueCat, Stripe, and other payment processors may retain transaction, receipt, tax, fraud, chargeback, or dispute records for periods required by law or their compliance obligations."

One more line covers how records reach you at all. The Terms say you consent to "ELECTRONIC DELIVERY OF NOTICES, POLICIES, AND RECORDS OF TRANSACTIONS INITIATED OR COMPLETED BY US OR VIA THE SERVICES." Records of transactions are electronic by agreement, and no self-serve invoice document is produced anywhere in the Whizi product surface.

#### What a Whizi charge should match

To reconcile a line on a card statement, compare it against the plan you selected. These are the USD amounts the pricing page shows. Each tier also carries a weekly product identifier, offered on mobile only, and no weekly price is published anywhere, so a weekly line is not covered by this table.

| What it covers | Amount shown |
| --- | --- |
| The first 7 days after signup | $0.99 |
| Starter, monthly | $15.99/month |
| Starter, annual | $10.99/mo billed annually at $131.88 |
| Pro, monthly | $29.99/month |
| Pro, annual | $19.99/mo billed annually at $239.88 |
| Powerhouse, monthly | $49.99/month |
| Powerhouse, annual | $34.99/mo billed annually at $419.88 |

Read the annual figures carefully. The total shown for an annual plan is the displayed per-month rate multiplied by twelve, kept as a display string rather than a figure read back from Stripe. Treat it as the expected value for a yearly charge, not as a guarantee of the exact number on the statement.

The trial is not free. The pricing page footnote states that the first 7 days are charged at the intro rate, $0.99 in USD, after which the plan you selected renews unless you cancel.

An account with no subscription is not charged anything. A free account gets a lifetime allowance of 7 messages, which does not reset, and then the paywall. The short form of these price and plan questions sits in the [Whizi FAQ](https://whizi.io/docs/whizi-faq).

The Terms describe the cycle plainly: "The length of your billing cycle is monthly, or annually," and your subscription "will continue and automatically renew unless canceled."

#### Tax and currency on the charge

Two passages in the Terms bear on what the number looks like, and both should be read before you assume a charge is wrong.

On tax and currency: "Sales tax will be added to the price of purchases as deemed required by us. We may change prices at any time. All payments shall be in US dollars."

On variation by market: "Prices, billing periods, introductory offers, renewal terms, and applicable taxes are shown before purchase and may vary by country, currency, store, and plan."

What the site itself renders today is the USD table above: every visitor is shown, and charged, those USD amounts. Whizi does not publish a per country tax breakdown, so if you need one, that is a question for the processor holding your transaction rather than something documented here.

The accepted payment methods listed in the Terms are Visa, Mastercard, and American Express.

#### When the charge and the account disagree

If the charge is on your statement but the account still behaves as unpaid, that is a different symptom with its own page: read [you paid but the plan features are still locked](https://whizi.io/docs/subscription-not-active), which also carries the two Settings strings people mistake for billing failures.

If the charge never completed at all, the checkout errors are catalogued in [card declined or payment failed](https://whizi.io/docs/payment-declined).

If you believe the amount itself is wrong, the Terms give one address: "If you have any questions or are unsatisfied with our Services, please email us at whizihelp@gmail.com." What that route does and does not promise is set out in [refunds](https://whizi.io/docs/refunds).

#### What deleting your account does to the record

Deleting your Whizi account does not remove the transaction record, and for this page that changes nothing: the retained record is the processor's, not a document Whizi will hand you. The Terms make deletion "subject to temporary provider caches, protected recovery history, separately stored safety reports, and limited records retained for legal, tax, payment, fraud-prevention, security, or dispute-resolution obligations."

The full retention picture, including what a retained payment record explicitly does not keep alive, is quoted from the Privacy Policy in [refunds](https://whizi.io/docs/refunds).

One practical point. Deletion and cancellation are separate actions: the Terms say "Account deletion and subscription cancellation are separate," and that Apple or Google billing is cancelled through the applicable store. The sequence is in [billing, trials and how to cancel](https://whizi.io/docs/billing-and-cancellation).

Key points:

- Whizi has no invoice, receipt, or billing history screen anywhere in Settings
- Who holds the transaction record, quoted from the Terms, for web and for store purchases
- The USD amounts the pricing page shows, and how to treat an annual total when you compare it

Checklist:

- Whizi has no invoice, receipt, or billing history screen
- Web payments are processed by Stripe, which holds the transaction information
- The Terms put a purchase made in the mobile app with Apple or Google, not with Whizi
- Whizi never stores a full payment card number, on web or on mobile
- The account data export lists a subscription record, and no invoice or charge amount
- An annual total is a displayed figure rather than a Stripe readback, so compare it as an expected value
- Multi seat invoicing is a manual arrangement, not a self-serve route
- Payment records can be retained after account deletion for tax and legal obligations

FAQ:

Q: Where do I download a Whizi invoice?
A: You cannot, because Whizi has no invoice download. Settings offers Subscription, Cancel, Reset password, Export data, Import chats, Default model, Dark mode, Language, and support and legal links, and nothing that produces a billing document. For a web subscription the Terms say payment "may be processed by Stripe" and that Stripe "may retain transaction information under its terms and legal obligations," so the transaction record sits with the processor rather than in your Whizi account.

Q: Does Whizi email me a receipt?
A: Whizi does not document a receipt email of its own. The record sits with whoever processed the payment: Stripe for a website subscription, or the store for a subscription the Terms describe as purchased in the mobile app. If you were expecting something in your inbox and nothing arrived, look there rather than in Whizi Settings.

Q: I have a store receipt. Why does Whizi still show me as free?
A: A store receipt is proof that the store charged you, not the record Whizi reads. The Terms say RevenueCat "verifies receipts and entitlement status so Whizi can provide and restore the purchased access," so what unlocks the plan is the entitlement rather than the receipt sitting in your email. If the entitlement is not on the account you are signed into, that is the case covered by [you paid but the plan features are still locked](https://whizi.io/docs/subscription-not-active).

Q: How much should the charge on my statement be?
A: These are the USD amounts the pricing page shows. The first 7 days after signup are charged at $0.99, then the plan you selected renews. Monthly is $15.99/month for Starter, $29.99/month for Pro and $49.99/month for Powerhouse. Annual is charged once for the year: Starter at $10.99/mo billed annually at $131.88, Pro at $19.99/mo billed annually at $239.88, and Powerhouse at $34.99/mo billed annually at $419.88. Those annual totals are the per-month rate multiplied by twelve, held as a display string rather than read back from Stripe, so treat them as expected values. The Terms add that sales tax "will be added to the price of purchases as deemed required by us." Weekly products exist on mobile only and carry no published price, so a weekly line is not covered here.

Q: If I delete my account, do my payment records go too?
A: No. Deletion is subject to "limited records retained for legal, tax, payment, fraud-prevention, security, or dispute-resolution obligations," so a payment record can outlive the account. That still does not produce a document for you: what is retained is the processor's record, not an invoice Whizi issues. Cancel any store subscription separately, since the Terms say deletion and cancellation are different actions.

### Cannot log in to Whizi, or the session keeps expiring

URL: https://whizi.io/docs/login-and-account-access/
Updated: August 2026

Quick answer: Whizi access failures fall into four groups: the sign-in attempt failed, the browser could not get a fresh session token, another session is already open, or the account sharing guard refused the request. Every authentication failure from the backend carries the code unauthorized with HTTP 401, and signing out and back in clears the session cases.

Whizi sign-in, sign-up, session and account access error strings quoted verbatim, the exact condition that triggers each one, and the fix.

#### The short answer

Whizi access problems fall into four groups: the sign-in attempt itself failed, the browser could not get a fresh session token, another session is already open, or the account sharing guard refused the request. The message tells you which one, and each message below is quoted from the code that produces it.

The five below cover sign-in, session, single-session and sharing refusals:

| Message | What triggered it | Fix |
| --- | --- | --- |
| `Login failed. Please check your credentials.` | The sign-in attempt returned a status other than complete | Re-enter the email and password, or reset the password |
| `Invalid or expired session token.` | HTTP 401. Token verification failed for any reason: expired token, bad signature, wrong issuer | Sign out and sign in again |
| `Authentication required.` | HTTP 401. The request carried no Bearer token in the Authorization header | Sign in again |
| `Please sign out from other sessions and try again.` | The sign-in error text contained "single session mode" | Sign out of Whizi in the other browser or tab, then retry |
| `Account sharing limit reached. Please use your own Whizi account.` | HTTP 403. More than 3 distinct device or session keys were seen on the account in the last 24 hours | Use your own account. The window is rolling, so otherwise the only option is to wait |

Every authentication failure from the backend carries code `unauthorized` with HTTP 401. On the chat page and in the share modal the backend message is shown verbatim, so the string in a network trace and the string on screen are the same string there.

#### What "session expired" actually means

It means the page asked for a fresh session token and did not get one. That single condition surfaces under a different sentence depending on where you were when it happened.

| Where you see it | Message |
| --- | --- |
| Any backend call with no token | `Authentication required.` |
| Any backend call whose token fails verification | `Invalid or expired session token.` |
| A token that verifies but is missing a claim | `The Clerk user_jwt token must include sub and email claims.` |
| Chat list, chat load, delete, rename or upload | `Failed to retrieve session token.` |
| Side by side compare, at send time | `Your session expired. Please refresh and try again.` |
| Importing a ChatGPT or Claude export | `Your session ended. Sign in again, then start the import.` |
| Settings, before a cancellation | `Authentication failed. Please sign in again.` |
| The upgrade modal, before checkout | `Authentication failed. Please sign in again.` |
| The share modal | `Not signed in.` |
| The voice overlay | `Sign in to Whizi to use voice mode.` |
| The session call itself | `Failed to fetch session ({status}).` |

Sign out and sign in again. Two of the strings say so themselves: `Your session ended. Sign in again, then start the import.` and `Authentication failed. Please sign in again.`

#### Sign-in and password reset errors

Authentication and account management are provided by Clerk, so these strings come from the sign-in flow rather than from a Whizi password store.

| Message | Trigger condition |
| --- | --- |
| `Login failed. Please check your credentials.` | The sign-in attempt returned a status other than complete |
| `An error occurred during login.` | The sign-in call threw and no readable message came back |
| `Please sign out from other sessions and try again.` | The error text contained "single session mode" |
| `Failed to authenticate with social login.` | The Google redirect handoff threw before leaving the page |
| `Please enter your email address.` | The password reset form was submitted with an empty email field |
| `Failed to send reset email. Please try again.` | The password reset request threw |
| `Current password is required.` | The current password field was empty on the reset password page |
| `Passwords do not match.` | The two new password fields differ |
| `Failed to reset password. Please try again.` | The reset failed and returned no readable message |

`Please sign out from other sessions and try again.` is the only one on that list whose fix is somewhere else: close the other tab or browser where Whizi is signed in, then retry here.

The Settings page has its own copy of the reset flow, with the same two strings: `Please enter your email address.` when the field is empty and `Failed to send reset email. Please try again.` when sending fails. Retrying is the whole fix. The flow itself is walked through step by step in [reset your Whizi password](https://whizi.io/docs/reset-your-password). If it keeps failing, contact support rather than creating a second account.

Retrying in a tight loop can also hit the rate limiter, which answers HTTP 429 with `Too many requests. Please wait and try again.` The windows behind that string are in [Whizi rate limits](https://whizi.io/docs/rate-limits).

#### Sign-up and email verification errors

| Message | Trigger condition |
| --- | --- |
| `Registration service is not available.` | The sign-up object was not ready when the form was submitted |
| `Registration failed. Please check your information.` | The sign-up call threw with no readable message |
| `Verification service is not available.` | The email code step ran with no sign-up object |
| `Verification failed. Please try again.` | The verification attempt returned a status other than complete |
| `Invalid verification code. Please check your email.` | The verification call threw |
| `Could not start Google sign-up. Please try again.` | The Google sign-up redirect threw |
| `Missing session token after social signup.` | No token was available after a social signup completed |
| `Your account was created, but checkout could not start. Please sign in and choose your plan again.` | A registration carrying a pending checkout could not get a token afterwards |

The last one is not a failed signup: the account exists, and only the pending checkout was lost. What to do next is in [card declined or payment failed at checkout](https://whizi.io/docs/payment-declined).

The two "not available" strings mean the Clerk sign-up object was not ready when the form was submitted, rather than that your details were rejected. Wait for the page to finish loading, then submit again.

#### Temporary email addresses are refused

`Temporary email addresses are not allowed. Please use a permanent email address.` is HTTP 403 with code `disposable_email_blocked`. It is raised at account creation when the email domain is on the disposable list.

Accounts that already exist are unaffected by the list. The fix is a permanent address, and there is no override to request.

The same string can also come back at checkout with a different status. That side of it is covered in [card declined or payment failed at checkout](https://whizi.io/docs/payment-declined).

#### Account sharing limit reached

`Account sharing limit reached. Please use your own Whizi account.` is HTTP 403 with code `account_sharing_limit`. It fires when more than 3 distinct device or session keys have been seen on one account in the last 24 hours.

Ordinary multi-device use is expected and supported: an account can be signed in on more than one device, and both reach the same backend and the same conversation store. Only one generation runs per account at a time, enforced by a per-user lock, so two devices can never generate at once on one account.

There is no setting for this and no self-serve reset. The window covers the last 24 hours, so it clears on its own as older device keys age out. If you are one person on more than three browsers or devices in a day and you keep hitting it, contact support rather than trying to work around it.

#### Signed in, but Whizi says the account is not there

`User not found.` is raised when the users row behind a session has been deleted. The checkout route returns the same string as HTTP 404 `not_found` when it cannot read the caller row.

Account deletion is a two-step confirmation in Settings, and it removes the live Whizi account record, synchronized chats and messages, uploaded and generated media, saved mobile memory, referral records, mobile entitlement and usage state, rate-limit state, push tokens, and the Clerk authentication account, subject to the limited retention set out in the privacy policy.

If you see this and did not delete anything, [contact support](https://whizi.io/docs/contact-whizi-support). There is no self-serve repair for it. What deletion covers in full is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

#### Guest profiles and linking

Guest use is signed-out use. Creating an account is optional in the mobile app, and continuing without one creates an anonymous, device-bound guest profile that needs no name or email address. It has its own set of access errors.

| Message | Trigger condition |
| --- | --- |
| `A valid guest session is required.` | HTTP 401. The guest token is absent, malformed, or names a guest profile that does not exist |
| `This guest profile has already been linked. Sign in to continue.` | HTTP 409. The guest token resolves to a profile already claimed by a different account |
| `This guest profile is linked to a different Whizi account.` | HTTP 409. The claim targets a guest profile already bound elsewhere |
| `Wait for the current reply to finish, then try creating your account again.` | HTTP 409. The profile was claimed while a generation was still running |

Only the last one has an easy fix: wait for the reply to finish, then create the account. The other three mean the guest profile is already bound to an account, so sign in to that account, or contact support if you do not recognize it.

Guest data is not stranded by this. Data export works for guest sessions too, because guests hold real chats and memory that are worth getting out.

#### When the account itself will not do something

These are account-level failures shown in Settings rather than sign-in failures.

| Message | What it is |
| --- | --- |
| `Your account could not be deleted. Please try again, or contact support.` | The deletion request failed, shown in Settings |
| `Account deletion failed ({status}).` | The raw error behind that toast, with the HTTP status |
| `Failed to export your data. Please try again.` | The data export failed, shown in Settings |
| `Export request failed ({status}).` | The raw error behind that toast |

Data export is rate limited to 5 requests per minute, because an export reads everything the account owns. Retrying faster than five times a minute will be refused by that limit.

There is no public or developer API, no API key issuance, and no personal access tokens. Every route authenticates as an end user session, so there is no key to rotate, revoke or paste when access breaks. If a script or third-party tool cannot reach your account, that is why.

Key points:

- Sign-in, sign-up and session error strings quoted verbatim, with the HTTP code where the backend sets one
- What "session expired" actually means, and which string each surface shows
- The account sharing guard: more than 3 device or session keys in 24 hours

Checklist:

- A session error means the request carried no usable token: sign out and sign in again
- Invalid or expired session token is HTTP 401 with code unauthorized
- A single session message means the other Whizi tab has to sign out first
- The account sharing guard trips above 3 device or session keys in 24 hours, and clears on its own
- Disposable email domains are refused at signup as HTTP 403 disposable_email_blocked
- User not found means the account record behind the session is gone: contact support
- A 429 means a rate limit window refused the request, so wait rather than retrying in a loop
- There are no API keys or personal access tokens to rotate

FAQ:

Q: Will resetting my password clear `Invalid or expired session token.`?
A: No. That string is returned when JWT verification against the Clerk JWKS throws, so it is about the token the browser is carrying rather than the password on the account. The password reset flow has its own separate strings, including `Current password is required.` and `Passwords do not match.` Sign out and sign in again instead.

Q: Is `Failed to retrieve session token.` a Whizi outage?
A: It is not a backend refusal. The website throws it itself whenever `getToken` returns nothing before a chat list, chat load, delete, rename or upload call, so it carries no HTTP status and no error code. The backend equivalents, which do carry a status, are `Authentication required.` and `Invalid or expired session token.`, both HTTP 401 `unauthorized`.

Q: What does `Failed to fetch session ({status}).` mean, and what is the number in it?
A: It is the raw error thrown by the website when the session call itself fails, with the HTTP status of that call filled in where `{status}` sits. It is not a message written for readers, which is why it looks like an internal string. Quote it with the number intact when you contact support.

Q: I cannot get in at all and none of these strings match. What now?
A: Contact support and paste the exact text you saw, along with what you were doing at the time. Each of these strings maps to one place in the code, so the verbatim wording is the fastest thing support has to work with. There is no self-serve account unlock.

### A Whizi message will not send, or the reply never arrives

URL: https://whizi.io/docs/message-not-sending/
Updated: August 2026

Quick answer: A Whizi message that will not send is one of five things: another reply is still generating, you sent the same text twice, you crossed a per-minute or per-hour rate limit, the message failed validation, or your session token expired. Whizi names the reason in the error text, and a rate-limited request costs nothing.

A send is refused by the per-user generation lock, by the duplicate-message check, or by a rate limit that is unwound so the refused request costs nothing. Here is the string for each case, with its HTTP code.

#### The short answer

A Whizi message that will not send is one of five things: another reply is still generating, you sent the same text twice, you crossed a per-minute or per-hour rate limit, the message itself failed validation, or your session token expired. Whizi names the reason in the error text, so the string you see decides the fix.

| What you see | Code | Why | What to do |
| --- | --- | --- | --- |
| `Generation already in progress. Please stop it or wait for it to finish.` | 409 | Another reply is running on your account | Stop it, or wait |
| `Duplicate message ignored. Please wait for the current response to finish.` | 409 | The same text (and image) was already sent to this chat recently | Wait for the reply already running |
| `Too many requests. Please wait and try again.` | 429 | A per-minute window refused the request | Wait, then resend |
| `Hourly message limit reached. Please try again later.` | 429 | The hourly chat ceiling refused it | Wait, then resend |
| `A non-empty message is required.` | 400 | The message is missing or whitespace only | Type something |
| `This message is {count} characters, over the 100,000 character limit. Attach a shorter file, or ask about one section at a time.` | 400 | One message went over 100,000 characters | Split it, or ask about one section |
| `The model is temporarily rate-limited. Please retry shortly.` | 429 | The model provider answered the connect with a 429 | Retry, or switch models |
| `Failed to send message.` | none, a website toast | The send failed and the backend supplied no message | Check your connection and resend |
| `Something went wrong. Please retry.` | 500 | Any unhandled error on the backend | Resend, then report it if it repeats |
| `Invalid or expired session token.` | 401 | Token verification failed | Refresh the page and sign in again |

Rate-limit refusals are unwound, so a refused request never spends a message off your allowance.

#### Nothing happens when you press Enter

If Enter does nothing at all, a reply is still generating. The composer deliberately ignores Enter while a reply is in flight, so this is not a dropped keypress. Shift and Enter still inserts a newline.

Whizi runs one generation per account at a time, enforced by a per-user lock, so two devices can never generate at once for the same account. A second attempt is refused with `Generation already in progress. Please stop it or wait for it to finish.`, HTTP 409, code `generation_in_progress`. The same string is returned when the generation coordinator reports a conflict at connect time.

Sending the identical text twice returns `Duplicate message ignored. Please wait for the current response to finish.`, HTTP 409, code `duplicate_message`. This fires when the same message text (and image) was already sent to this conversation recently, which is what a double click or an impatient second Enter produces.

The fix for both is the same: stop the running reply, or let it finish. Stopping is handled on the server. The coordinator cancels the provider read, saves whatever text arrived, releases the lock, and refunds an undelivered turn. If a lock is somehow left behind, it recovers on its own: a lock becomes stealable once its holder is judged orphaned or stalled, and expired locks are swept on a schedule.

If a stop leaves a gap, the assistant bubble reads `Generation stopped.`, which is the text persisted when a generation is stopped and no partial text was captured. `Generation stopped, but it could not be saved.` means the stop worked but persisting the partial reply did not.

#### The message is refused before any model sees it

These are HTTP 400 validation refusals raised before anything reaches a model. Nothing was sent upstream and nothing was charged.

| Message | Code | Trigger |
| --- | --- | --- |
| `A non-empty message is required.` | `empty_message` | `content` is missing or whitespace only |
| `A model is required.` | `invalid_model` | `model` is missing or blank |
| `That model is not available for text chat.` | `unsupported_model` | The model id does not resolve to a text model in the catalog |
| `This message is {count} characters, over the 100,000 character limit. Attach a shorter file, or ask about one section at a time.` | `message_too_long` | A single message crossed 100,000 characters |
| `That system prompt is missing or over the 32,000 character limit.` | `invalid_system_prompt` | The system prompt is missing or over 32,000 characters |
| `That image reference is not valid.` | `invalid_image_url` | The image reference is not a string, or is longer than 2,048 characters |
| `This conversation has {count} messages, over the 100 message limit.` | `too_many_messages` | The app chat route refused a transcript carrying more than 100 messages in one request |

The 100,000 character cap counts the whole message, including any text extracted from an attachment. For what Whizi does when extracted attachment text does not fit the message budget, see [file upload failed](https://whizi.io/docs/file-upload-failed).

#### Rate limits, and one refusal that is not about speed

Both 429 rate-limit strings are in the table above, and both carry the code `rate_limit_reached`. A refused request is unwound, so a 429 never spends a message off your allowance: wait out the window and resend. The window figures and the plan allowances behind them are on [rate limits](https://whizi.io/docs/rate-limits) and [out of credits](https://whizi.io/docs/out-of-credits).

One refusal in the same family is not about speed at all: `Account sharing limit reached. Please use your own Whizi account.`, HTTP 403, code `account_sharing_limit`, fires when more than 3 distinct device or session keys have been seen on one account in the last 24 hours. Waiting a minute does not clear it.

#### The reply starts, then stops or never renders

Once the connection is open, a failure comes either from the model provider or from the machinery carrying the stream. The strings differ by where the failure landed.

| Message | Where it comes from |
| --- | --- |
| `The model is temporarily rate-limited. Please retry shortly.` | HTTP 429 `provider_rate_limited`, when the provider answers the connect with a 429 |
| `The model provider rejected the request.` | Used when the provider error body cannot be parsed, or parses to an empty message |
| `The model provider returned an empty stream.` | HTTP 502, upstream answered OK but sent no body |
| `The model provider returned an error.` | In-stream error event when the provider sends an error frame with no message of its own |
| `The model stream was interrupted.` | In-stream error event for any other mid-stream failure |
| `Generation failed mid-stream.` | The generation ended in failure after bytes had already been sent |
| `Generation failed to start.` | The internal fallback text on a coordinator start failure that carries no message of its own |
| `The generation stream is unavailable.` | Raised when the coordinator cannot hand back a stream to attach to |
| `Generation aborted before it started.` | Raised as an AbortError when the client disconnects before the upstream connect completes |
| `Something went wrong. Please retry.` | HTTP 500 `internal_error`, the body for any unhandled backend error |
| `Sorry, there was an error generating the response.` | Written into the assistant bubble when a stream ends with no final message and no accumulated content |

Any upstream failure other than a 429 is HTTP 502 `provider_error` carrying the provider message, truncated to 300 characters. On the mid-stream path there is no in-band error event: the connection simply breaks.

One of those strings is a record rather than a message. `Generation aborted before it started.` is raised as an AbortError when the client disconnects before the upstream connect completes, and the route answers that case with HTTP 408 and an empty body, so the string is what the server writes down, not what the browser prints. In a network trace, an empty 408 on the generate route means the request went away before any model was reached.

The catch-all sits underneath all of this: anything the backend does not recognize ends as HTTP 500 `internal_error` with the body `Something went wrong. Please retry.` HTTP 500 does not identify a cause. Resend once, then contact support if it repeats.

What to do: resend, and if it repeats, switch to a different model in the same conversation. These errors originate upstream of Whizi, not in your account or your settings.

One thing you do not have to do is keep the tab open. Generation runs in a per-conversation Durable Object, so closing the tab mid-stream does not kill the answer and the reply is still persisted. A reopened tab or a second device can reattach to a generation already in flight. `The generation stream is unavailable.` is raised when the coordinator cannot hand back a stream to attach to. Reload the conversation and read whatever was persisted instead of waiting on the stream.

In the side-by-side compare view the per-side strings are different again: `The model did not respond. Please try again.` when a generate response is not OK and carries no readable error, `Could not start a comparison chat.` when the backing chat cannot be created, and `Something went wrong.` as the last resort. That last one is the short form, with no retry sentence after it, which is how you tell it apart from the `Something went wrong. Please retry.` body of a 500.

#### Your session expired

Auth failures look like send failures because they surface at send time. `Authentication required.` is HTTP 401 `unauthorized` and fires when the request carries no bearer token. `Invalid or expired session token.` is the same status and code when token verification fails for any reason: expired token, bad signature, wrong issuer.

On the website side you may instead see `Failed to retrieve session token.`, which the site throws whenever the token fetch returns nothing before a chat list, chat load, delete, rename, or upload call. The compare view has its own wording: `Your session expired. Please refresh and try again.`

The fix is a page refresh, then signing in again if the refresh does not restore the session. Nothing on the account is lost, and no message was sent.

#### It is not a failure, it is a limit

An allowance refusal and a tier refusal are not send failures. The request was formed correctly and Whizi declined it on purpose, so there is nothing to fix in the composer: allowance refusals are covered in [out of credits](https://whizi.io/docs/out-of-credits), and a model that shows as locked is covered in [troubleshooting](https://whizi.io/docs/troubleshooting).

A prompt refused by content screening is also not a send failure. Screening runs before any quota is reserved and before the provider is called, so a refused prompt costs no allowance, and the refusal arrives as HTTP 400 with the reason shown in the chat.

#### Editing a sent message fails instead

Editing a prompt and resending is how you re-run a turn, since Whizi has no regenerate control. That path has its own two failures.

`That message is no longer part of this conversation. Reload and try again.` is HTTP 404 `replace_anchor_not_found`, returned when the edit names a message id that is no longer in the conversation. Reload the conversation. The client is holding a message id the server no longer has.

`Conversation not found.` is HTTP 404 `not_found`, returned when the conversation id does not exist or is not owned by the caller. It shows on the generate route and on the read, delete, rename, message-append, version-switch and share routes too.

Key points:

- The exact error string for each way a send can fail, quoted verbatim
- Which failures cost you an allowance and which cost nothing
- The two limits that refuse a send: one generation at a time, and 10 messages a minute

Checklist:

- Enter doing nothing means a reply is still generating, not a dropped keypress
- One generation runs per account at a time, across all your devices
- A duplicate send is ignored on purpose, with a 409 and a message saying so
- Chat has a per-minute and a per-hour cap that no plan lifts
- A rate-limited request costs nothing off your allowance
- One message is capped at 100,000 characters
- Provider errors are fixed by resending or switching model, not by settings
- Closing the tab mid-reply does not kill the reply
- Auth errors are fixed by refreshing and signing in again

FAQ:

Q: What does "Generation already in progress. Please stop it or wait for it to finish." mean?
A: Stop the running reply, or wait for it to finish, then send again. The refusal is HTTP 409 with code generation_in_progress, and it also appears when a different device is generating on the same account. Stopping is handled on the server: the partial answer is saved, the lock is released, and an undelivered turn is refunded.

Q: How many messages can I send per minute?
A: Ten per minute and 60 per hour, per user, flat across every plan including Powerhouse. There is also a shared ceiling of 30 requests per minute across all routes and a per-IP ceiling of 60 per minute. Crossing them returns HTTP 429 with either "Too many requests. Please wait and try again." or "Hourly message limit reached. Please try again later." Nothing lifts these limits, but a refused request is unwound and never spends a message off your allowance.

Q: My reply stopped halfway through. Did I get charged?
A: A mid-stream failure breaks the connection with no in-band error event, and whatever text arrived before the break is kept. If you stop a generation yourself, the server cancels the provider read, saves the partial answer, and refunds an undelivered turn. For a repeated provider failure, switch to a different model in the same conversation, since the error originates upstream of Whizi. If what you get back is HTTP 500 with "Something went wrong. Please retry.", that is the generic body for an unhandled backend error rather than a model problem, so resend once and report it if it keeps happening.

Q: Does closing the tab cancel my reply?
A: No, with one narrow window at the very start. If the client disconnects before the upstream connect completes, the server records "Generation aborted before it started." and answers HTTP 408 with an empty body, which is why that request looks like it returned nothing at all. Past that point the reply survives the client, because generation runs in a per-conversation Durable Object and the answer is still persisted. "The generation stream is unavailable." is raised when the coordinator cannot hand back a stream to attach to, and reloading the conversation shows whatever was saved.

Q: Whizi says my message is over the 100,000 character limit. What now?
A: A single message is capped at 100,000 characters and the refusal is HTTP 400, code message_too_long. The count covers the whole message, including text extracted from an attachment. Ask about one section at a time, or attach a shorter file. When extracted attachment text does not fit the message budget, Whizi truncates instead of refusing and warns you that only the first part was sent.

### A model is unavailable in Whizi: greyed out, missing, or erroring

URL: https://whizi.io/docs/model-unavailable/
Updated: August 2026

Quick answer: A model is unavailable in Whizi for one of three reasons: your plan does not include it, the model id did not resolve to a usable model in the catalogue, or the provider behind it failed on that request. A 402 or 403 means the plan, a 400 means the model id, and a 502 means the provider.

A model refuses for one of three reasons: your plan, the model id, or the provider behind it. This page splits the model error messages Whizi returns by cause, and gives the tier ladder and the picker sections that decide which models you see.

#### The short answer

A model is unavailable in Whizi for one of three reasons: your plan does not include it, the model id did not resolve to a usable model in the catalogue, or the provider behind that model failed on that request. The error text tells you which of the three you hit.

| Cause | What you see | What fixes it |
| --- | --- | --- |
| Plan gate | A 403 or a 402 | Upgrade, or pick a model your tier includes |
| Model id | A 400 saying the model is not available or not supported | Pick a model from the picker rather than reusing an old id |
| Provider failure | A 429 or a 502 mentioning the model provider | Retry, or send the same message on another model |

Every backend error arrives as JSON in the shape `{"error": {"code": "...", "message": "..."}}`, and the website shows the backend message verbatim, so the string on your screen is the string in the table below.

#### Plan gate errors, word for word

These mean the account is real and the model is real, but the tier does not reach it. Access is resolved on the server by model identifier, so this is an entitlement rather than an interface bug.

| Message | Status and code | Trigger |
| --- | --- | --- |
| `Upgrade your plan to use this model.` | 403 `tier_upgrade_required` | A paid account picked a text model its tier does not include |
| `This model is only available on the Powerhouse plan.` | 403 `powerhouse_required` | A creative model on the Powerhouse-only list was requested from a lower tier |
| `Choose a subscription to use Whizi.` | 402 `subscription_required` | The route requires a paid plan and the account is not subscribed |
| `Your free messages are used up. Start a subscription to keep chatting.` | 402 `free_limit_reached` | A free account is already at its lifetime message cap |

The free tier lifetime allowance is 7 messages, counted for the life of the account with no reset path. Once it is spent, every send returns the 402 in the last row above.

#### Which plan opens which models

Tier rank runs free 0, Starter 1, Pro 2, Powerhouse 3, and a higher tier always satisfies a lower requirement, so each tier includes everything below it.

| Plan | What it opens |
| --- | --- |
| Free | A lifetime allowance of 7 messages with no reset. After that, `Your free messages are used up. Start a subscription to keep chatting.` |
| Starter, $15.99/month | Exactly four picker entries: Auto, Whizi AI, one OpenAI GPT model, and one fast Google Gemini model |
| Pro, $29.99/month | Starter's four plus 37 further model identifiers, so 41 listed identifiers |
| Powerhouse, $49.99/month | Everything else in the 280+ catalogue |

The house model Whizi AI sits at Starter tier, so the default assistant is never paywalled above Starter. The tier fallback is Powerhouse rather than Pro: any identifier the catalogue gains lands on the top tier until it is deliberately promoted into the Pro list. A model that appeared this week and sits on Powerhouse is usually not a mistake.

Pro is a curated set rather than everything cheap: one or two current flagships per family, each family's fast tier, and the high-volume workhorses. Claude Sonnet is the Pro ceiling and the Claude Opus class is Powerhouse. Agent personas are a Pro feature gated by identifier ahead of the Powerhouse fallback, so they do not inherit the top-tier default. Which rows a plan reaches is listed family by family, including [Grok](https://whizi.io/docs/using-grok-in-whizi), [DeepSeek](https://whizi.io/docs/using-deepseek-in-whizi) and [Qwen](https://whizi.io/docs/using-qwen-in-whizi).

The reliable way to avoid this class of error entirely is Auto. Auto is available on every paid plan, it is a router rather than a model, and it only ever picks from models the subscriber can already reach, never above the tier.

#### The model is missing from the picker

Check which section you are looking at before assuming it is gone. The picker has six: Recommended, Agents, Image Generation, Video Generation, Audio Generation, and All models. Recommended is deliberately constrained to models reachable at Pro or below, so the top of the picker is never Powerhouse-locked. A model that needs Powerhouse will not be in Recommended.

If you see image models where you expected text models, the composer is in creative mode, which switches the picker to image models and swaps the placeholder to an image prompt. Media generators are not gated per model at all: they skip the per-model gate and are tiered by picker section instead.

A model that was there before still answers. A model never leaves the catalogue once it has been offered, because a conversation carries the model it was created with. The identifier keeps resolving even when the model behind it has been replaced: a chat pinned to a retired model is silently answered by its replacement rather than refused, so an old thread still opens. The picker contents are served by the Whizi backend at `GET /api/models`, and the lists hardcoded into the web and mobile clients are offline fallbacks only, so what you see is the live catalogue rather than a list baked into your browser.

#### Model id errors, word for word

These fire before any provider is called. They mean the id in the request did not resolve to a model that can do the thing being asked.

| Message | Status and code | Trigger |
| --- | --- | --- |
| `A model is required.` | 400 `invalid_model` | The `model` field is missing or blank |
| `That model is not available for text chat.` | 400 `unsupported_model` | The requested model id does not resolve to a text model in the catalog |
| `A title and supported text model are required.` | 400 `invalid_conversation` | A conversation was created with a missing title, or a model that is neither a text nor a creative model |
| `Choose Whizi Image, Nano Banana, FLUX, or Stable Diffusion.` | 400 `invalid_image_model` | The requested model is not an image generation model |
| `Choose Whizi Music AI to generate a song.` | 400 `invalid_music_model` | A music job named something other than the music model |
| `Choose Whizi Video AI to generate a video.` | 400 `invalid_video_model` | A video job named something other than the video model |

`That model is not available for text chat.` is the one people hit most, and the usual cause is asking a generation model for prose. The Image Generation section is Whizi Image, Nano Banana, Flux and Stable Diffusion, Video Generation is one row, and Audio Generation is one row. None of them answer text turns. Pick a text model from Recommended or All models and resend.

#### Provider failures, where retrying is the fix

These are not your account and not your plan. The model exists, your tier reaches it, and the upstream provider failed on that request.

| Message | Status and code | Trigger |
| --- | --- | --- |
| `The model is temporarily rate-limited. Please retry shortly.` | 429 `provider_rate_limited` | The upstream provider answered the connect with a 429 |
| `The model provider rejected the request.` | 502 `provider_error` | The provider's error body could not be parsed, or parsed to an empty message |
| `The model provider returned an empty stream.` | 502 | The upstream response was OK but carried no body |
| `The model provider returned an error.` | In-stream error event | The provider sent an error frame mid-stream with no message of its own |
| `The model stream was interrupted.` | In-stream error event | Any other mid-stream failure |
| `We could not generate this image.` (or video, or song) | 502 `provider_error` | Every provider tried for a creative generation failed |
| `Every image provider is rate-limited right now.` | 429 `provider_rate_limited` | Every attempt was rate-limited |
| `No image provider would run this prompt.` | 400 `provider_rejected_prompt` | Every provider refused the prompt itself. Reword it |
| `This image took too long to generate.` | 504 `provider_timeout` | The wait budget ran out before any provider finished |

No setting on your side clears any of these, and Whizi publishes no retry window. Resend in a moment, or send the same message on a different model in the same conversation. Switching model does not lose the thread.

The four creative rows work differently from the text rows above them. A failed image, video or song is retried on the model you chose and then re-run on the other models of the same type before you are told anything, so seeing one of those messages means every provider failed rather than one. Your credits are returned in all four cases, which is why each message says so. Because a fallback can answer, the model named on a finished image is the one that actually produced it, which is not always the one you picked. Every string on that path, and whether it returns your allowance, is catalogued in [image generation failed](https://whizi.io/docs/image-generation-failed).

Two more strings belong to the same family. `Failed to send message.` is the website's fallback toast when a send fails and the backend supplied no message at all, which usually means a dropped connection rather than a model problem. `Sorry, there was an error generating the response.` is written into the assistant bubble when a stream ends with no final message and no accumulated content. In the side-by-side compare view, a failed side reads `The model did not respond. Please try again.`

#### Errors that look like model errors but are not

Three strings that name no model get read as the model being broken: `Your monthly message limit has been reached.` and `Your weekly message limit has been reached.` are HTTP 429 `message_limit_reached`, `Too many requests. Please wait and try again.` and `Hourly message limit reached. Please try again later.` are HTTP 429 `rate_limit_reached`, and `Generation already in progress. Please stop it or wait for it to finish.` is HTTP 409 `generation_in_progress`. None of them is a model gate, and switching model fixes none of them.

The allowance side is covered in [out of credits](https://whizi.io/docs/out-of-credits), and the rate limit and concurrency side in [a message will not send](https://whizi.io/docs/message-not-sending).

Key points:

- The three causes pulled apart: a plan gate, an unresolvable model id, and a provider failure
- Which plan opens which part of the picker, and why new models land on the top tier
- The failures where the only fix is to retry or switch models

Checklist:

- A 402 or 403 means the plan, a 400 means the model id, a 502 means the provider, and a 429 means either the provider or your own allowance depending on the code
- The free lifetime allowance is 7 messages, and it never resets
- Starter opens exactly four picker entries, Pro opens 41 identifiers
- Anything new in the catalogue defaults to Powerhouse until it is promoted
- Auto never picks a model above your tier, so it never triggers a plan gate
- Recommended is limited to models reachable at Pro or below
- A model never leaves the catalogue once it has been offered
- For provider failures the only fixes are retry or switch models

FAQ:

Q: Why is a model greyed out in Whizi?
A: Because your tier does not reach it. Sending anyway returns `Upgrade your plan to use this model.` as HTTP 403 `tier_upgrade_required`, or `This model is only available on the Powerhouse plan.` as HTTP 403 `powerhouse_required` for a creative model on the Powerhouse-only list. Upgrade, or pick a model your tier includes.

Q: What does "That model is not available for text chat." mean?
A: Almost always one of two things you did without noticing. Creative mode was left on, which points the picker at image models, so the row you picked was an image row. Or the chat was started on a row from Image Generation, Video Generation or Audio Generation and later used for prose. Turn creative mode off, choose a text model, and resend.

Q: How long do I wait after "The model is temporarily rate-limited. Please retry shortly."?
A: Whizi does not publish a wait time, so any number would be a guess. That string is HTTP 429 `provider_rate_limited` and it means the upstream provider refused the connect, not that your account did anything wrong. Resend in a moment, or send the same message on a different model in the same conversation, which keeps the thread and its context intact.

Q: A model disappeared from my model picker. Was it removed?
A: Almost certainly not. Start with which of the six sections you are in, because the picker splits into Recommended, Agents, Image Generation, Video Generation, Audio Generation, and All models, and Recommended is limited to models reachable at Pro or below. If the rows in front of you are image models and the composer placeholder is asking for an image prompt, creative mode is on and the picker has switched with it. Turn it off, then open All models.

Q: How many messages do I get before Whizi asks me to subscribe?
A: Seven. The free lifetime allowance is 7 messages, counted for the life of the account with no reset path, so it is not a monthly free tier. After that, sends return `Your free messages are used up. Start a subscription to keep chatting.` as HTTP 402 `free_limit_reached`, and no amount of waiting clears it. Starter at $15.99/month opens four picker entries, including Auto and the house model Whizi AI.

Q: A model I could use last month is locked now. What changed?
A: Model access is resolved per request by model identifier, through the same lookup every time: the Starter set, then the agent set, then the Pro set, and otherwise Powerhouse. So the answer can move for two reasons. Your tier is one. The other is the catalogue: any identifier it gains lands on Powerhouse until it is deliberately promoted into the Pro list, so a row that looks familiar but is new to the catalogue sits on the top tier by default. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

### The Whizi models list: which AI models are included on each plan

URL: https://whizi.io/docs/models/
Updated: August 2026

Quick answer: Whizi includes more than 280 AI models from OpenAI, Anthropic, Google, xAI, DeepSeek, Meta, Mistral, Moonshot, Qwen, Z.ai and others under a single subscription. Starter covers three 1 credit models, Pro adds Claude Sonnet 5, GPT-5.6 Terra and roughly 37 more, and Powerhouse opens the full catalogue.

The full model catalogue by plan: what Starter, Pro and Powerhouse each unlock, what every model costs in credits, and how access is enforced.

#### The short answer

Whizi includes more than 280 AI models from OpenAI, Anthropic, Google, xAI, DeepSeek, Meta, Mistral, Moonshot, Qwen, Z.ai and others under a single subscription. Access is tiered across three plans.

| Plan | Price per month | Text models | Monthly credits |
| --- | --- | --- | --- |
| Starter | $15.99, or $10.99 billed yearly | Whizi AI, GPT-5.6 Luna, Gemini 3 Flash | 400 |
| Pro | $29.99, or $19.99 billed yearly | Everything in Starter plus Claude Sonnet 5, GPT-5.6 Terra, Grok 4.6, Kimi K3 and roughly 37 more | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | The full catalogue, including Claude Opus 5 and GPT-5.6 Sol | 8,000 |

Every plan covers web, iOS and Android from one login. New accounts start with a 7 day trial for $0.99.

#### What Starter includes

Starter is deliberately narrow: three text models chosen because they are fast, cheap to run, and cost 1 credit per message, so a 400 credit allowance goes a long way.

| Model | Provider | Credits per message |
| --- | --- | --- |
| Whizi AI | Whizi house assistant | 1 |
| GPT-5.6 Luna | OpenAI | 1 |
| Gemini 3 Flash | Google | 1 |

Whizi AI is the default assistant. It runs on GPT-5.6 Luna with a Whizi system prompt, and the raw model is offered separately in the full model list for anyone who would rather have no persona layer on top.

Starter also covers image, PDF and Word document upload, which is the feature most people actually came for. What it does not include is image generation, side-by-side comparison, or any frontier model.

#### What Pro adds

Pro is a curated tier rather than "everything cheap". The rule is one or two current flagships per model family, each fast tier, and the high volume workhorses. It works out to roughly 40 text models.

**OpenAI.** GPT-5.6 Terra, GPT-5.2, GPT-5.5, GPT-5 Mini, GPT-4.1, GPT-4.1 Mini, GPT-4o, GPT-4o Mini, GPT-OSS-120B.

**Anthropic.** Claude Sonnet 5, Claude Sonnet 4.6, Claude Sonnet 4.5, Claude Haiku 4.5. Sonnet is the Pro ceiling, and the Opus class is Powerhouse.

**Google.** Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash, Gemini 3.1 Pro Preview, Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 2.5 Flash Lite.

**Everyone else.** Grok 4.6 from xAI, DeepSeek V3.2, DeepSeek Chat V3.1, DeepSeek R1, Llama 4 Maverick, Llama 4 Scout, Llama 3.3 70B, Mistral Large 2512, Mistral Medium 3.1, Mistral Small 3.2, Mistral Nemo, Qwen3.7 Plus, Qwen3 235B, Kimi K3, Kimi K2.5, GLM-5, and Sonar Pro.

Pro also unlocks the agent personas, which are a feature rather than a model tier: the Canadian CPA agent, the study mentor, the writing coach and the math tutor all run on the same base model, and what you are buying is the prompt, the token budget and the reasoning depth. Pro adds 100 image generations per month on top.

#### What Powerhouse unlocks

Powerhouse opens the remainder of the catalogue. Anything not named in the Starter or Pro lists above resolves to Powerhouse, which is roughly 240 further models. The categories held back on purpose:

- **The Opus class.** Claude Opus 5, and the 4.x Opus generation. 20 credits per message.
- **GPT-5.6 Sol.** Held back deliberately as the trophy next to Claude Opus, at 20 credits per message. Terra takes the recommended row on Pro, because the top of the picker should never be locked.
- **The frontier reasoning tier.** Grok 4.5, Qwen3 Max Thinking, GLM-5.1, Kimi K2 Thinking, Intellect-3.
- **Coding specialists.** GPT-5.2 Codex, Kimi K2.7 Code, and the Qwen3 Coder and Devstral families.
- **The 400B class open weights, and the routing models.**

Powerhouse is also where the non-text generation lives: side-by-side model chats, AI music generation, 500 image generations, 500 minutes of AI voice, and 100 AI video generations per month.

#### What a model costs in credits

Every plan spends from a single monthly credit allowance, and the cost of a message depends on which model answered. Most of the catalogue sits at 1 or 2 credits, and the scale runs up to 80 for the oldest Opus generation. The multiplier tracks what the model actually costs to run, so the cheap workhorses stay effectively unlimited and the frontier models are metered.

| Model | Credits per message |
| --- | --- |
| Whizi AI, GPT-5.6 Luna, Gemini 3 Flash, DeepSeek V3.2, Qwen3.7 Plus | 1 |
| GLM-5 | 2 |
| GPT-5.6 Terra, Claude Haiku 4.5 | 4 |
| Grok 4.6 | 6 |
| Claude Sonnet 5, GPT-5.2, Gemini 3.1 Pro, Kimi K3 | 10 |
| Claude Opus 5, GPT-5.6 Sol, Sonar Pro | 20 |

The practical consequence: 2,000 Pro credits is 200 Claude Sonnet 5 messages, or 500 GPT-5.6 Terra messages, or 2,000 messages on the 1 credit tier, or any mix of the three. 8,000 Powerhouse credits is 400 Claude Opus 5 messages.

Image, video and music generation are not credit charged. They draw on their own monthly counts, listed per plan above. The full per model breakdown is in the [credits reference](https://whizi.io/docs/credits).

For a single model, the [model reference](https://whizi.io/models) carries a page each: the list rate per million tokens, the cost of one standard answer, the context window, the credit cost, and the plan that unlocks it. For example [Claude Opus 5](https://whizi.io/models/claude-opus-5), [Claude Sonnet 5](https://whizi.io/models/claude-sonnet-5) and [GPT-5.6 Terra](https://whizi.io/models/gpt-5-6-terra).

#### How access is enforced

Model access is decided on the server, not in the interface. The worker ships each model minimum tier on the models endpoint, and the same function backs the gate on the chat route, so a locked model cannot be reached by editing a request.

The fallback for an unrecognised model id is Powerhouse rather than Pro. That matters when the catalogue gains a model: a new arrival is Powerhouse-only until it is explicitly promoted, which is the safe direction. It also means the Starter and Pro lists above are exhaustive rather than indicative.

The web app and the mobile apps read the same catalogue, so a model available on the iPhone app is available on the web on the same plan.

Key points:

- The exact model list for each of the three plans, not a marketing summary
- What every model costs in credits, and why the number varies from 1 to 20
- Which models are deliberately held back for Powerhouse, and what that buys

Checklist:

- Starter covers three 1 credit models plus file upload
- Pro is the tier that adds Claude Sonnet and the GPT-5.6 Terra class
- Claude Opus and GPT-5.6 Sol require Powerhouse
- Credits are one shared allowance, and most models spend 1 or 2 per message
- Image, video and music generation are counted separately, not in credits
- Anything not named in the Starter or Pro list requires Powerhouse

FAQ:

Q: Does Whizi include both ChatGPT and Claude in one subscription?
A: Yes. Pro and Powerhouse both include OpenAI and Anthropic models in the same workspace, so you can send a prompt to GPT and then switch the same conversation to Claude without losing context. Pro reaches Claude Sonnet 5 and GPT-5.6 Terra, and Powerhouse reaches Claude Opus 5 and GPT-5.6 Sol. Starter includes OpenAI and Google models but no Anthropic model.

Q: How many AI models does Whizi have?
A: The catalogue is over 280 text models plus the image, video, music and voice generators. Three are on Starter, roughly 40 on Pro, and the remainder on Powerhouse. The number moves as providers add and retire models, and the model picker inside the app is always the live list.

Q: Is Claude Opus available on the Pro plan?
A: No. Sonnet is the Pro ceiling for Anthropic models, and the entire Opus class requires Powerhouse. This is a deliberate split rather than a technical limit: Powerhouse needs a frontier model to sell, and Opus at 20 credits per message costs several times what Sonnet costs to run.

Q: What happens when I run out of credits?
A: Sending stops until the allowance resets at the start of your next billing month, and the app offers an upgrade at that point. Because roughly half the catalogue costs 1 credit while the frontier tier costs 20, the cheapest way to stretch an allowance is to route routine work to the 1 credit models and save the expensive ones for the messages that actually need them.

Q: Do the mobile apps have the same models as the web app?
A: Yes. The web app, the iOS app and the Android app read the same server-side catalogue and the same tier gate, so a model available on one is available on all three at the same plan level. Your subscription and your credit allowance are shared across every device on one login.

### How to organize conversations in Whizi: rename, delete, search and projects

URL: https://whizi.io/docs/organize-conversations/
Updated: August 2026

Quick answer: Organizing a Whizi chat history comes down to four actions: rename a chat, delete one chat or several at once, search the chat list from the sidebar search box, and group chats into projects. The sidebar search matches the chat title only, so renaming is how you index your own history.

What keeps a long chat history usable: renaming, deleting one chat or several, the sidebar search box, and projects as folders with their own files and instructions.

#### The short answer

Organizing a Whizi chat history comes down to four actions: rename a chat, delete one chat or several at once, search the chat list from the sidebar search box, and group chats into projects, which are folders of conversations that also carry pinned files and per-project instructions.

The one detail that changes how you should use all four: the sidebar search matches the chat title only. Message bodies are not searched. So a chat you never renamed is findable only by scrolling, and renaming is effectively how you index your own history.

| Tool | Where it lives | What it actually does |
| --- | --- | --- |
| Rename | The chat list | Sets the title, 1 to 120 characters. The title is also the only thing search matches |
| Delete | The chat list | Removes one chat, or several in one action |
| Search | A search box over the chat list in the sidebar | Matches titles across projects and the ungrouped list at once |
| Projects | The sidebar, listed above the chat list | A folder of chats, plus up to 10 pinned files and instructions applied to every chat in it |

#### Renaming a chat, and the two different length limits

A conversation title has to be between 1 and 120 characters. The backend refuses anything outside that with `Conversation titles must be between 1 and 120 characters.`, HTTP 400, code `invalid_title`.

The website stops you earlier than the backend does. It refuses a blank name with the toast `Chat name cannot be empty.` and a long one with `Chat name is too long (max 100 characters).` So the practical ceiling when you rename from the website is 100 characters, while the field itself accepts 120.

If the rename call itself fails, the toast is `Failed to rename chat. Please try again.`

Because search matches titles, a good title is a search query you are writing for your future self. Names that work are the ones carrying the noun you will actually type: a client name, an invoice number, a file name, a project code. Names that do not work are the ones the first message generates for you, which describe the opening question rather than what the thread became.

#### What the search box does and does not do

The sidebar has a search box over the chat list. It matches the chat title, and it cuts across projects and the ungrouped list together, so a chat is findable by name whichever folder it sits in.

Message bodies are not searched. There is no full-text search over what was said inside a conversation, so a phrase you remember from an answer will not find the chat that contains it. This is the single most useful thing to know before you build a filing habit, because it means the work belongs in the title rather than in the folder structure.

There is also no keyboard shortcut to jump to it. Whizi binds Enter to send and Shift plus Enter to a newline, and Escape closes an open dialog, and beyond those two conventions there is no shortcut system and no shortcuts panel.

The habits that make a title-only search actually find things are set out in [search your chat history](https://whizi.io/docs/search-your-chat-history).

#### Projects: folders that carry files and instructions

A project is a folder of conversations that also holds pinned files and its own custom instructions. Projects are listed in the sidebar above the chat list, and they are available to any signed-in account including the free tier. The per-project caps are abuse guards rather than a paid gate.

A chat belongs to at most one project. Anything not in a project stays in the ordinary chat list, which is why the sidebar can show projects and an everything-else view side by side.

| Cap | Value |
| --- | --- |
| Projects per account | 50 |
| Project name | 120 characters |
| Project instructions | 32,000 characters |
| Pinned files per project | 10 |
| Size of one pinned file | 10 MB |
| Extracted text from one pinned file | 32,000 characters |
| The whole project prompt block | 120,000 characters |
| Creating or editing a project | 20 per minute |
| Uploading a project file | 10 per minute |

Pinned files accept PDF, plain text, Markdown, CSV, JSON, .docx and .xlsx. Images are deliberately excluded, because a pinned file rides along as prompt text on every turn rather than being looked at once.

Two consequences follow from that, and they are the reason a project is more than a folder. The first is cost of context: the project instructions and the pinned file text are rebuilt into the prompt on every single message in that project, inside the same input budget as the conversation itself. The detail on what that spends is in [when the conversation is too long](https://whizi.io/docs/context-too-long). The second is retention: pinned file bytes are stored until you delete them and do not expire, unlike chat attachments, which are configured to expire after up to 30 days.

Project instructions are also the only user-authored standing instructions in the product. There is no account-level custom instructions setting, so if you want the same brief applied to every chat about one piece of work, a project is where that lives.

#### Deleting chats, and what deletion actually removes

A chat can be deleted from the chat list, one at a time or as a bulk selection. A failed single delete shows `Failed to delete chat. Please try again.`, and a failed bulk delete shows `Failed to delete some chats`, with no trailing full stop on the second one.

Deleting a conversation is not the same action as deleting your account. Individual conversations can be deleted on their own, and cancelling a subscription does not delete anything: it stops billing and ends access at the end of the paid term while the account and its contents remain. Removing the data is the separate account deletion action.

Deleted rows do not vanish from every layer at once. Cloudflare D1 keeps protected point-in-time recovery history, and deleted database records may remain in that history for up to 30 days before aging out. Where each kind of content lives and how long it is kept is set out in [where Whizi stores your data](https://whizi.io/docs/where-whizi-stores-data).

One gotcha specific to imported history: deleting an imported chat and importing the same export again brings the chat back, rather than being skipped as a duplicate. Imports are keyed by account plus source plus source id, so re-running an import is normally safe and idempotent, but it is also the one way a chat you deleted returns. The rest of the import behavior is on [chat import errors](https://whizi.io/docs/chat-import-errors).

#### When organizing actions fail for a reason that is not about organizing

One error shows up across all of these and has nothing to do with the chat itself. `Failed to retrieve session token.` is thrown by the website whenever the token fetch returns nothing before a chat list, chat load, delete, rename or upload call. It carries no HTTP status and no error code, because the site throws it before the request is made.

`Conversation not found.` is the backend counterpart, HTTP 404 code `not_found`, returned when the conversation id does not exist or is not owned by the caller. It appears on the read, delete, rename, message-append, version-switch and share routes alike.

Both are sign-in problems rather than data problems. The fixes, and the full list of session error strings, are on [login and account access](https://whizi.io/docs/login-and-account-access).

Key points:

- Search matches chat titles only, which makes renaming the thing that decides what you can find later
- Projects are folders that also carry pinned files and their own instructions, on every plan including free
- Every cap and every error string, including the rename limit that differs between the website and the backend

Checklist:

- Search matches chat titles only, never message bodies
- Rename accepts 1 to 120 characters, though the website refuses anything over 100
- Search cuts across projects and the ungrouped list at once
- Projects are available on every plan including free, up to 50 per account
- A chat belongs to at most one project
- Pinned project files ride along on every turn, so 10 files is a context cost as well as a cap
- Pinned project files do not expire, unlike chat attachments
- Deleting an imported chat and re-importing the same export brings it back
- Deleting a chat is not the same as deleting your account

FAQ:

Q: Can I search inside my Whizi conversations?
A: You can search the chat list, but the match is on the chat title only. Message bodies are not searched, so a phrase you remember from an answer will not find the thread it came from. The search does cut across projects and the ungrouped chat list together, so a chat is findable by name whichever folder it is in. In practice this makes renaming the important habit: the title is the index.

Q: How do I rename a chat in Whizi?
A: Rename it from the chat list. A title must be between 1 and 120 characters, and the backend refuses anything else with `Conversation titles must be between 1 and 120 characters.` The website is stricter and stops you at 100 with the toast `Chat name is too long (max 100 characters).`, and refuses an empty name with `Chat name cannot be empty.`

Q: Does Whizi have folders for chats?
A: Yes, they are called projects. A project is a folder of conversations that also carries up to 10 pinned files and its own instructions, up to 32,000 characters, applied to every conversation in it. You can have 50 projects per account, and they are available to any signed-in account including the free tier. A chat belongs to at most one project, and anything unfiled stays in the ordinary chat list.

Q: Do pinned project files count against the conversation length?
A: Yes. The project instructions and the extracted text of every pinned file are rebuilt into the prompt on every single turn in that project, inside the same input budget as the conversation. Each file contributes at most 32,000 characters of text and the whole project block is capped at 120,000 characters. If a project chat seems to lose the thread faster than an ordinary one, that is why, and unpinning what you are not asking about is the fix.

Q: If I delete a chat, is it really gone?
A: The conversation is deleted from your account without touching anything else, and you do not need to delete your account to remove one chat. Deleted database rows may persist for up to 30 days in Cloudflare D1 point-in-time recovery history before aging out. The exception worth knowing is imported history: deleting an imported chat and importing the same export again restores it rather than skipping it as a duplicate.

Q: Are there keyboard shortcuts for managing chats?
A: No. Enter sends a message, Shift plus Enter inserts a newline, and Escape closes an open dialog. Beyond those two conventions there is no shortcut system in Whizi, no chord bindings, and no shortcuts help panel.

### Out of credits in Whizi: what happens, and when they reset

URL: https://whizi.io/docs/out-of-credits/
Updated: August 2026

Quick answer: When your credit allowance runs out, Whizi refuses the next message as HTTP 429 with the code message_limit_reached, and the refused turn is not charged. The allowance comes back when the UTC calendar month rolls over rather than on your billing anniversary. There is no top-up, no credit purchase and no gifting.

The exact message Whizi shows when your credit allowance is spent, why a message can be refused while credits remain, and the calendar rule that decides when the allowance comes back.

#### The short answer

When your credit allowance runs out, Whizi refuses the next message with `Your monthly message limit has been reached.`, returned as HTTP 429 with the code `message_limit_reached`. The refused turn is not charged, the conversation is intact, and everything you have already written stays readable.

The allowance comes back when the period changes, and the period is the UTC calendar month rather than your billing anniversary. There is no top-up, no credit purchase and no gifting path anywhere in Whizi, so there are exactly two moves: wait for the reset, or move to a plan with a larger allowance.

Find the string you actually saw here first.

| What you see | Code | What it means | What to do |
| --- | --- | --- | --- |
| `Your monthly message limit has been reached.` | HTTP 429 `message_limit_reached` | Your credit allowance for this UTC month is spent | Wait for the reset, or upgrade |
| `Your weekly message limit has been reached.` | HTTP 429 `message_limit_reached` | The same refusal on a weekly billing cycle | Wait for Monday, or upgrade |
| `Your free messages are used up. Start a subscription to keep chatting.` | HTTP 402 `free_limit_reached` | A free account spent its 7 lifetime messages | Nothing resets this. A subscription is the only path |
| `A Deep Research run costs 200 credits and you do not have enough left this period.` | HTTP 429 `message_limit_reached` | One research run needs 200 credits and your balance is short | Send a normal message instead, or wait for the reset |
| `Upgrade your plan to use this model.` | HTTP 403 `tier_upgrade_required` | Not a credit problem at all: the model needs a higher tier | Pick a model on your plan, or upgrade |

The allowances are fixed per tier, and the gaps between them are large:

| Plan | Credits per month | Monthly price |
| --- | --- | --- |
| Free | 0 credits, 7 messages for the life of the account | not applicable |
| Starter | 400 | $15.99/month |
| Pro | 2,000 | $29.99/month |
| Powerhouse | 8,000 | $49.99/month |

#### The wording is the message limit wording, on purpose

There is no separate "out of credits" string in ordinary chat. When the credit balance is what refused the turn, Whizi still says `Your monthly message limit has been reached.`, because both allowances are enforced through the same period check. Deep Research is the one place with its own credit-specific wording.

The noun changes with your billing cycle and nothing else: `Your weekly message limit has been reached.` on a weekly cycle, `Your monthly message limit has been reached.` otherwise. Weekly billing is not sold on the website.

On the web app the refusal opens the upgrade dialog rather than a toast. That dialog is triggered by an HTTP 402, by the code `message_limit_reached`, or by any message containing the text "message limit reached". Its badge reads `Monthly Limit Reached`, and its description reads `You have reached your monthly limit of {limit} credits on the {plan} plan. Upgrade now to continue using Whizi without interruption.`, with the limit and plan filled in from your account.

Two shorter variants of that prompt exist. A paid account can see `You've reached your monthly limit on the {plan} plan. Upgrade now to continue using Whizi without interruption.` A free account sees `Choose a plan to unlock Whizi with higher limits and premium models.` instead, which is the tell that you are on the free tier rather than out of credits: free accounts are given 0 credits and a 7 message lifetime allowance instead.

#### When credits reset

There is no grant event and no top-up job. Usage is summed against a period key, so the arrival of a new period key is itself the reset. Nothing has to run, and nothing can be late.

| Billing cycle | Period key | Format | The allowance returns |
| --- | --- | --- | --- |
| Monthly | UTC calendar month | `YYYY-MM` | At 00:00 UTC on the 1st |
| Weekly | ISO-8601 UTC week, Monday to Sunday | `YYYY-Www` | At 00:00 UTC on Monday |

The boundary is the calendar, not the day you subscribed. If you subscribed on the 20th, your credits do not return on the 20th. They return when the UTC month rolls over. In a timezone behind UTC the reset lands during the last day of your local month rather than at your local midnight.

Credits do not roll over. The balance query sums only the rows carrying the current period key, so an unspent balance is simply unreachable once the key changes. Spending nothing in one month does not make the next one larger.

The free tier has no reset at all. Its allowance is 7 messages counted for the life of the account, and waiting does not restore it. A subscription is the only thing that changes that number.

#### Why a message is refused while you still have credits

A turn that does not fit in the remaining balance is refused whole. Partial charging never happens, so a 20 credit message sent with 3 credits left is rejected rather than discounted down to 3.

That is deliberate. Charging 3 credits for a 20 credit model would mean the same model cost different amounts depending on when in the month you sent it, which makes the price impossible to reason about. A turn also never costs less than one credit: the charge is floored at 1 and rounded down.

A balance too small for the model you picked is often still enough for a cheaper one, and the cheapest rung is 1 credit. With 3 credits left you cannot send anything on the 20 credit rung, but you can send three more messages on a 1 credit model. Roughly half the catalogue costs 1 credit. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation), and the per family references that list the cost of every row, such as [DeepSeek](https://whizi.io/docs/using-deepseek-in-whizi) and [Qwen](https://whizi.io/docs/using-qwen-in-whizi).

The balance is a sum of credits rather than a count of messages, which is why it can drop by 20 in one turn. Every model row served to the web app carries a cost signal, so the price of the next message is visible before you send it. Auto is the exception: it ships a per-tier credit range rather than a single number, because a turn costs whatever model Auto picked. The full scale is in [how credits work](https://whizi.io/docs/credits).

A 20 credit model spends a 2,000 credit Pro allowance in 100 turns, while a 1 credit model gives you 2,000. Which model answers your routine messages therefore moves the balance far more than message count does. The fix is in [troubleshooting](https://whizi.io/docs/troubleshooting).

#### What running out does not affect

Image generation and voice are separate meters with their own counts and their own strings, so an empty credit balance does not produce either of these, and hitting either of these does not touch your credits.

| Meter | Its own error | Code |
| --- | --- | --- |
| Image generation | `Your monthly image generation limit has been reached.` | HTTP 429 `image_limit_reached` |
| Voice minutes | `Your monthly voice minutes have been used up.` | HTTP 429 `voice_limit_reached` |

Both have weekly variants with the same substitution the chat string uses. Neither is a credit refusal, so neither is fixed by waiting for the credit reset if your cycle boundary has not arrived. The other reasons an image will not generate are in [image generation failed](https://whizi.io/docs/image-generation-failed).

#### Charges, refunds, and the retry trap

A generation that fails or is aborted is refunded by deleting its usage row, and applying the same refund twice restores exactly one turn rather than two. So a reply that broke mid-stream should not leave you paying for nothing.

Deep Research states this in the failure text itself. `The research run failed. You have not been charged.` is written to a run that threw for a reason other than a provider error, `Planning the research run failed. You have not been charged.` when the plan stage cannot be parsed after one retry, and `Web searches failed. You have not been charged.` when too few searches complete. A run refused with `A Deep Research run is already in progress. Wait for it to finish.` (HTTP 409 `research_in_progress`) also has its reserved credits refunded.

A chat turn rejected for being out of quota still spends its rate-limit tokens, because otherwise an out-of-quota client could retry without limit. Hammering send after the refusal therefore earns a rate-limit block on top of the refusal you already have, and it never produces a reply. The strings, codes and caps for that second block are in [rate limits](https://whizi.io/docs/rate-limits).

If the balance looks wrong rather than spent, [support](https://whizi.io/support) is the place to raise it, and reading the figure correctly in the first place is covered in [check your credit balance](https://whizi.io/docs/check-your-credit-balance). The usage rows behind a balance are kept for 400 days.

Key points:

- The verbatim string and HTTP code for each way a spent allowance refuses a message
- The exact reset boundary: the UTC calendar month, not your billing anniversary
- Why a 20 credit message is refused whole while a 1 credit message still sends

Checklist:

- Out of credits reads `Your monthly message limit has been reached.`, HTTP 429
- The refused message is not charged
- The reset is the UTC calendar month, not your billing anniversary
- Weekly cycles reset at 00:00 UTC on Monday
- Credits never roll over to the next period
- A turn is refused whole, so a cheaper model may still send
- Free accounts get 7 lifetime messages and no reset at all
- No top-up, no credit purchase, no gifting: wait or upgrade

FAQ:

Q: What does "Your monthly message limit has been reached." mean in Whizi?
A: Your allowance for the current period is spent. It is HTTP 429 with the code `message_limit_reached`, and it is the same string whether the credit allowance or the message allowance is what refused the turn, because both are enforced through one period check. The refused message is not charged, and the conversation is untouched.

Q: When exactly do Whizi credits reset?
A: At 00:00 UTC on the 1st of the month, or at 00:00 UTC on Monday if you are on a weekly cycle. It is never your billing anniversary, so subscribing on the 20th does not mean credits return on the 20th.

Q: Can I buy more credits when I run out?
A: No. Nothing adds credits to a period that is already running. The balance changes only when the period rolls over, or when you move to a plan carrying a bigger one: 400 credits on Starter, 2,000 on Pro, 8,000 on Powerhouse.

Q: Why was my message refused when I still had credits left?
A: Because a turn is charged whole or not at all. A 20 credit model asked to run on a 3 credit balance is rejected rather than part-charged, so the same model never costs different amounts depending on when you sent it. Those 3 credits are still spendable on a cheaper model, and roughly half the catalogue costs 1 credit.

Q: I am on the free plan and out of messages. When does that reset?
A: It does not, which is why the string you get is `Your free messages are used up. Start a subscription to keep chatting.` rather than a limit-reached message. Claiming a guest session into an account carries its usage across as well, so spending the allowance as a guest and then registering starts you at zero rather than at seven.

Q: Was I charged for an answer that failed halfway through?
A: No. To confirm the refund landed, check the credit balance rather than a message count: the balance is the sum of the credits on the rows for this period, so a refunded 20 credit turn puts 20 back and not one message. Deep Research says so in its own failure text, for example `The research run failed. You have not been charged.`

### Card declined or payment failed at Whizi checkout

URL: https://whizi.io/docs/payment-declined/
Updated: August 2026

Quick answer: If your card was genuinely declined at Whizi checkout, the refusal came from your bank through Stripe rather than from Whizi, so the fix is a different card or a call to your bank. Whizi has no card declined message anywhere in it, so Whizi wording on the error means it was not a decline.

A real decline comes from your bank through Stripe, so Whizi never sees the reason. Every error Whizi itself can show at checkout, quoted exactly, with the condition that triggers it and what fixes it.

#### The short answer

If your card was genuinely declined, the refusal came from your bank through Stripe, not from Whizi, so the fix is a different card or a call to your bank. Why banks block AI subscription payments in the first place, and how to tell which of the six causes you hit, is in [paying for AI when your card is declined](https://whizi.io/resources/paying-for-ai-when-your-card-is-declined). Whizi does not have a card declined message anywhere in it. Card details are entered on Stripe's hosted checkout page, which sits outside Whizi's codebase, and Whizi's Stripe webhook handler has no declined payment message to show you.

So if you saw an error with Whizi wording on it, it was not a decline. It was one of the strings below, and each one has a different cause and a different fix.

The three that are easiest to mistake for a decline:

| What you saw | What it means | What to do |
| --- | --- | --- |
| `A $0.99 intro trial has already been used with this payment method.` | That card has already claimed the intro trial | Subscribe at the normal rate, or use a card that has not claimed it. Retrying will not work |
| `Pricing is temporarily unavailable.` | Whizi could not read the price table from Stripe | Wait and try again later. Nothing on your side fixes this |
| `Failed to create Stripe checkout session.` | The checkout call failed and the backend supplied no message | Retry once. If it repeats, contact support |

#### Every error the Whizi backend returns at checkout

These come from the server, so they arrive with an HTTP status and an error code that support can match in a network trace. Every backend error body is JSON in the shape `{"error": {"code": "...", "message": "..."}}`, so the code sits next to the message. The rows below are ordered by HTTP status.

| Message | HTTP | Code | Triggered when |
| --- | --- | --- | --- |
| `A plan and billing cycle are required.` | 400 | `invalid_checkout_request` | The checkout body could not be parsed |
| `Invalid plan or billing cycle.` | 400 | `invalid_plan_or_cycle` | The plan or cycle does not map to a known Stripe price |
| `No subscription item found for the active subscription.` | 400 | `subscription_item_missing` | An in place plan change ran against a subscription with no item on it |
| `The supplied discount must be an active 100% promotion code or coupon ID.` | 400 | `invalid_discount` | The code supplied is not an active 100 percent promotion code or coupon ID |
| `Temporary email addresses are not allowed. Please use a permanent email address.` | 400 | `disposable_email_blocked` | The signed in account's email is on the disposable domain list |
| `A $0.99 intro trial has already been used with this payment method.` | 403 | `duplicate_trial_payment_method` | The card fingerprint has already claimed the intro trial |
| `User not found.` | 404 | `not_found` | The checkout route could not read your user row |
| `Pricing is temporarily unavailable.` | 503 | `pricing_unavailable` | The localized price table could not be read from Stripe |

What to do with each. The intro trial refusal has its own section below. `Pricing is temporarily unavailable.` is server side: HTTP 503, returned when the localized price table cannot be read from Stripe, and there is no setting on your account that changes it. `Failed to create Stripe checkout session.` is not a server message at all, which is why it names no reason: it is the website's own fallback text, shown when the checkout call fails and the backend supplied no message. Both are wait and retry.

`A plan and billing cycle are required.` and `Invalid plan or billing cycle.` both mean the request did not carry a plan the server recognises, so start again from the pricing page and pick the plan and the monthly or annual toggle there rather than reusing an old or edited checkout link.

`User not found.` means your session points at a user row the server cannot read. Sign out, sign back in, then start checkout again.

`No subscription item found for the active subscription.` fires during a plan change rather than a first purchase, and it is not fixable from the interface. Contact support.

#### Errors the website shows before checkout even opens

These come from the browser, not the server, so they have no HTTP code. They all mean the same broad thing: something failed while Whizi was preparing your account for the handoff to Stripe, so no charge was ever attempted.

| Message | Triggered when |
| --- | --- |
| `Failed to create Stripe checkout session.` | The checkout call failed and the backend supplied no message |
| `Could not open checkout. Please try again.` | The upgrade modal's checkout call threw with no message |
| `Authentication failed. Please sign in again.` | The token fetch returned nothing before checkout |
| `Failed to save user before checkout.` | The pre checkout user save call was not OK |
| `Could not prepare your account for checkout.` | The pricing page could not save the user before checkout |
| `Missing user or Stripe customer details.` | The user id was absent at checkout time |
| `Checkout response did not include a Stripe URL.` | Checkout succeeded but returned neither a URL nor a subscription id |
| `Something went wrong while starting your trial. Please try again.` | The pricing page's generic trial start failure |

For anything on this list, sign out, sign back in, and start checkout again from the pricing page. That clears the cases where the token fetch returned nothing before checkout, which is what `Authentication failed. Please sign in again.` reports.

`Checkout response did not include a Stripe URL.` is the one worth pausing on. It means the call succeeded, so before you try to pay again, open your account settings and check whether a subscription is already there.

Two more appear on the sign in and sign up pages when you pick a plan first and authenticate second: `Signed in, but checkout could not start. Please choose your plan again.` and `Your account was created, but checkout could not start. Please sign in and choose your plan again.` Both mean the account part worked and only the handoff failed. Your account exists. Go to pricing and pick the plan again.

One last cause of a mysterious failure: retrying checkout many times in a row can trip the rate limiter, which returns HTTP 429 and `Too many requests. Please wait and try again.` Every request on the account, whatever the route, counts against a ceiling of 30 per minute. Wait a minute rather than clicking again.

#### The intro trial refusal, which no retry will clear

`A $0.99 intro trial has already been used with this payment method.` is HTTP 403 with the code `duplicate_trial_payment_method`. It is the one checkout error on this page that a retry will never clear.

The check is keyed to the card rather than to the account: the refusal is returned when the card fingerprint has already claimed the intro trial.

Your two options: subscribe at the standard rate for the plan you want, or pay with a payment method that has not claimed the trial before.

Plan prices, the trial terms and what happens at conversion are all in [billing, trials and how to cancel](https://whizi.io/docs/billing-and-cancellation).

#### Payment went through but the account is still on the free plan

A charged card plus an account that still behaves as free is a different symptom from checkout failing, and it has its own page.

Read [you paid but the plan features are still locked](https://whizi.io/docs/subscription-not-active) for the strings that appear in that state, the check that tells you which account holds the subscription, and what to send support.

#### What Whizi cannot tell you

Three gaps, so you do not spend time hunting for answers that are not there.

Whizi never sees why your bank declined a card, because that exchange happens on Stripe's hosted checkout page, outside Whizi's codebase. A sweep of both Whizi repos found no card declined or payment declined string of any kind, and the Stripe webhook handler carries no declined payment message. So there is no decline reason stored in your account, no retry button, and nothing for Whizi to override. The reason sits with your bank.

Whizi also has no message for a failed or past due renewal. No user-facing string for one exists in either repo, so if a renewal charge fails later you will not get a Whizi error explaining it. The one renewal adjacent notice that does exist is a push notification reading "Your subscription is no longer active." under the title "Whizi subscription updated", sent when a RevenueCat webhook syncs an account down to the free tier. That is the RevenueCat webhook path, so it is not a report of a Stripe renewal failing.

Whizi billing runs through Stripe, so which cards can pay at all is decided there rather than in Whizi. A Russian or Iranian card cannot complete registration. Ukraine, Kazakhstan, Turkey, Indonesia and India work end to end.

Key points:

- Whizi contains no card declined message, so a real decline is your bank rejecting the charge
- Every checkout error Whizi can return, quoted exactly, with its HTTP code and trigger
- The one refusal that no retry will ever clear: the intro trial is keyed to the card

Checklist:

- A real card decline comes from your bank through Stripe, not from Whizi
- Whizi has no card declined message and no decline reason to show you
- Whizi wording on the error means it was not a decline
- Session errors clear by signing out and back in, then retrying from pricing
- Pricing is temporarily unavailable is server side: wait, do not retry in a loop
- The intro trial is keyed to the card fingerprint rather than to the account
- Charged but still on the free plan is a different symptom, covered in subscription-not-active

FAQ:

Q: Why does Whizi say my payment failed but my bank shows no attempt?
A: Because the string you saw was raised by the website before Stripe was ever reached, so no charge was attempted. The quick way to tell one of those apart from a real backend refusal: a backend refusal arrives as JSON carrying an error code, such as `pricing_unavailable` or `invalid_plan_or_cycle`, and shows an HTTP status in a network trace. A browser-side string has neither. If what you saw carried no code, nothing reached your bank.

Q: What does `A $0.99 intro trial has already been used with this payment method.` mean?
A: It means the card you entered has already claimed the intro trial. It is not a decline: Whizi has no card declined string at all, and this one comes from Whizi's own checkout route as HTTP 403 with the code `duplicate_trial_payment_method`, which is worth quoting if you write in. Either subscribe at the standard rate or use a payment method that has not claimed the trial.

Q: I got `Pricing is temporarily unavailable.` Is my card the problem?
A: No. It is HTTP 503 `pricing_unavailable`, and it comes from the route that reads the localized price table from Stripe, not from the checkout route that would take a card. There is no fix on your side, so wait rather than trying a different card. If you saw `Failed to create Stripe checkout session.` instead, that is a separate string with a separate origin: the website's fallback when the checkout call failed and the backend supplied no message.

Q: My promo code was rejected. Why?
A: Checkout returns `The supplied discount must be an active 100% promotion code or coupon ID.` as HTTP 400 `invalid_discount`. The discount field at checkout only accepts an active 100 percent promotion code or coupon ID, so an expired code or a partial discount code will be refused there. Remove it and check out at the normal price if you need access now.

Q: Whizi rejected my email at checkout. What happened?
A: You saw `Temporary email addresses are not allowed. Please use a permanent email address.` The checkout route returns it as HTTP 400 `disposable_email_blocked` when the signed in account's email domain is on the disposable list. The same string is used at sign up as HTTP 403. Use a permanent address on the account and checkout will proceed.

Q: I was charged twice. What should I do?
A: Contact support with both Stripe receipts rather than cancelling blindly. Cancelling is a separate action with its own outcome: for an active subscription the cancel route answers `Subscription cancellation scheduled for end of billing period.`, so that subscription runs to the end of the period it is already paid for.

### Whizi plans, prices and limits

URL: https://whizi.io/docs/plans-and-limits/
Updated: August 2026

Quick answer: Whizi has three plans: Starter at $15.99 per month, Pro at $29.99 and Powerhouse at $49.99, or $10.99, $19.99 and $34.99 per month billed yearly. All three include web, iOS and Android access on one login and file upload. Monthly credits are 400, 2,000 and 8,000. Every new account starts with a 7 day trial.

Every number in one place: what Starter, Pro and Powerhouse cost, what each includes, and every monthly limit that applies to an account.

#### The short answer

Whizi has three plans. All of them include web, iOS and Android access on one login, and all of them include file upload. The difference is which models you can reach and how much you can send.

| | Starter | Pro | Powerhouse |
| --- | --- | --- | --- |
| Monthly | $15.99 | $29.99 | $49.99 |
| Yearly, per month | $10.99 | $19.99 | $34.99 |
| Yearly, billed once | $131.88 | $239.88 | $419.88 |
| Credits per month | 400 | 2,000 | 8,000 |
| Text models | 3 | about 40 | 280 plus |
| Image generations | none | 100 | 500 |
| AI voice | 10 minutes | 80 minutes | 500 minutes |
| AI video | none | none | 100 |
| AI music | none | none | included |
| Side-by-side chats | no | no | yes |
| Agent personas | no | yes | yes |

For context, a single ChatGPT Plus subscription is $20 per month and a single Claude Pro subscription is $20 per month. Whizi Pro at $19.99 billed yearly covers both model families plus Gemini, Grok, DeepSeek, Llama, Mistral, Kimi and Qwen.

#### The trial

New accounts start with a 7 day trial for $0.99. It is a paid trial rather than a free one, which filters out most of the abuse that forces other products into aggressive rate limits.

The trial converts to the plan you selected at signup unless you cancel before it ends. Cancelling during the trial keeps access to the end of the 7 days and then stops, with nothing further charged.

You can cancel from account settings on the web at any time. Subscriptions bought through the iOS or Android app are cancelled through Apple or Google instead, because the store owns that billing relationship. See [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

#### What counts against a limit and what does not

Four separate meters exist on an account, and they do not share a pool.

| Meter | What spends it | Reset |
| --- | --- | --- |
| Credits | One per text message, multiplied by the model rate | Monthly, on the billing date |
| Image generations | One per generated image | Monthly |
| Voice minutes | Time spent in voice mode | Monthly |
| Video generations | One per generated video | Monthly |

Things that spend nothing: uploading a file, turning on web search, reading or re-reading a past conversation, sharing a conversation, switching models mid-thread, and editing a prompt before you send it.

One thing that costs double: side-by-side comparison. Two models each produce an answer, and each is billed at its own credit rate, so comparing Claude Sonnet 5 against GPT-5.6 Sol costs 30 credits rather than 10 or 20.

#### Which plan actually fits

The honest version, given the model tiers rather than the marketing.

**Starter** makes sense if you want a cheap, fast assistant and the ability to ask questions about your own documents, and you do not care which model answers. Three models, all at 1 credit, 400 messages per month. It does not include any Anthropic model, so if Claude is the reason you are here, this is not the plan.

**Pro** is the plan most people want. It is the tier that adds Claude Sonnet 5 and the GPT-5.6 Terra class, which together cover the work that a ChatGPT Plus and Claude Pro pair would cover, at less than either one alone when billed yearly. 2,000 credits is a serious allowance: 200 Claude Sonnet messages plus effectively unlimited fast-tier use.

**Powerhouse** is for people who specifically need the frontier tier, the media generation, or side-by-side comparison. Claude Opus 5, GPT-5.6 Sol, AI video and AI music all live here and nowhere else. If you are not going to use any of those four things, Pro is the better value and the model list is not the reason to upgrade.

Key points:

- All three plans priced monthly and yearly, with the real yearly total
- Every quota that exists on an account, in one table
- What the $0.99 trial does and does not include

Checklist:

- Yearly billing is roughly a 30 percent discount across all three plans
- The trial is 7 days for $0.99 and converts unless cancelled
- Starter includes no Anthropic model
- Pro is the tier that adds Claude Sonnet and GPT-5.6 Terra
- Claude Opus, GPT-5.6 Sol, video, music and side-by-side are Powerhouse only
- Credits, images, voice minutes and videos are four separate meters
- Uploading files and using web search cost nothing

FAQ:

Q: How much does Whizi cost?
A: Starter is $15.99 per month or $10.99 per month billed yearly, Pro is $29.99 or $19.99 billed yearly, and Powerhouse is $49.99 or $34.99 billed yearly. Every new account starts with a 7 day trial for $0.99. Yearly billing charges once at $131.88, $239.88 or $419.88 respectively.

Q: Is Whizi cheaper than paying for ChatGPT Plus and Claude Pro separately?
A: Yes, substantially. ChatGPT Plus and Claude Pro are $20 per month each, so $40 per month for two model families. Whizi Pro at $19.99 per month billed yearly covers both plus Gemini, Grok, DeepSeek, Llama, Mistral, Kimi and Qwen, which is roughly half the price for far more models. The trade is the credit allowance: unlimited use of one model is not the same shape of deal as a metered allowance across forty.

Q: Can I change plans mid-month?
A: Yes. Upgrading takes effect immediately so the new models and the larger allowance are available right away. Downgrading takes effect at the end of the current paid term, so you keep what you paid for until the period you already bought runs out.

Q: What happens to my conversations if I downgrade?
A: They stay. Downgrading changes which models you can send new messages to, not what you can read, so past conversations with a model you no longer have access to remain fully readable. You just cannot continue them on that model until you upgrade again, and the app will offer a model you do have access to instead.

Q: Is there a free plan?
A: There is no ongoing free tier. The entry point is the 7 day trial for $0.99, which is deliberate: a paid trial is a much stronger filter against automated abuse than a free one, and the alternative is the aggressive rate limiting that free tiers force on everybody else.

### Whizi rate limits: the numbers, the errors, and what to do

URL: https://whizi.io/docs/rate-limits/
Updated: August 2026

Quick answer: Whizi limits chat to 10 messages per minute and 60 messages per hour per user, and every request also passes a shared ceiling of 30 requests per minute per user, plus 60 per minute per IP address. Those burst limits are flat across every plan, and a refused request costs nothing off your allowance.

Whizi allows 10 chat messages per minute and 60 per hour per user, plus a 30 request per minute ceiling across every route. Every 429 string, what trips it, and the fix.

#### The short answer

Whizi limits chat to 10 messages per minute and 60 messages per hour per user, and every request of any kind also passes a shared ceiling of 30 requests per minute per user. Those burst limits are flat across free, Starter, Pro, and Powerhouse. Upgrading does not raise them.

A refused request costs nothing. Rate limit refusals are unwound, so a 429 never spends a message off your plan allowance.

| Limit | Value | Scope |
| --- | --- | --- |
| Chat messages per minute | 10 | Per user |
| Chat messages per hour | 60 | Per user |
| All routes combined | 30 requests per minute | Per user |
| All routes combined | 60 requests per minute | Per IP address |

The per IP ceiling is deliberately looser than the per user one because IP addresses are shared.

The fix for all four is the same: wait. The windows are short, so in practice this means a minute.

#### Every rate limit error string, verbatim

Rate limit refusals return HTTP 429 with the code `rate_limit_reached`. Image uploads carry their own codes, `upload_rate_limited` and `upload_daily_limit_reached`.

| Message | HTTP and code | What tripped it | What to do |
| --- | --- | --- | --- |
| `Too many requests. Please wait and try again.` | 429 `rate_limit_reached` | The per minute window for the route, the all routes ceiling, or the per IP window refused the request | Wait a minute and send again |
| `Hourly message limit reached. Please try again later.` | 429 `rate_limit_reached` | The hourly chat ceiling of 60 messages was the window that refused | Wait for the hour window to roll over |
| `Upload rate limit reached. Please wait a moment before uploading another image.` | 429 `upload_rate_limited` | More than 10 image uploads attempted in one UTC minute | Wait for the next UTC minute |
| `Daily upload limit reached. Please try again tomorrow.` | 429 `upload_daily_limit_reached` | 60 image uploads already made in one UTC day | Wait for the next UTC day |
| `That is a few in a row. Give it a minute, then send this again.` | Feedback request returned 429 | The feedback route allows 6 submissions per minute | Wait a minute, the text is not lost |
| `Too many reports right now. Please try again shortly.` | Share report endpoint returned 429 | The shared conversation report route allows 10 per minute, keyed by the share token | Wait and report again |

`Too many requests. Please wait and try again.` is the default message for every window except the hourly chat one, so it can mean the route limit, the account wide ceiling, or your network. `Hourly message limit reached. Please try again later.` is used only when the 60 per hour chat ceiling is the window that refused, which makes it the more precise diagnosis of the two.

#### The per route limits, in full

Each route carries its own per minute window on top of the account wide ceiling of 30 requests per minute. These are per user unless noted.

| Action | Limit |
| --- | --- |
| Sending a message on a given model | 10 per minute |
| Creating a conversation | 10 per minute |
| Searching conversations | 20 per minute |
| Importing conversations | 20 per minute |
| Creating or editing a project | 20 per minute |
| Uploading a project file | 10 per minute |
| Uploading an image | 10 per minute, and 60 per day |
| Starting a voice session | 6 per minute |
| Ending a voice session | 6 per minute |
| Submitting feedback | 6 per minute |
| Reporting an AI output | 6 per minute |
| Reporting a shared conversation | 10 per minute, keyed by the share token |
| Syncing a subscription | 6 per minute |
| Exporting your account data | 5 per minute |

The account data export is the tightest of them at 5 per minute, because an export reads everything the account owns.

One route retries for you rather than surfacing the error. When a chat import hits a 429, the client waits 20 seconds and retries, up to 4 attempts per batch, so a large ChatGPT or Claude import is expected to pause and continue on its own. See [chat import errors](https://whizi.io/docs/chat-import-errors).

#### 429s that are not about how fast you are sending

Three other refusals share the 429 status code and are routinely mistaken for burst limits. The code in the error body is what tells them apart.

| Message | HTTP and code | What it actually means |
| --- | --- | --- |
| `Your monthly message limit has been reached.` | 429 `message_limit_reached` | Your plan allowance for the period is spent, not a burst limit |
| `The model is temporarily rate-limited. Please retry shortly.` | 429 `provider_rate_limited` | The upstream model provider answered with a 429. Nothing on your account caused it |
| `Generation is paused on this account after repeated requests that Whizi cannot fulfil. It reopens automatically. Contact support from Settings if you think this is wrong.` | 429 `content_policy_locked` | A lockout earned by repeatedly tripping the content screener on generation prompts |

The plan allowance case is easy to identify without reading a network trace: the website opens its upgrade modal instead of showing a toast, badged `Monthly Limit Reached`. Waiting does not clear it inside the period. Either the billing period rolls over or you upgrade. The allowances are Starter 400 messages per monthly period, Pro 800, and Powerhouse 5,000. On the web client, which is metered in credits rather than messages, the same tiers are 400, 2,000, and 8,000 credits. Full detail in [plans and limits](https://whizi.io/docs/plans-and-limits) and [how credits work](https://whizi.io/docs/credits).

The provider case has a matching string on the voice path: `Realtime voice is temporarily rate-limited. Please retry shortly.` is the same 429 `provider_rate_limited` code, returned when the voice token mint upstream answers 429. The media generation paths have their own: `Image generation could not be started. Please try again.` is 429 `image_provider_rate_limited` when the image provider refuses, and `Image generation status is temporarily unavailable.` is 429 `image_status_failed` when the status poll is the call that gets refused. Video and music carry the same pair with their own codes. On the image submit path the reservation is refunded before the error is returned, and a generation that fails or is aborted is refunded. The rest of those strings are on [image generation failed](https://whizi.io/docs/image-generation-failed).

The content policy lockout is the only one on this page where the honest answer is to contact support. It reopens automatically, but no duration is published, and if you believe the screener is wrong about your prompts, support is the route.

#### The 409, 403 and 413 refusals that are not 429s

Four other refusals share the look of a rate limit and carry different status codes. The code in the error body is what tells them apart.

| Message | HTTP and code | What it actually is |
| --- | --- | --- |
| `Duplicate message ignored. Please wait for the current response to finish.` | 409 `duplicate_message` | The same message text was already sent to this conversation recently |
| `Generation already in progress. Please stop it or wait for it to finish.` | 409 `generation_in_progress` | The per user generation lock is already held |
| `Account sharing limit reached. Please use your own Whizi account.` | 403 `account_sharing_limit` | More than 3 distinct device or session keys have been seen on one account in the last 24 hours |
| `Request is too large.` | 413 `request_too_large` | A JSON body exceeded the byte cap for that route |

The first three belong to the send path and are covered in [a message will not send](https://whizi.io/docs/message-not-sending). For the 413, send less in one request rather than fewer requests.

Key points:

- The exact per minute and per hour limits, which are identical on every plan
- Every rate limit error string Whizi can return, with the condition that triggers it
- The three other 429s that are not burst limits: plan allowance, provider, and a content policy lockout

Checklist:

- Chat is 10 messages per minute and 60 per hour, per user
- Every route also passes a 30 requests per minute per user ceiling
- The per IP ceiling is 60 requests per minute, looser because IPs are shared
- Burst limits are identical on free, Starter, Pro, and Powerhouse
- A 429 never spends a message off your allowance
- Code rate_limit_reached means wait, message_limit_reached means plan
- provider_rate_limited is upstream and nothing on your account caused it
- Chat imports retry a 429 themselves after 20 seconds, up to 4 attempts

FAQ:

Q: What does "Too many requests. Please wait and try again." mean in Whizi?
A: It is HTTP 429 with the code rate_limit_reached, and it means one of three windows refused the request: the per minute limit for that route, the account wide ceiling of 30 requests per minute, or the per IP ceiling of 60 per minute. Wait a minute and send again.

Q: How many messages can I send per minute?
A: Ten per minute and 60 per hour, per user. Sending on a given model is separately capped at 10 per minute, and creating a conversation is also 10 per minute. Those numbers are the same on every plan, so a Powerhouse subscriber has exactly the same burst limits as a free account.

Q: Does upgrading raise my rate limit?
A: No. The 10 per minute and 60 per hour chat limits are flat across free, Starter, Pro, and Powerhouse, and they sit in front of the plan quota layer rather than inside it. What upgrading raises is the period allowance: Starter 400 messages per monthly period, Pro 800, and Powerhouse 5,000, or 400, 2,000, and 8,000 credits on the web client. If the error says your monthly message limit has been reached, upgrading helps. If it says too many requests, it does not.

Q: Why does it say "Hourly message limit reached. Please try again later." when I have credits left?
A: Because that string is the burst limiter, not the allowance. It fires only when the 60 messages per hour ceiling is the window that refused, and it is unrelated to how many credits or messages your plan has left for the period.

Q: The same model keeps returning "The model is temporarily rate-limited. Please retry shortly." What can I do?
A: Send the message on a model from a different provider family, or use Auto, which is a router rather than a model and only picks from models your plan already reaches. The refusal comes from the upstream provider rather than from your account, and Whizi does not publish a wait time for it, so retrying the same model is the only other option. The voice path returns the same code with its own wording: "Realtime voice is temporarily rate-limited. Please retry shortly."

Q: Do failed or rate-limited requests use up my credits?
A: No. Rate limit refusals are unwound, so a 429 never spends a message off your allowance, and a generation that fails or is aborted is refunded. Content policy refusals are screened before any quota is reserved and before the provider is called, so those cost nothing either.

### Does Whizi give refunds? What the Terms of Service say

URL: https://whizi.io/docs/refunds/
Updated: August 2026

Quick answer: Whizi publishes no refund policy for a subscription bought on whizi.io, and the Terms of Service contain no refund clause for one. The only remedy the Terms describe is cancellation, which takes effect at the end of the current paid term rather than returning the last charge. A store purchase is decided by Apple or Google.

Whizi publishes no refund policy for a subscription bought on whizi.io. What the Terms provide instead, what cancelling returns and what it does not, and where a billing dispute goes.

#### The short answer

Whizi publishes no refund policy for a subscription bought on whizi.io. The Terms of Service, last updated July 16, 2026, contain no refund clause for one. The only remedy the Terms describe is cancellation, and cancelling stops the next charge rather than returning the last one.

The Cancellation clause of section 6 is the whole of it, quoted exactly: "You can cancel your subscription at any time by logging into your account. Your cancellation will take effect at the end of the current paid term. If you have any questions or are unsatisfied with our Services, please email us at whizihelp@gmail.com."

| What people ask for | What the Terms say |
| --- | --- |
| Money back for the unused part of a paid term | No clause promises one. Cancellation takes effect at the end of the current paid term |
| Money back after changing plan | No clause promises one. Section 6 covers billing and renewal, the paid trial, cancellation and fee changes, and none of the four returns funds |
| Money back because a model was removed | No clause promises one. Whizi may add, replace, suspend, or remove a model or feature, and the Terms disclaim liability for any modification, price change, suspension, or discontinuance |
| A route to raise it with a person | Yes. The Cancellation clause names whizihelp@gmail.com for anyone unsatisfied |

A subscription bought inside the mobile app is a different route with a different answer. The same Terms say "Apple or Google controls payment processing, renewal, cancellation, and store refunds under the applicable store terms," and section 11 adds that on a failure of the App to conform to a warranty, "the App Distributor, in accordance with its terms and policies, may refund the purchase price, if any, paid for the App." That decision belongs to the store, under the store terms, not to Whizi. Everything below describes purchases made on whizi.io.

#### The clauses that decide what happens to a charge

Section 5 (Purchases and Payment) and section 6 (Subscriptions) are the two that decide what happens to a charge. Quoted, with the clause each line comes from:

| Clause | What it says |
| --- | --- |
| Section 5, accepted payment | Visa, Mastercard and American Express. "All payments shall be in US dollars." |
| Section 5, sales tax | "Sales tax will be added to the price of purchases as deemed required by us." |
| Section 5, price changes | "We may change prices at any time." |
| Section 5, authorising the charge | "You agree to pay all charges at the prices then in effect for your purchases and any applicable shipping fees, and you authorize us to charge your chosen payment provider for any such amounts upon placing your order." |
| Section 5, pricing errors | "We reserve the right to correct any errors or mistakes in pricing, even if we have already requested or received payment." |
| Section 5, refusing an order | "We reserve the right to refuse any order placed through the Services." |
| Section 6, Billing and Renewal | "Your subscription will continue and automatically renew unless canceled." The billing cycle is monthly or annually |
| Section 6, Paid Trial | "We offer a 7-day introductory trial for $0.99 to new users who register with the Services. The account will be charged according to the user's chosen subscription at the end of the introductory trial unless canceled." |
| Section 6, Cancellation | Effective at the end of the current paid term, with whizihelp@gmail.com named for anyone unsatisfied |
| Section 6, Fee Changes | Whizi may change the subscription fee and "will communicate any price changes to you in accordance with applicable law" |

None of them sends money back for a subscription bought on whizi.io.

Two more sit further down and matter if the reason you want money back is that the product changed. The Mobile and AI Service Terms say Whizi may add, replace, suspend, or remove a model or feature when a provider changes availability, pricing, safety, law, or technical requirements, followed by "We will not intentionally describe a paid plan in a materially misleading manner." Section 17 then says Whizi "will not be liable to you or any third party for any modification, price change, suspension, or discontinuance of the Services."

#### What cancelling returns

Cancelling ends the renewal and leaves the current term paid for. On a monthly plan that is the rest of the month. On a yearly plan it is the rest of the year, so cancelling in month three still ends the plan in month twelve. The Terms describe no mechanism that returns the months in between.

The cancel route answers with one of three messages, and none of the three mentions money. They report scheduling, not settlement.

When the paid term runs out you drop to the free tier. A free account is not model gated: it can try any text model, limited only by a 7 message lifetime allowance. The yearly cycle, where to cancel, and when a plan change takes effect are in [billing and cancellation](https://whizi.io/docs/billing-and-cancellation), and what the free tier does and does not include is answered in short in the [Whizi FAQ](https://whizi.io/docs/whizi-faq).

#### The other meaning of refund

Inside the product, "refunded" means credits rather than money, so a turn you paid credits for and did not receive is a balance question rather than a billing one, and it is answered in [out of credits](https://whizi.io/docs/out-of-credits).

#### If you believe a charge is wrong

The address the Terms name is whizihelp@gmail.com, in the Cancellation clause and again in section 28, which also gives the postal address: Whizi, 238 Featherstone Crescent, Kitchener, Ontario N2R 1Z3, Canada. What to put in that email, and the other routes to a person, are in [how to contact Whizi support](https://whizi.io/docs/contact-whizi-support).

Before writing, rule out the two cases that look like a billing error and are not. A card that was genuinely declined was declined by your bank rather than by Whizi, which is [payment declined](https://whizi.io/docs/payment-declined). A payment that went through while the plan stays locked is [subscription not active](https://whizi.io/docs/subscription-not-active).

Website payments are processed by Stripe, and Whizi does not store full payment-card numbers. What Whizi keeps after that, and for how long, is in [data and privacy](https://whizi.io/docs/data-and-privacy).

If it escalates, the Terms set the venue. They are governed by the laws of Canada, and section 18 says the courts of Canada have exclusive jurisdiction. Section 19 refers disputes to binding arbitration before the International Commercial Arbitration Court under the European Arbitration Chamber, with two arbitrators, the seat in Waterloo, Canada, and proceedings in English, limited to the dispute between the two parties with no class action basis.

Section 19 also carves three kinds of dispute out of arbitration: "(a) any Disputes seeking to enforce or protect, or concerning the validity of, any of the intellectual property rights of a Party; (b) any Dispute related to, or arising from, allegations of theft, piracy, invasion of privacy, or unauthorized use; and (c) any claim for injunctive relief." Those go to a court within the jurisdiction named above instead.

California residents get one extra named route. Section 26 says that if a complaint is not satisfactorily resolved, you can contact the Complaint Assistance Unit of the Division of Consumer Services of the California Department of Consumer Affairs, in writing at 1625 North Market Blvd., Suite N 112, Sacramento, California 95834, or by telephone at (800) 952-5210 or (916) 445-1254.

Key points:

- The Terms of Service contain no refund clause for a subscription bought on whizi.io
- Cancellation returns access to the end of the paid term, not money
- The clauses in sections 5 and 6 that decide what happens to a charge, quoted

Checklist:

- The Terms of Service contain no refund clause for a subscription bought on whizi.io
- Cancellation takes effect at the end of the current paid term
- A yearly plan is billed as one annual total, and cancelling does not return the unused months
- `Trial subscription canceled immediately.` is the one cancellation that is not scheduled
- A refund of credits is not a refund of money, and it is a different page
- whizihelp@gmail.com is the address the Terms name for anyone unsatisfied
- Disputes are governed by the laws of Canada and go to binding arbitration, with named exceptions

FAQ:

Q: Does Whizi offer refunds?
A: No refund policy is published for a subscription bought on whizi.io, and no clause in the Terms of Service returns money for one. Cancellation is the remedy they name. The one place the Terms attach a number to money owed to you runs the other way: section 22 limits Whizi liability for any cause of action "TO THE AMOUNT PAID, IF ANY, BY YOU TO US DURING THE six (6) mONTH PERIOD PRIOR TO ANY CAUSE OF ACTION ARISING."

Q: Can I get money back if I cancel in the middle of a month?
A: The Terms provide access rather than money. Cancellation takes effect at the end of the current paid term, so you keep the plan for the days you already paid for and nothing further is charged. There is no clause describing a return of the unused part, and no pro-rating mechanism is documented anywhere in the Terms.

Q: What about a yearly plan I barely used?
A: A yearly plan is billed as one annual total, shown as $131.88 for Starter, $239.88 for Pro or $419.88 for Powerhouse. Cancelling ends it at the close of that term, and the Terms describe no mechanism for returning the months you did not use. Monthly billing is the cycle that limits exposure to one term at a time, at $15.99/month for Starter.

Q: Is the $0.99 introductory trial charge refundable?
A: The Terms do not say. What they do say is that the trial is 7 days for $0.99 and that the account is charged for the chosen subscription at the end of it unless cancelled. Cancelling during the trial is the one cancellation the server handles on the spot rather than scheduling, and it answers "Trial subscription canceled immediately." Whether the introductory charge itself comes back is not stated in the Terms, so ask at whizihelp@gmail.com rather than assuming either answer.

Q: Whizi removed a model I was paying for. Can I get my money back?
A: The Terms reserve that right explicitly: Whizi may add, replace, suspend, or remove a model or feature when a provider changes availability, pricing, safety, law, or technical requirements, and section 17 disclaims liability for any modification, price change, suspension, or discontinuance of the Services. The one commitment attached to it is that Whizi "will not intentionally describe a paid plan in a materially misleading manner." A removed model is not a documented refund trigger, so the practical move is to cancel before the next renewal.

Q: I paid and my plan is still locked. Is that a refund question?
A: Not on its own. The cancel route is the quickest test: start a cancellation in account settings, and if the answer is "Subscription is already canceled. Your account is on the free plan." then this account has no Stripe customer on file, or a customer with no active subscription, so there is nothing on it to refund. The rest of that diagnosis is in the subscription not active page.

### Regenerate a response in Whizi, and edit a message you already sent

URL: https://whizi.io/docs/regenerate-and-edit-messages/
Updated: August 2026

Quick answer: Whizi has no regenerate control. Re-running a turn is done by editing your own message and sending it again, which replaces the reply rather than appending a second one and files the previous attempt as a version you can flip back to. The Edit control sits on your own messages only, never on an assistant reply.

Whizi has no regenerate button. Editing your own message and sending it again is how a turn is re-run, and each edit files the old reply as a version you can flip back to.

#### The short answer

Whizi has no regenerate control, and nothing in the worker serves one. Re-running a turn is done by editing your own message and sending it again, which replaces the reply rather than appending a second one, and files the previous attempt as a version you can flip back to.

The Edit control sits on your own messages only, never on an assistant reply, and the resend is charged as one ordinary turn at the rate of whichever model answers it.

| What you want | What Whizi has | Where it lives |
| --- | --- | --- |
| Re-run the same prompt | No regenerate control | Edit the message and send it again |
| Fix a prompt you got wrong | Edit and resend | The Edit control on your own messages |
| Compare the old answer with the new one | Version arrows reading like 2/2 | Under the edited turn |
| A different answer without rewording | Switch the model in the same conversation | The model picker |

#### What an edit actually changes

A sent user message can be edited and resent, and the reply that follows replaces the old one instead of stacking underneath it.

Each edit creates a version of the turn, and the arrows under it read like 2/2. The important part is what a version contains. It is not just the edited message: it is that message and every reply that followed it, so switching versions swaps the whole tail of the conversation rather than one bubble.

That has a consequence people meet by surprise. Editing a turn discards the version history of every turn after it, because those versions belong to a conversation that no longer exists. If you edited message three twice and then go back and edit message one, the two versions of message three go with it.

Versions are also why two branches of one thread cannot both stay live, which is the question answered on [can you fork a conversation](https://whizi.io/docs/fork-a-conversation).

Version switching is a server operation, not a client toggle: it is served by `POST /api/conversations/:id/versions`, and it takes the generation lock for the duration. Whizi runs one generation per account at a time, so a version switch and a reply in flight cannot happen at once.

Resending the identical text is not a second route to a new answer. The same message text sent to the same conversation recently is refused as a duplicate, so a re-roll of unchanged words is unavailable even in principle. That refusal is documented with the rest of the send failures on [a message will not send](https://whizi.io/docs/message-not-sending).

#### What re-running a turn costs

There is no regenerate price, because there is no regenerate. A resend is an ordinary turn, charged at the fixed per-message rate of whichever model answers it.

A model credit cost is a fixed integer per message. Nothing about message length, conversation length, or answer length enters the calculation, so a re-run of a long thread costs exactly what the first attempt cost on that model. A turn is floored at one credit and rounded down, so nothing is ever charged less than one.

Editing a prompt before you send it costs nothing in credits, and neither does switching the model you are about to send it to. The charge lands when a model answers.

| Action | Credits |
| --- | --- |
| Editing the text of a message before sending | None |
| Sending the edited message | One turn at the answering model rate |
| Stopping a reply part way | Refunded as an undelivered turn |

The refund is the reason to correct early rather than late. A generation that fails or is aborted is refunded by deleting the usage row, so an answer you stop in its first two lines costs less than one you let finish and then re-run. How a stop is handled on the server is on [a message will not send](https://whizi.io/docs/message-not-sending).

A different model is the other route to a different answer, and that turn is charged at the new model rate rather than the old one. The patterns worth using are on [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

On a paid plan this comes out of the credit allowance for the period. Free accounts have no credit allowance at all: they get a lifetime allowance of 7 messages that never resets, and a resend is a message like any other. The full scale and the per-plan allowances are in the [credit reference](https://whizi.io/docs/credits).

#### When an edit or a version switch fails

When the edit itself is refused, the strings and codes for it live with the other send failures on [a message will not send](https://whizi.io/docs/message-not-sending). The strings below belong to the version arrows and appear nowhere else.

| What you see | Code | Trigger |
| --- | --- | --- |
| `That version is no longer available. Reload and try again.` | 404 `version_not_found` | The version you switched to has been archived away |
| `A message position and a version number are required.` | 400 `invalid_version` | The version-switch request lacked a valid anchor sequence or version index |

The website has its own wording for a failed switch: `That version is no longer available. Reload the chat and try again.` Both mean the same thing and both are fixed by reloading the conversation.

`Editing a message is not supported on this backend.` is thrown when an edit-and-resend would route to the legacy backend. The Edit control is gated so users should not reach it.

Key points:

- Why there is no regenerate button, and what replaces it
- What a version actually contains, and which history an edit throws away
- What re-running a turn costs, and when a turn is refunded

Checklist:

- No regenerate control exists in Whizi, and the worker carries no route behind one
- Edit and resend is the re-run: the reply is replaced, not repeated
- The Edit control appears on your own messages only
- A version is the message plus every reply after it, so switching swaps the whole tail
- Editing an earlier turn discards the version history of every turn after it
- Editing costs nothing; the resend is charged at the answering model rate
- A stopped generation is refunded as an undelivered turn
- A version switch takes the generation lock, so it cannot run while a reply is generating

Steps:

1. Find the message you sent: The Edit control appears on your own messages only. Assistant replies do not carry one.
2. Edit the prompt: Change the wording, add the constraint the first answer missed, or leave it and switch the model instead.
3. Send it again: The reply is replaced rather than appended, so the conversation does not fill up with near-duplicate answers.
4. Flip between versions: Arrows under the turn read like 2/2 and move between the versions of it, tail of the conversation included.

FAQ:

Q: How do I regenerate a response in Whizi?
A: You cannot, in the sense of a button that re-rolls the same prompt. The equivalent is to edit your own message and send it again, which replaces the reply rather than appending a second one. Resending the words unchanged is not a workaround either: the same text sent to the same conversation recently comes back as HTTP 409 `duplicate_message`, so a changed prompt or a different model is the only way to a second attempt.

Q: Can I edit a message I already sent?
A: Yes, on messages you sent. Assistant replies carry no Edit control, so there is nothing to press on the answer itself, and the reply under an edited message is replaced rather than joined by a second one. The same edit path exists on the creative (media) route, where a stale message id fails with the same 404 a text edit gets. What a version holds, and which history an edit throws away, is covered under what an edit actually changes.

Q: Does re-running a turn cost credits?
A: Editing the text costs nothing, since the charge lands only when a model answers. The resend is one ordinary turn at the fixed per-message rate of whichever model answers it, with no separate re-run price, and it is floored at one credit. If the remaining balance does not cover that turn, it is refused whole rather than part-charged.

Q: Do I lose the original answer when I edit?
A: Not for that turn. The previous attempt is kept and the arrows move between the two. Lower down the thread you do: an edit drops the stored versions of every turn beneath it. Work backwards if the branches matter, meaning edit the earliest message last, and copy anything from a lower branch you want to keep before you touch a message above it.

Q: What does "That version is no longer available. Reload and try again." mean?
A: HTTP 404 with code `version_not_found`. You moved to a message version that has been archived away, so there is nothing left to show. Reload the conversation and the arrows rebuild from what is actually stored. If the string you got says "Reload the chat" instead, that is the website toast for the same failure and it means the same thing.

Q: Can I get a different answer without editing anything?
A: Yes, by sending it on a different model in the same conversation. The thread comes with you and the turn is charged at the new model rate. The gate to watch is your plan: picking a model above your tier returns HTTP 403 `tier_upgrade_required` with "Upgrade your plan to use this model." instead of an answer.

### How to reset your Whizi password

URL: https://whizi.io/docs/reset-your-password/
Updated: August 2026

Quick answer: The mobile app is the route that replaces a forgotten Whizi password: select Forgot password? on its sign-in screen and Whizi emails a 6-digit code you enter with the new password. On the website, Forgot password? emails a sign-in link aimed at a change form whose first field is Current Password.

Where the reset control lives on the website and in the app, what Whizi emails you in each case, why starting from Settings signs you out first, and what a reset does not fix.

#### The short answer

The mobile app is the route that actually replaces a forgotten password. Select Forgot password? on its sign-in screen, type the address on the account, and Whizi emails a 6-digit code that you enter on the next screen along with the new password. The website route is a different thing wearing the same label: Forgot password? there asks the provider for an `email_link` sign-in email pointed at whizi.io/reset-password, and `/reset-password` is a change form whose first field is Current Password. So the website side replaces a password you still know. If the website password is genuinely gone, the app route or support is the way back, not the emailed link.

There are four places the reset can start, and they do not behave the same way:

| Where you start | The control | What gets sent | Signed out first |
| --- | --- | --- | --- |
| Website sign-in page | `Forgot password?`, on the Password label row | An `email_link` sign-in email aimed at `/reset-password` | No |
| Website Settings, Account section | `Reset Password` | The same email link, to the same address | Yes, the moment you select Send Reset Link |
| Mobile app sign-in screen | `Forgot password?` | A 6-digit code | You are already signed out there |
| Mobile app Settings | Reset password, behind a confirmation | A 6-digit code | Yes, after you confirm |

The one to know before you use it is the second row. It signs you out of Whizi on the spot, and it does that before it checks whether you typed anything in the email field.

#### Whizi runs its own reset screens

Authentication is provided by Clerk, but the screens are not Clerk hosted pages. The reset dialog on the sign-in page, the change form at `/reset-password`, and the reset steps inside the mobile sign-in screen are all Whizi screens that call the provider underneath, so the form you type into is Whizi code rather than a provider-hosted page.

That split decides which wording you get when something fails. Every one of those surfaces prefers the provider message when there is one, and falls back to a Whizi string when there is not. So a refusal about password length or strength arrives in the provider wording, quoted as it came. The fallback differs by surface, which is the fastest way to tell in a support thread where a failure happened:

| Surface | Fallback when no readable message came back |
| --- | --- |
| Website reset dialogs, sign-in page and Settings | `Failed to send reset email. Please try again.` |
| Website change form at `/reset-password` | `Failed to reset password. Please try again.` |
| Both reset steps in the mobile sign-in screen | `Couldn't sign in. Please try again.` |

The checks Whizi writes for itself are thin and unevenly spread. The website change form requires the current password field to be filled and the two new entries to match. The mobile reset step performs none: it has one new-password field and sends its contents as typed. The only other Whizi check anywhere in the flow is that the email field is not empty, on both website dialogs and on the app's first reset step. There is no minimum length, character rule or reuse rule written into Whizi at all. Those live at the provider and reach you through its quoted messages.

The full table of sign-in, session and reset error strings, with the trigger for each, is on [when you cannot sign in](https://whizi.io/docs/login-and-account-access). This page covers the flow, that page covers the wording.

#### Starting from Settings signs you out first

On the website, Settings has a `Reset Password` row in the Account section, next to Subscription and Cancel Subscription. It opens a dialog with an Email Address field and a `Send Reset Link` button.

Selecting that button signs you out of Whizi before it does anything else. The sign-out runs first, then the code checks whether the email field was filled. Submit it empty and you are signed out with nothing sent, and the only way back in is the sign-in page.

The field is also blank when the dialog opens. It is not filled in from the address on the account you are signed in to, so have the address to hand before you open it rather than after you have been signed out.

Success reads differently depending on where you started, which is a quick way to tell the two routes apart in a support thread:

| Surface | Success message |
| --- | --- |
| Sign-in page dialog | `Check your inbox for a reset link.` |
| Settings dialog | `Password reset email sent successfully!` |

The mobile app is built on the same sign-out-first design but tells you first. Selecting reset there raises a confirmation titled "Reset your password?" whose body says Whizi will sign you out and email a 6-digit code to the account address, naming that address in the sentence, then asks you to enter it on the next screen to set a new password. The buttons are Cancel and "Sign out and reset". It also carries your address across the sign-out, so the reset step opens already filled in. The website does neither of those things.

#### The mobile app sends a code, not a link

The app reset is a two-step flow on the sign-in screen and it ends with you signed in, which makes it the more predictable of the two routes.

1. The first step asks for the address. Submit it empty and the screen answers `Enter the email address on your account.`

2. Whizi requests the code, then confirms with `We sent a 6-digit code to` followed by the address you typed.

3. The second step takes the code and the new password together. If it succeeds, the app activates the new session and drops you into Whizi. There is no separate sign-in afterwards.

One account type cannot finish there. If the account has a second verification factor, the app has no screen for it and answers: "This account needs another verification step. Contact support and we'll help you back in." That is the real instruction, not a retry prompt.

Everything else the app shares with the website, including the plan and the conversation history you come back to, is on [Whizi on mobile](https://whizi.io/docs/whizi-on-mobile).

#### Changing a password you already know

`/reset-password` on the website is a change form rather than a recovery form. It has three fields, Current Password, New Password and Confirm New Password, and it refuses to run without the first one: `Current password is required.` when it is empty, and `Passwords do not match.` when the last two differ.

A successful change ends every other session on the account. The form asks the provider to sign out of other sessions as part of the update, so any other browser or device holding a Whizi session is signed out by the change, not merely left alone. It is silent: nothing on the form announces it, and no message is produced by it, so the only strings this screen can show are the two validation errors, the provider's own refusal text, and the success line below. Worth knowing before you change a password on a shared machine.

On success the page shows `Password reset successful! Redirecting to chat...` and moves you to the chat page after a two second pause.

Typing that URL in the address bar is not a way to reach it. The route sits behind the guard that sends signed-in visitors to the chat page, and `/reset-password` is not one of the four paths exempt from that guard (settings, pricing, terms of service and privacy policy are). Visited while signed out it renders `User not authenticated. Please log in.` instead of a form. The sign-in page dialog and the Settings dialog both send the same email link to that same address, so neither is a way around the Current Password field either.

#### What a password reset does not fix

A new password only replaces the credential. Four things people expect it to solve, it does not:

| What you want | Does a reset do it |
| --- | --- |
| Clear a session token error such as `Invalid or expired session token.` | No. That is token verification failing, covered on [when you cannot sign in](https://whizi.io/docs/login-and-account-access) |
| Change the email address you sign in with | No. There is no self-serve control for the address at all. See [change the email on your account](https://whizi.io/docs/switch-account-email) |
| Restore free messages you have used | No. The free tier allowance is 7 messages for the life of the account, with no reset path |
| Move a subscription onto a different account | No. The plan follows the account, not the password |

If the account address itself is one you can no longer read, the reset route is gone with it, because both the link and the code arrive in that inbox. Raise that with support while the old inbox still works rather than after.

Key points:

- The website emails a sign-in link aimed at the change form, the mobile app emails a 6-digit code, and both are reached from Forgot password? on the sign-in screen
- Starting the reset from Settings signs you out first, before the email is even sent
- Changing a password you already know signs out every other session on the account

Checklist:

- Forgot password? sits on the Password label row of the website sign-in page and opens a Reset password dialog
- In the app it is a right-aligned link below the email and password fields, which carry placeholders and no labels
- The website emails a sign-in link aimed at /reset-password, the mobile app emails a 6-digit code
- Settings, Reset Password signs you out before it sends, and before it checks the email field is filled
- The website reset field is never prefilled, so have the account address ready
- The mobile app warns first, prefills the address, and signs you in when the reset completes
- The website change form checks the current password field is filled and the new entries match, the mobile reset step validates nothing at all, and every length or strength rule comes from the auth provider
- Changing a password you know signs out every other session on the account
- A reset does not clear session token errors, change your email address, or return used free messages

Steps:

1. Open the sign-in screen: On the website, go to the Whizi sign-in page. In the mobile app, open the sign-in screen. Both carry the reset control, and neither needs you to be signed in already.
2. Select Forgot password?: On the website the link sits on the same row as the Password label, to the right of it. It opens a small dialog titled Reset password with one Email field.
3. Enter the address on the account and send: The website field is blank and is not prefilled from anything you typed on the sign-in form, so type the address in full, then select Send reset link. On success the dialog reads: Check your inbox for a reset link.
4. Open the email: On the website Whizi asks its authentication provider to email a link pointing back at whizi.io/reset-password. In the mobile app it is a 6-digit code instead, and the screen confirms with We sent a 6-digit code to followed by the address you typed.
5. Set the new password: In the app, type the code and the new password on the next screen, and a successful reset signs you straight in. The app checks nothing before it sends: there is one new-password field, and whatever is in it goes to the provider as typed. On the website change form Whizi checks two things of its own, that the current password field is filled and that the two new entries match. Every rule about length and strength comes from the authentication provider, and its wording is shown to you word for word.

FAQ:

Q: Where is Forgot password on Whizi?
A: On the website it sits on the sign-in screen, on the same row as the Password label and to the right of it, and opens a dialog titled Reset password with a single Email field and a Send reset link button. The mobile app places it differently: the sign-in fields there carry placeholders and no labels at all, so Forgot password? is a right-aligned link on its own row below the email and password fields. You do not need to be signed in on either, and there is no separate recovery URL to find.

Q: Does being signed out by the Settings reset mean it failed?
A: No. The sign-out is the first thing that route does, before the email is even attempted, so it happens whether the send succeeds or not. `Password reset email sent successfully!` in the dialog is the confirmation that it did. If that line never appears, sign back in from the sign-in page and start the reset there instead.

Q: Does resetting my password sign me out on my other devices?
A: Yes, when you change it from the change form at /reset-password. That update asks the authentication provider to sign out of other sessions as part of the change, so other browsers and devices holding a Whizi session lose it. Nothing warns you first and nothing reports it afterwards, so check what else is signed in before you change a password rather than after.

Q: I sign in with Google. What password am I resetting?
A: Whizi does not branch on how the account was created: the reset dialog takes an email address, asks the provider to send the reset, and shows whatever the provider answers word for word. If Continue with Google has always been your way in, that button is on the sign-in page and needs no password at all, which is the faster route back. If the provider refuses the reset, the message it returns is the accurate explanation and is worth quoting to support verbatim.

Q: Does a password reset give me my free messages back?
A: No. The free tier allowance is 7 messages counted for the life of the account, with no reset path, so nothing about credentials touches it. What the app says when you reach the end of it is on out of credits, and the paid allowances are on plans and limits.

### How to search your chat history in Whizi

URL: https://whizi.io/docs/search-your-chat-history/
Updated: August 2026

Quick answer: Yes, Whizi has chat search, and it searches chat titles only. The sidebar carries a search box over the chat list, and one search covers your projects and your ungrouped chats together. Message bodies are not searched, so the only route to words inside a message is the account data export.

Whizi searches chat titles and nothing else. What that means in practice, how to name chats so you can find them again, and what to do when the words you remember are inside a message.

#### The short answer

Yes, Whizi has chat search, and it searches chat titles only. The sidebar carries a search box over the chat list, and what you type is matched against the names of your chats. Message bodies are not searched, so a phrase you remember from inside a conversation will not find that conversation.

What the search box does and does not reach:

| What you might search on | Matched |
| --- | --- |
| The chat title | Yes |
| Text inside a message | No |
| Text inside an uploaded file | No |
| Chats that sit inside a project | Yes, projects and the ungrouped list are searched together |

That single design fact decides everything else on this page. Finding an old conversation in Whizi is a naming problem rather than a search problem.

#### One search covers every project at once

Searching cuts across projects and the ungrouped list, so a chat is findable by name whichever folder it sits in. You do not have to remember which project you filed something under, or open each one in turn.

That is worth knowing because projects are otherwise a hard partition: a chat belongs to at most one project, which is what makes the "everything else" view in the sidebar work. Search is the one place where that partition is ignored. Renaming, deleting and the project structure around it are covered in [organize your conversations](https://whizi.io/docs/organize-conversations).

There is no backend search route either. No endpoint in the worker accepts a search query over chats or messages, so the filtering you see is over the chat list in the sidebar rather than a query run across your whole account server side.

#### Rename the chats you will want back

Because the title is the only searched field, renaming is the lever. A chat called something you would actually type is a chat you will find; a chat whose title describes the first message rather than the work is one you will scroll past.

The rules the rename enforces, in the order you hit them:

- A blank name is refused with `Chat name cannot be empty.`
- Over 100 characters is refused in the browser with `Chat name is too long (max 100 characters).`
- The backend cap on the same field is 120 characters, and it answers `Conversation titles must be between 1 and 120 characters.`
- If the call itself fails you get `Failed to rename chat. Please try again.`

Two habits do most of the work. Put the nouns you would search for into the title, meaning the client, the file, the feature or the decision, not the verb. And rename at the end of a session rather than the start, because that is the point at which you know what the conversation turned out to be about.

#### Imported chats keep the titles the old tool gave them

Whizi imports from ChatGPT and Claude, and an imported conversation carries its original title across, trimmed to 120 characters. A conversation with no usable title in the export becomes `Imported chat`.

For search that has one consequence worth planning for: a large import can leave a run of rows all named `Imported chat`, and title search cannot separate them from each other. The fix is the same as everywhere else on this page, which is to rename the handful you expect to want again rather than all of them.

Imports are also idempotent, keyed by account plus source plus source id, so re-running the same export does not duplicate the rows you have already renamed. If the import itself is failing, the message you are seeing is covered in [chat import errors](https://whizi.io/docs/chat-import-errors).

#### When the words you remember are inside a message

There is no full text search over message bodies in the product. The honest answer is that Whizi does not offer it, and no combination of settings turns it on.

The one route that gets you there is the account export. Export Data sits in Settings under Account and returns your account as a single downloadable JSON file that covers chats and messages, along with memory, projects and file metadata, media, share links, subscription and referral profile. Open that file in any editor and search it there. Export is rate limited to 5 requests per minute, because it reads everything the account owns. The mechanics are in [exporting your chats and files](https://whizi.io/docs/export-chats-and-files).

The cheaper habit is prevention. An account can hold up to 50 projects, and a project narrows the chat list before you search it, so filing work that will recur into a project of its own does more for findability than any wording you could put in a title.

Key points:

- Search matches the chat title only, so message text never brings a chat back
- One search covers your projects and your ungrouped chats at the same time
- The one route to finding words inside a message, and the exact rename limits

Checklist:

- Search matches the chat title only, never message text
- One search covers projects and ungrouped chats together
- There is no server side search route, so the filter runs over the chat list
- Rename a chat to the words you would actually type to find it
- Renames are capped at 100 characters in the browser and 120 on the backend
- Imported chats keep their original title, and untitled ones become Imported chat
- For words inside a message, export the account JSON and search that file
- Up to 50 projects per account, and a project narrows the list before you search

FAQ:

Q: Can I search my Whizi chat history?
A: Yes, from the search box over the chat list in the sidebar, and it matches chat titles only. Searching cuts across projects and the ungrouped list at the same time, so you do not need to remember which project a chat was filed under. What it will not do is match text inside the messages themselves.

Q: Can I search inside my messages?
A: No. Search matches the chat title only, message bodies are not searched, and there is no backend search route that accepts a query over messages. The one way to search your message text is to take the account export from Export Data in Settings under Account, which returns chats and messages as one JSON file, and search that file in an editor.

Q: Why can I not find a chat I know exists?
A: Almost always because the words you are typing are in the conversation rather than in its title. Title search is the only search there is, so a chat whose name does not contain your search term will not appear however much of the term is in the transcript. Rename the chat to the words you would actually reach for, and it becomes findable from then on.

Q: Can I search my imported ChatGPT and Claude conversations?
A: Yes, on the same terms as everything else, which means by title. Imported conversations carry their original title across, trimmed to 120 characters, and any conversation with no usable title in the export becomes `Imported chat`. A big import can therefore produce several rows with that same name, and title search cannot tell them apart until you rename them.

Q: Does searching find chats inside a project?
A: Yes. Searching cuts across projects and the ungrouped list, so a chat is findable by name whichever folder it sits in. This is the one place where the project partition is ignored: everywhere else a chat belongs to at most one project, which is what makes the sidebar list projects separately from everything else.

### How to share an AI conversation link from Whizi

URL: https://whizi.io/docs/share-a-conversation/
Updated: August 2026

Quick answer: Sharing creates a public link to one conversation. Anyone with the link can read it in a browser without a Whizi account, and they cannot continue it, edit it, or see anything else in your workspace. Sharing costs no credits and works on every plan, and revoking a link breaks it immediately for everyone.

Create a public link to a conversation, what the recipient can and cannot see, and how to revoke access once you are done.

#### The short answer

Sharing creates a public link to one conversation. Anyone with the link can read it in a browser without a Whizi account, and they cannot continue it, edit it, or see anything else in your workspace.

Sharing costs no credits and is available on every plan. You can revoke a link at any time, which immediately breaks it for everyone who has it.

#### What the recipient sees

A shared link shows the conversation as it stood when you shared it: the prompts, the answers, and which model produced each answer. That last detail matters more than it first appears, because it is what makes a shared conversation useful as evidence rather than just as a transcript.

What a recipient does not get: the ability to send messages into your conversation, access to any other chat, anything about your account, your plan, or your usage. The link is scoped to one conversation and is read-only.

| What the link carries | Visible to a recipient |
| --- | --- |
| The prompts and the answers | Yes |
| Which model produced each answer | Yes |
| Text extracted from an uploaded file, where the thread contains it | Yes, it reads like any other message |
| The ability to send, edit or continue the thread | No |
| Any other conversation in your workspace | No |
| Your account, your plan or your usage | No |

Recipients need no account and no app. The link opens in a browser like an ordinary web page, which is the point: sharing an AI conversation with someone who does not use AI tools is one of the more common reasons to do it.

#### What sharing is actually good for

Three uses that come up repeatedly.

**Showing your working.** When an AI answer informed a decision, sending the conversation is more honest and more useful than pasting the conclusion. The reader can see what was asked, which model answered, and where the reasoning was thin.

**Getting a second opinion on a prompt.** If a model keeps producing something unhelpful, sharing the thread lets someone else see what you actually asked, which is usually where the problem is. Sharing gives the recipient a read-only view rather than a copy they can continue, and why there is no copy is covered in [can you fork a conversation](https://whizi.io/docs/fork-a-conversation).

**Comparing models in public.** A conversation where the same question went to Claude and to GPT, with both answers visible and labelled, is a far stronger argument than describing the difference. This is the shape that tends to travel.

#### Revoking and what to check first

Revoke from the same dialog you used to share. Once revoked, the link stops working for everyone, including people who already opened it. There is no partial revocation and no per-person access, because the link is the credential.

That is worth stating plainly, because it decides how a link behaves once it leaves your hands. Access travels with the URL: anyone it is forwarded to reads the conversation on the same terms as the person you sent it to, without an account and without being named anywhere. So the only control you keep after sending is revoking, and revoking is all or nothing.

Two things worth checking before you create a link. First, whether the conversation contains anything from an uploaded file that should not leave your organisation, since extracted document text appears in the thread like any other message. Second, whether any earlier turn in the conversation went somewhere you have forgotten about, which is common in long threads that changed subject.

If in doubt, start a fresh conversation with just the exchange you want to show and share that instead. It takes a minute and removes the question entirely.

Key points:

- What a recipient sees, and what stays private
- Why sharing is read-only, and what that prevents
- How to revoke a link you have already sent

Checklist:

- Sharing creates a public read-only link to one conversation
- Recipients need no Whizi account
- The link shows which model produced each answer
- Recipients cannot continue the chat or see anything else
- Sharing costs no credits and works on every plan
- Revoking breaks the link immediately for everyone
- Check for uploaded document text before sharing a work thread

Steps:

1. Open the conversation: Go to the chat you want to share.
2. Open the share dialog: Use the share control in the chat header.
3. Create the link: Whizi generates a public URL for this conversation.
4. Send it: Anyone with the link can read the conversation without an account.
5. Revoke when done: Return to the same dialog and revoke the link to cut off access.

FAQ:

Q: Do people need a Whizi account to open a shared link?
A: No. A shared conversation opens in any browser like an ordinary web page, with no account, no login and no app required. This is deliberate, because a large share of sharing is aimed at people who do not use AI tools themselves and would not create an account to read one conversation.

Q: Can someone continue a conversation I shared with them?
A: No, shared links are read-only. A recipient can read every message and see which model produced each answer, but cannot send anything into the thread, edit it, or reach any other part of your workspace. If you want a genuine back and forth, copy the relevant prompt to them rather than sharing the thread.

Q: How do I unshare a conversation?
A: Open the share dialog on that conversation again and revoke the link. It stops working immediately for everyone who has it, including people who already opened it once. Because the link itself is the credential, there is no way to revoke it for one person and not another.

Q: Does a shared link update if I keep chatting?
A: Treat a shared link as a snapshot of what you meant to show rather than a live window onto an ongoing thread. If a conversation is going to keep moving and you only wanted to show one exchange, the cleaner approach is to start a fresh conversation containing just that exchange and share it instead.

### You paid for Whizi but the plan features are still locked

URL: https://whizi.io/docs/subscription-not-active/
Updated: August 2026

Quick answer: If you paid and Whizi still refuses you, the status code decides the fix. HTTP 402 means the account you are signed into has no active subscription, so the payment is on a different account or never completed. HTTP 403 with a tier code means the subscription is active and the request sits above that tier.

Why a paid Whizi account can still behave like a free one, how the 402 and 403 refusals differ, and what each error string means.

#### The short answer

If you paid and Whizi still refuses you, the status code behind the message decides the fix. HTTP 402 means the account you are signed into has no active subscription on it, so the payment is either on a different account or never completed. The two tier codes, `tier_upgrade_required` and `powerhouse_required`, arrive as HTTP 403 and mean a subscription is active and the request sits above that tier, which is a plan question rather than a billing failure. Other 403 codes exist that are neither billing nor tier, `account_sharing_limit` among them, so the code matters as much as the status.

| What you see | Code | What it means | Group |
| --- | --- | --- | --- |
| `Choose a subscription to use Whizi.` | 402 `subscription_required` | The subscriber guard found no paid plan on this account | No subscription |
| `Your free messages are used up. Start a subscription to keep chatting.` | 402 `free_limit_reached` | The free-tier lifetime message counter is already at its cap | No subscription |
| `Choose a subscription to generate images.` | 402 `subscription_required` | A free account asked for image generation | No subscription |
| `Upgrade your subscription to talk with Whizi in voice mode.` | 402 `subscription_required` | A free account started a voice session | No subscription |
| `Upgrade your plan to use this model.` | 403 `tier_upgrade_required` | A paid account picked a text model its tier does not include | Wrong tier |
| `This model is only available on the Powerhouse plan.` | 403 `powerhouse_required` | A creative model on the Powerhouse-only list was requested from a lower tier | Wrong tier |
| `Upgrade to Pro or Powerhouse to generate images.` | 403 `tier_upgrade_required` | A paid account whose tier has no image allowance asked for an image | Wrong tier |
| `Account sharing limit reached. Please use your own Whizi account.` | 403 `account_sharing_limit` | More than 3 distinct device or session keys were seen on this account in the last 24 hours | Neither |

Only the first group is a "I paid and nothing works" problem. The rest mean the subscription is doing its job and the request was above what it buys.

#### Why the upgrade popup is not evidence of anything

On the website, a send that comes back with any HTTP 402 opens the upgrade modal rather than a toast. The badge on that modal reads `Monthly Limit Reached`, so it says the same thing whether you actually exhausted a monthly allowance or the server simply found no subscription on the account. Do not read the badge as a diagnosis.

The line underneath is more useful, because it is chosen from the account state rather than from the error. A free account gets `Choose a plan to unlock Whizi with higher limits and premium models.` A paid account gets `You have reached your monthly limit of {limit} credits on the {plan} plan. Upgrade now to continue using Whizi without interruption.`, or the shorter `You've reached your monthly limit on the {plan} plan. Upgrade now to continue using Whizi without interruption.`, with your real plan name filled in.

A modal with no plan name in it means this browser is signed into an account with no subscription on it.

#### A subscription belongs to one account, and only that one

Billing in Whizi is per individual account, and the tier is derived in exactly one place from that user's own subscription. There is no team, organization, workspace or seat concept, so a plan cannot be spread across two logins and one person paying does not lift anyone else.

One cause is that the checkout ran under one email address and the session you are using is signed in as another. Compare the address on the payment receipt with the address shown in account settings before assuming anything is broken.

The check that settles it: open account settings and start a cancellation. If there is no Stripe customer on file, or the customer has no active subscription, the cancel route answers `Subscription is already canceled. Your account is on the free plan.` That string is the server telling you plainly that this account has nothing attached to it. If instead you get `Subscription cancellation scheduled for end of billing period.` then the subscription is here and you have just cancelled it, so re-subscribe or contact support.

Two adjacent strings from the same screen are not the answer to this question, so do not over-read them: `Authentication failed. Please sign in again.` means the session token could not be fetched, and `Unable to retrieve user or subscription details. Please try again.` means the user id or Stripe customer id was missing on that request.

#### Checkout that looked finished but was not

A subscription is created on Whizi's side from the completed Stripe checkout session. If that session never completed, or completed without the metadata the worker needs, no subscription record exists no matter what the card statement looks like.

One string on that path is a refusal rather than a retry case: `Checkout session is missing customer, plan, or billing cycle for subscription creation.` It is raised inside webhook handling when a completed session lacks the metadata needed to create the subscription. You cannot see or retry it. Nothing in the interface reports it, and no subscription record is created. If your card was charged and settings still says you are on the free plan, this is the case to hand to support.

The other refusal on the checkout path, the intro trial having already been used with this payment method, ends before any charge, so it belongs to a different symptom. It is covered in [payment declined](https://whizi.io/docs/payment-declined).

#### A 403 means the plan is working and the tier is the limit

If the message is `Upgrade your plan to use this model.` then your subscription is active and the server accepted it. The model you picked simply sits above your tier. Access is resolved per request on the server from the subscription on your own account, so reloading the page, clearing the cache or signing out and back in will not change the answer.

Which models your tier includes is a mapping question rather than a billing one, and it is listed per model in [the model reference](https://whizi.io/docs/models), and family by family in the model guides, [Grok](https://whizi.io/docs/using-grok-in-whizi) and [Qwen](https://whizi.io/docs/using-qwen-in-whizi) among them.

The workaround while you decide: switch to Auto. Auto is available on every paid plan, and it only ever picks from models your plan can already open, walking back down the ladder rather than stopping at a paywall. The hub version of this symptom is in [troubleshooting](https://whizi.io/docs/troubleshooting).

Agent personas are a separate gate of the same kind: they resolve to a minimum tier of Pro by identifier, so a persona locked on Starter is behaving as designed.

#### The notification that says your subscription is no longer active

`Your subscription is no longer active.` is a push notification body, sent under the title `Whizi subscription updated`. It fires when a RevenueCat webhook syncs the account down to the free tier.

That is a report of a sync, not a bug: the notification goes out at the moment the account is moved onto the free tier. If you did not cancel and you believe it is still paid, [billing and cancellation](https://whizi.io/docs/billing-and-cancellation) is where to pick it up.

One related refusal, if you subscribed as a guest and the account had not finished linking: `The guest store identity is not linked yet. Reopen Whizi and try again.`, HTTP 409 `guest_store_identity_missing`. Here the instruction in the string is the actual fix. Reopen Whizi so the identity binding commits, then let the subscription sync run again.

#### What support has to fix

There is no self-serve way to attach a payment to an account it did not land on. If the cancel route reports `Subscription is already canceled. Your account is on the free plan.` on the account you are actually using, and your statement shows a charge, the record has to be fixed on Whizi's side. Entitlements can be granted by hand, but that is an operator action, not a setting you can reach.

The pair of email addresses, the one you are signed in with and the one on the payment receipt, is what support checks first, because those are often two different accounts. The full list of what to send is in the FAQ below.

What will not help: refreshing, reinstalling, clearing cookies, or sending the message again, for the reason given in the 403 section above. Nothing on the client decides this.

Key points:

- HTTP 402 and HTTP 403 mean two different problems with two different fixes
- Every subscription and tier refusal string, quoted verbatim with its code
- The one check that tells you whether the payment reached the account you are using

Checklist:

- HTTP 402 means no active subscription on this account
- `tier_upgrade_required` and `powerhouse_required` are HTTP 403 and mean the subscription is active and the tier is the limit
- The `Monthly Limit Reached` badge opens on any 402, so it is not a diagnosis
- A modal with no plan name in it means this account is on the free plan
- A subscription belongs to one account only, with no team or seat sharing
- `Subscription is already canceled. Your account is on the free plan.` confirms nothing is attached here
- A charged card plus a free-plan account is a support case, not a retry case

FAQ:

Q: I paid for Whizi but it still says to choose a subscription. What is wrong?
A: `Choose a subscription to use Whizi.` is HTTP 402 `subscription_required`, which the subscriber guard throws when the account making the request has no paid plan on it. One cause is that the payment landed on a different account, because billing is per individual account and the tier is read from that user's own subscription. Check which email you are signed in with against the email on the receipt. If they match, checkout may not have completed, in which case support has to look at the record.

Q: What is the difference between "Choose a subscription to use Whizi." and "Upgrade your plan to use this model."?
A: The first is HTTP 402 and means no subscription is attached to this account at all. The second is HTTP 403 `tier_upgrade_required` and means a subscription is attached, it was accepted, and the model you picked sits above the tier you bought. The first is a billing problem, the second is a plan choice. Switching to Auto works around the second one immediately, because Auto only picks models your plan can already open.

Q: Does signing out and back in refresh my plan?
A: No, for the reason set out in the 403 section above. What has to change is the subscription record on the account. If you just completed checkout and the account still reads as free, that is a support case rather than something a reload can clear.

Q: Why does the upgrade popup say "Monthly Limit Reached" when I never hit a limit?
A: Because the website opens that same modal for any HTTP 402 response, and the badge text is fixed. The line beneath it is the informative part: a free-plan account is offered `Choose a plan to unlock Whizi with higher limits and premium models.`, while a paid account sees its own plan name interpolated into the monthly-limit wording. No plan name means no plan on this account.

Q: I was charged but my account is on the free plan. What do I send support?
A: Four things: the email address you are signed in with, the email address on the payment receipt, the verbatim error string in front of you, and the status code if you can see it. Send that rather than retrying, since the webhook metadata failure described above leaves nothing for a retry to find.

Q: Can two people share one Whizi subscription?
A: Sharing one login trips an anti-abuse gate. More than 3 distinct device or session keys seen on one account within 24 hours returns `Account sharing limit reached. Please use your own Whizi account.`, HTTP 403 `account_sharing_limit`, and that is neither a billing refusal nor a tier refusal. So no, and there is no seat concept to buy instead.

### Whizi supported file types: what you can upload

URL: https://whizi.io/docs/supported-file-types/
Updated: August 2026

Quick answer: Whizi accepts images, PDFs, Word documents, Excel spreadsheets, PowerPoint decks, and around forty plain text and source code formats. File upload is included on every plan, including Starter, and uploading a file costs no credits. PDF, Word, Excel and PowerPoint are extracted to text in your browser before anything is sent.

The complete list of accepted file extensions, what happens to each type once uploaded, and how long files are kept.

#### The short answer

Whizi accepts images, PDFs, Word documents, Excel spreadsheets, PowerPoint decks, and around forty plain text and source code formats. File upload is included on every plan, including Starter, and uploading a file costs no credits.

The four groups, and what happens to each:

| Group | Extensions | What Whizi does |
| --- | --- | --- |
| Images | .png, .jpg, .jpeg | Passed to the model as an image, so the model sees it directly |
| Documents | .pdf, .docx, .doc | Text extracted in your browser, then sent as text |
| Spreadsheets | .xlsx, .xls | Cell contents extracted and sent as structured text |
| Presentations | .pptx | Slide text, tables and speaker notes extracted, slide by slide |
| Text and code | 40 plus extensions, listed below | Sent as text as-is |

#### The complete extension list

**Images.** `.png`, `.jpg`, `.jpeg`

**Documents, spreadsheets and decks.** `.pdf`, `.docx`, `.doc`, `.xlsx`, `.xls`, `.pptx`

**Source code.** `.js`, `.ts`, `.jsx`, `.tsx`, `.py`, `.java`, `.cpp`, `.c`, `.cs`, `.rb`, `.php`, `.go`, `.rs`, `.swift`, `.kt`, `.kts`, `.scala`, `.lua`, `.pl`, `.pm`, `.r`, `.m`, `.sql`, `.graphql`

**Markup and styles.** `.html`, `.css`, `.scss`, `.sass`, `.less`, `.xml`, `.md`

**Config and data.** `.json`, `.yaml`, `.yml`, `.env`, `.ini`, `.conf`, `.csv`, `.txt`

**Shell scripts.** `.sh`, `.bash`, `.zsh`, `.fish`

If a format is not on this list, the fastest route is usually to export or save it as PDF, CSV or plain text first. A Google Doc becomes a `.docx` or a PDF, a Keynote deck becomes a `.pptx` or a PDF, and almost any data export becomes a `.csv`.

#### How extraction actually works

For PDF, Word, Excel and PowerPoint files, the text is extracted in your browser before anything is sent. The model receives the extracted text rather than the original binary file. This has two consequences worth knowing.

The first is that a scanned PDF with no text layer produces nothing useful. If a PDF is a photograph of a page rather than a document containing text, extraction returns empty and the model has nothing to work with. The workaround is to upload it as an image instead, since the vision-capable models can read a page image directly.

The second is that layout is partly lost. Extraction preserves the words and the reading order, not the visual arrangement, so a complex multi-column report or a heavily formatted table may arrive flatter than it looks on screen. For tables specifically, a `.csv` or `.xlsx` upload gives the model far cleaner structure than the same table inside a PDF.

Images are different: they are passed through to the model directly rather than extracted, so charts, diagrams, screenshots and handwriting all work as long as the model you selected can see images.

#### Retention and deletion

Uploaded attachments and generated media are configured to expire after up to 30 days. You can delete them sooner yourself, and account deletion removes them along with everything else.

Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models, and does not sell that content as training data. The full detail is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

Key points:

- Every accepted extension, grouped by what Whizi does with it
- Which formats are read as text and which are read as images
- How long an uploaded file is kept, and how to remove it sooner

Checklist:

- Images, PDF, Word, Excel, PowerPoint and 40 plus text and code formats are accepted
- File upload is on every plan, including Starter
- Uploading costs no credits
- PDF, Word, Excel and PowerPoint are extracted to text in your browser first
- Scanned PDFs with no text layer should be uploaded as images instead
- Tables work far better as .csv or .xlsx than inside a PDF
- Uploads expire after up to 30 days

FAQ:

Q: Can I upload a PDF to Whizi?
A: Yes, on every plan including Starter, and it costs no credits. The text is extracted in your browser and sent to whichever model you selected, so you can ask a PDF question of GPT, then switch the same conversation to Claude and ask again without re-uploading. A scanned PDF with no text layer is the one case that fails, and uploading those pages as images works instead.

Q: Can Whizi read Excel files?
A: Yes, both .xlsx and .xls. Cell contents are extracted and sent as structured text, which gives the model much cleaner data than the same table pasted into a message or embedded in a PDF. For anything analytical, a spreadsheet upload is the format that produces the most reliable answers.

Q: Is there a file size limit?
A: The practical constraint is the context window of the model you selected rather than a fixed upload cap. A very long document consumes context that the model then cannot use for anything else, so extremely large files are better split into the sections you actually want analysed. If a document is too long for the model you picked, switching to a larger-context model in the same conversation usually solves it.

Q: Can I upload more than one file at once?
A: Yes, and this is one of the more useful things to do with a multi-model workspace: upload a spreadsheet and the PDF report it came from together, and ask the model to reconcile them. Drag and drop works anywhere in the chat window, not just on the attach button.

Q: Are my uploaded files used to train AI models?
A: Whizi does not use your prompts, files, conversations, voice transcripts or generated content to train Whizi-owned AI models, and does not sell that content as training data. Uploads are configured to expire after up to 30 days, and you can delete them sooner or remove everything through account deletion.

### Changing the email address on your Whizi account

URL: https://whizi.io/docs/switch-account-email/
Updated: August 2026

Quick answer: No setting in Whizi changes the email address on an account. Settings offers Subscription, Cancel, Reset password, Export Data and Import Chats, and none of them edits the address you sign in with. The only published route is support, at whizihelp@gmail.com or whizi.io/support, and Whizi may verify identity first.

Settings has no email-change control. What it does offer, why the request has to go to support, and what a second account would cost you.

#### The short answer

Settings has no control that changes the email address on an account. Under Account it offers Subscription, Cancel, Reset password, Export Data and Import Chats. Under Preferences it offers Default model, Dark mode and Language, alongside the support and legal links. None of them edits the address you sign in with.

The only published route is to ask support, at whizihelp@gmail.com or through [whizi.io/support](https://whizi.io/support). The privacy policy frames that as a correction request: depending on where you live you may have rights to access, correct, delete, or receive a copy of personal information, and Whizi may need to verify identity before completing the request. No procedure is documented and no turnaround time is published.

What you can and cannot change yourself:

| What | Self-serve | Where |
| --- | --- | --- |
| Password | Yes | Reset password, in Settings, or the reset link on the login page |
| The email address on the account | No | No control for it in Settings. Ask support |
| Plan, or cancelling it | Yes | Settings, under Account |
| A copy of your data | Yes | Export Data, in Settings under Account |
| The whole account | Yes | Account deletion under Settings, with a two-step confirmation |

#### The Terms ask you to keep the address current anyway

Section 5 of the Terms of Service, on purchases and payment, puts an obligation on you that the product gives you no control for: "You further agree to promptly update account and payment information, including email address, payment method, and payment card expiration date, so that we can complete your transactions and contact you as needed."

A separate section, on user representations, adds that "If you provide any information that is untrue, inaccurate, not current, or incomplete, we have the right to suspend or terminate your account and refuse any and all current or future use of the Services (or any portion thereof)."

So an address you can no longer read is not a cosmetic problem to leave alone. Since there is no setting to fix it with, the practical move is to raise it with support while you still control the old inbox, because that inbox is what password recovery uses.

The error surface around signing in belongs to [when you cannot sign in](https://whizi.io/docs/login-and-account-access), including the password reset strings.

#### Making a second account does not move anything

The obvious workaround is to register again with the new address and carry your history across. A second account gets you none of your chats, none of your plan, and no second intro trial.

**Chats cannot be moved between Whizi accounts.** Export Data returns your account as one JSON file covering chats and messages, memory, projects and their file metadata, media, share links, subscription and referral profile. Import accepts exactly two sources: ChatGPT and Claude. A Whizi export is not a Whizi import format, so there is no supported path from one Whizi account into another. The export carries media as metadata plus a signed URL, and the bytes stay out of the file. Uploaded attachments and generated media are configured to expire after up to 30 days, so take your own copies of anything you want to keep.

**The subscription does not follow the address.** Billing is per individual account, and the tier is derived in exactly one place, from the subscription sitting on that account.

**The intro trial is bound to the card, not the account.** A second signup on the same payment method is refused with `A $0.99 intro trial has already been used with this payment method.`, HTTP 403 with code `duplicate_trial_payment_method`. The rest of the checkout failure surface is on [payment declined](https://whizi.io/docs/payment-declined).

**A new account starts on the free tier.** The allowance is 7 messages, counted for the life of the account with no reset path. Free accounts are not model-gated inside that allowance, so the count is the only limit. What the message says when you reach it is on [out of credits](https://whizi.io/docs/out-of-credits).

**The new address has to be a permanent one.** Disposable domains are refused at account creation, and the same refusal comes back from the checkout route when the signed-in account address is on the disposable list. Existing accounts are not blocked at sign-in, but a disposable address still blocks checkout. The string and both status codes are on [when you cannot sign in](https://whizi.io/docs/login-and-account-access).

**On mobile, guest data links once.** A guest token that resolves to a profile already claimed by a different account is refused, so device guest data cannot be re-linked into a second account after the first claim. The guest table on [when you cannot sign in](https://whizi.io/docs/login-and-account-access) carries that string and its code.

One further limit, from the Terms: if an account has been terminated or suspended, "you are prohibited from registering and creating a new account under your name, a fake or borrowed name, or the name of any third party, even if you may be acting on behalf of the third party."

#### What to do instead, by situation

| Situation | The route that exists |
| --- | --- |
| You want a different password | Reset password in Settings, or the reset link on the login page. See [reset your Whizi password](https://whizi.io/docs/reset-your-password) |
| You are losing access to a work address | Export your data and raise the change with support while the old inbox still works |
| Your subscription is on the wrong account | Compare the address on the payment receipt against the one you are signed in with, then see [subscription not active](https://whizi.io/docs/subscription-not-active) |
| You want to stop paying and start clean | Cancel first, confirm the cancellation, then decide about the account. See [billing and cancellation](https://whizi.io/docs/billing-and-cancellation) |
| You want the account and its contents gone | Account deletion under Settings, with a two-step confirmation |

Cancelling and deleting are different actions, and neither of them changes an address. What each one covers is set out on [billing and cancellation](https://whizi.io/docs/billing-and-cancellation) and [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

If you delete and register again, you are in the second-account case above: nothing carries over, because deletion is not a migration and no import path accepts a Whizi export.

Key points:

- No setting in Whizi changes the email address on an account
- What Settings does let you change
- Why a second account moves neither your chats nor your plan

Checklist:

- There is no setting in Whizi that changes your account email
- Nothing in Settings edits the sign-in address: not Subscription, Cancel, Reset password, Export Data, Import Chats, Default model, Dark mode or Language
- The only published route is support, at whizihelp@gmail.com or whizi.io/support
- The Terms still ask you to keep the address current
- Export Data first: it returns the account as one JSON file
- Import accepts only ChatGPT and Claude exports, so chats cannot move between Whizi accounts
- A subscription does not follow you, because billing is per individual account
- A new account starts on the free tier, with 7 messages for the life of the account

Steps:

1. Export your data first: Settings, under Account, has Export Data. It returns the account as one downloadable JSON file. Take the copy before anything changes.
2. Note whether a subscription is attached: Billing is per individual account, so record which plan the account holds and where it was purchased.
3. Send the request to support: The published contact route is whizihelp@gmail.com, or the support page at whizi.io/support.
4. Expect an identity check: The privacy policy states Whizi may need to verify identity before completing a request to access, correct, delete, or receive a copy of personal information.
5. Keep using the current address until you hear back: Nothing changes on the account by itself, and there is no published turnaround time to plan around.

FAQ:

Q: What happens to my subscription and my chats while a support request is outstanding?
A: Nothing moves on its own. Account information and synchronized chat history are retained while the account remains active, so the chats stay where they are, and the subscription stays on the account that bought it, because the tier is derived from the subscription sitting there. Cancelling is a separate decision you can still make in the meantime, and it has no bearing on the address either way.

Q: What should I do with the export file once I have downloaded it?
A: Keep it as your own copy. The export endpoint is the access counterpart to account deletion, and it works for guest sessions too, so a guest can take their chats and memory out before the device is gone. It is rate limited to 5 requests per minute, so take one clean copy instead of retrying. Quota counters, rate windows and push tokens are deliberately left out of it, so what you get is your content, not your bookkeeping.

Q: I am losing access to the email I signed up with. What should I do first?
A: Export your data while you still can, then raise the change with support before the old inbox goes away. Password recovery runs by email, so once you cannot read that inbox you have lost the self-serve recovery route as well as the address itself.

Q: Does changing my password change my email address?
A: No. They are separate. Reset password is available in Settings and from the login page, and it is the only credential you can change yourself. Its failure strings and every other sign-in error are collected on when you cannot sign in.

### How to switch AI models mid-conversation in Whizi

URL: https://whizi.io/docs/switch-models-mid-conversation/
Updated: August 2026

Quick answer: Yes. You can switch between GPT, Claude, Gemini and any other model inside one Whizi conversation, changing the active model per message from the picker. The whole thread carries across: your messages, the previous model's responses and your attached files. Each turn is charged at the credit cost of the model that answers it.

Change models inside a single chat without losing context. Use GPT to brainstorm, Claude to draft, and Gemini to verify, all in the same thread.

#### The short answer

Yes: you can use GPT, Claude, Gemini and any other model in the same conversation, switching between them per message from the model picker. The whole thread carries across the switch: your messages, the previous model's responses, and your attached files, so Claude can critique what GPT just wrote without you re-pasting anything.

Each turn is charged at the credit cost of the model that answers it, and the model must be included in your plan. The rest of this page is when switching pays off and when it does not.

#### What happens when you switch

The new model receives the full conversation so far, including your messages, the previous model's responses, and any files you attached. From its perspective it is joining a discussion already in progress, with everything visible.

That is why this is different from copying text between three browser tabs. When you paste into a new tool you bring the output but lose the reasoning, the constraints you established over five messages, and the document you uploaded. Switching in place keeps all of it.

Two things are worth knowing. The new model sees the previous model's output as part of the conversation, which can anchor it toward agreement. When you want a genuinely independent view, say so explicitly: `Evaluate the answer above on its merits. I want your assessment, not a synthesis of what has already been said.` And every model gets the same per turn budget of 40,000 input tokens, so a very long thread is trimmed to fit no matter which model you switch to; switching to a larger-context model does not send more of the thread. If a switch produces an answer that has forgotten the early part of the discussion, [the context budget](https://whizi.io/docs/context-too-long) is usually why.

#### The four patterns worth learning

**1. Phase routing.** The workflow most people converge on. Different stages of one task go to different models.

Structure in GPT, prose in Claude, verification in a web-connected model. It applies to writing a post, a proposal, a spec, or a report, and it works because these really are three different skills rather than three difficulty levels of the same one.

**2. The second opinion.** Get an answer, then switch and ask the new model to attack it.

`A colleague proposed the answer above. Find what is wrong with it: factual errors, missed edge cases, a simpler approach, or an assumption that does not hold. If it is actually sound, say so plainly rather than inventing objections.`

Two useful outcomes. Either a real flaw surfaces before you act on it, or the second model agrees despite being pushed to disagree, which is meaningful confirmation. Iterating with the same model produces neither, because models tend to agree with themselves.

**3. The reframe.** When an explanation does not land, switch rather than re-reading.

Claude and GPT explain the same concept with genuinely different framings, and the second framing is frequently the one that clicks. This is the single most underrated reason to have more than one model available, and it costs one click.

**4. Capability routing.** Switch because only one model can do the thing.

A 300 page document goes to Gemini because of context window. An image goes to a multimodal model. A question about last week goes to a web-connected model. This is not preference, it is a hard constraint, and it is the case where a single-model subscription simply stops.

#### Which model for which phase

A starting point rather than a rule. Your own tasks should override it within a week of paying attention.

| Phase | Model | Why |
| --- | --- | --- |
| Brainstorm, structure, outline | GPT | Fast, follows a brief tightly, clean hierarchies |
| Draft prose, tone, persuasion | Claude | Least filler, sustains a voice, best at difficult tone |
| Research, recent facts, sources | Gemini | Live web material with links you can open |
| Long documents, large corpora | Gemini | Largest context window |
| Strict formats, tables, JSON | GPT | Most reliable at obeying an exact schema |
| Critique and red teaming | Anything that did not write it | Independence is the whole point |

The table names the three families most people arrive with, and the picker is wider than that. The open weight families each have their own reference page, with the credit cost per message and the plan that opens each row: [Llama](https://whizi.io/docs/using-llama-in-whizi), [Mistral](https://whizi.io/docs/using-mistral-in-whizi), [Kimi](https://whizi.io/docs/using-kimi-in-whizi) and [GLM](https://whizi.io/docs/using-glm-in-whizi).

The one rule worth holding regardless of task: the model that critiques should never be the model that drafted.

#### Making the switch actually pay off

Switching helps most when you tell the new model what you want from it. Dropping in with no framing gets you a continuation of the same conversation in a slightly different voice.

**Handing off a phase**

`We now have the outline above. Draft section 2 only. Use the evidence listed under it. Voice: [paste sample]. Do not restate the outline back to me.`

**Asking for independence**

`Ignore the framing established above and answer this from scratch: [restate the question]. I want a genuinely independent take, not a refinement of what is already here.`

**Changing the job entirely**

`Stop drafting. Switch to reviewing. Here is the criterion: [state it]. Report only problems, do not rewrite.`

One further habit: when a switch produces something noticeably better, note which model and which phase. After a couple of weeks you will have a routing map that is specific to your work rather than to a benchmark, and that map is worth more than any published comparison.

#### When not to switch

Switching is cheap but not free, and there are cases where it makes things worse.

- **Mid-draft, for style.** Changing model halfway through a long piece produces a visible seam, because voice consistency is exactly what changes between models. Finish the draft, then switch for the critique.
- **When you have not defined the task.** Switching does not fix a vague prompt. If the first answer was unhelpful because the question was unclear, a different model gives you a different unhelpful answer.
- **To find agreement.** Cycling until a model tells you what you want is a way of laundering a decision you had already made. If two of three disagree, that is the finding.
- **On a very long thread, into a smaller-context model.** If the early context matters and the thread is huge, either stay with the large-context model or summarise the thread first and start fresh.

For running two models on the same prompt simultaneously rather than sequentially, see [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side).

Key points:

- Four switching patterns that cover most real work, with the prompts for each
- What the new model does and does not inherit when you switch
- When switching helps and when it just adds a step

Checklist:

- Assign a model to each phase of the task before you start
- Tell the new model what job it is taking over, rather than just continuing
- Ask explicitly for independence when you want a real second opinion
- Never let the model that drafted something also be the one that critiques it
- Switch when an explanation does not land, instead of re-reading the same one
- Finish a draft before switching, so the voice stays consistent
- Keep notes on which model wins which phase for your own work

Steps:

1. Start a chat with one model: Pick the model that best fits the first task, for example GPT for brainstorming.
2. Click the model selector in the composer: Change the active model to your next pick, for example Claude to draft the final copy.
3. Continue the conversation: Whizi carries the full context forward so the new model sees everything that was said.
4. Repeat as needed: Switch again to Gemini for a final fact-check, or back to GPT for a structured summary.

FAQ:

Q: Does the next model see the full conversation?
A: Yes. Whizi passes the full conversation context forward by default, including your messages, previous model responses, and attached files, so the new model joins with everything visible. The one caveat is the per turn context budget of 40,000 input tokens, which applies to every model equally: a very long thread is trimmed to fit, oldest messages first. An answer that seems to have forgotten the early part of a long discussion is usually hitting that budget, not a property of the model you switched to.

Q: Can I hide part of the thread from a specific model?
A: You can fork the thread or start a new chat with only the context you want to carry over, which is the right approach for sensitive snippets and also for long threads that have accumulated irrelevant detail. Starting clean with a short summary of what matters often produces better answers than carrying forty messages of history, since older context competes for attention with the actual question.

Q: Will switching models mid-draft change the writing style?
A: Yes, noticeably, and that is a reason to finish a draft in one model before switching. Voice consistency is precisely what differs between models, so a handover halfway through a long piece leaves a visible seam. The productive pattern is to draft entirely in one model, then switch for the critique, the fact check, or a structured extraction.

Q: Is there automatic model routing?
A: Yes, as a single picker row called Auto. It reads each message, sorts it into one of six fixed rungs, and answers on the model pinned to that rung, never above what your plan can open. There are no routing rules to write: nothing in Settings edits the ladder, weights a rung, or pins Auto to a model. Picking a model by hand stays a per-message decision, and the manual habit of choosing per phase tends to teach you more about which model actually suits your work.

Q: Does switching count as extra messages?
A: Switching itself costs nothing, but each turn is charged at the credit cost of the model that answers it, so the price of the next answer changes with the switch. A message answered by a 1 credit model costs 1 credit and the same message answered by a 20 credit model costs 20; the [credit scale](https://whizi.io/docs/credits) lists every model's cost. Asking the same question of a second model is a second charged turn, which is worth it when the answer matters and unnecessary when it does not.

### Does Whizi have team or workspace plans?

URL: https://whizi.io/docs/team-accounts/
Updated: August 2026

Quick answer: No. Whizi has no team, workspace, organization or seat concept: no admin console, no invites, no member management. Billing is per individual account, and tier is derived from that account's own subscription. The three plans are individual, Starter at $15.99, Pro at $29.99 and Powerhouse at $49.99 per month.

No. Whizi sells three individual plans, and nothing in the product represents more than one person on an account. What that means for a group, and what to do instead.

#### The short answer

No. There is no team, workspace, organization or seat concept anywhere in Whizi: not in the data model, not in the routes, and not in the pricing path. Billing is per individual account, and a subscriber's tier is derived in exactly one place, from that user's own subscription. Everyone who wants to use Whizi needs their own account and their own subscription.

The three plans that exist are individual plans: Starter at $15.99/month, Pro at $29.99/month and Powerhouse at $49.99/month. Those three tier names are the only plans the pricing path will even parse, so no fourth paid tier exists to be a team tier.

| What people look for | What Whizi has |
| --- | --- |
| Team or workspace plan | Does not exist. Three individual tiers only |
| Per-seat pricing | Does not exist. No seat-based, enterprise, education or student tier |
| Admin console, invites, member management | Does not exist. Nothing represents more than one user on an account |
| Pooled credits across people | Does not exist. Allowances are per account |
| Shared chat library or shared prompt library | Does not exist. A conversation can be published as a read-only link |
| A developer API to build your own team layer on | Does not exist. No public API, no API keys, no personal access tokens |

For a buyer: Whizi is priced and built for one person per account. A team of five is five accounts and five subscriptions, bought separately.

#### What "billing is per individual account" actually means

Tier resolution happens once, on the server, from the subscription attached to the requesting user. Nothing reads a parent organization, a group, or a payer other than the account making the request. So one person paying does not lift anyone else, and a plan cannot be spread across two logins.

Every allowance follows the same rule. Credits, message limits, voice minutes and image generations are counted against the account that spent them, not against a pool. Starter carries 400 credits per monthly billing period, Pro 2,000 and Powerhouse 8,000, and those are one person's allowance each. There is no pool, no transfer and no rollover between accounts.

Projects do not change this. Projects are folders of conversations with pinned files and per-project instructions, and they are available to any signed-in account, but a project belongs to the account that created it. There is no member list on a project and no way to hand one to a colleague.

A colleague on their own free account is not blocked from trying Whizi first: a free account is not model-gated and can try any text model, but it has a lifetime allowance of 7 messages that never resets. After that the reply is `Your free messages are used up. Start a subscription to keep chatting.`, HTTP 402 `free_limit_reached`.

#### Sharing one login instead: three things that happen

The first is an anti-abuse gate. When more than 3 distinct device or session keys have been seen on one account in the last 24 hours, the server answers `Account sharing limit reached. Please use your own Whizi account.`, HTTP 403 `account_sharing_limit`. That is neither a billing refusal nor a tier refusal, and there is no seat concept to buy your way past it.

The second is the generation lock, which bites long before the sharing gate does. Only one generation can run per account at a time, enforced by a per-user lock, so two people on one login can never generate concurrently. The second person gets `Generation already in progress. Please stop it or wait for it to finish.`, HTTP 409 `generation_in_progress`. On a shared login this is not an edge case, it is the normal state whenever two people are working at once.

The third is the terms. The Whizi Terms of Service grant "a non-exclusive, non-transferable, revocable license" to access the Services, "solely for your personal, non-commercial use or internal business purpose". On registration you also "agree to keep your password confidential and will be responsible for all use of your account and password". Handing the login round is on the account holder.

There is also nothing to gain from it privately. The chat history, memory and projects of one shared login are visible to everyone who holds that password.

#### What a group can actually do today

**One account and one subscription per person.** This is the supported shape. Each person subscribes from their own account, on whichever tier suits their work, and nothing about one account depends on another. Plans can differ per person: one Powerhouse and four Starters is a perfectly normal arrangement, because there is no group to keep consistent.

**Share individual conversations read-only.** A conversation can be published behind an unguessable public link that anyone can read in a browser without a Whizi account. The link is read-only, stripped of user ids, emails and internal fields, screened for safety before it goes public, and revocable at any time. Revoking deletes the row, so a revoked link and a link that never existed both return `This shared chat is no longer available.` as HTTP 404. The walkthrough is in [share a conversation](https://whizi.io/docs/share-a-conversation).

**Hand work over as files rather than as access.** A chat answer can produce a real .xlsx or .docx that downloads in the browser, which is usually the right way to get a result to someone who is not on Whizi. Separately, `GET /api/account/export` returns the account's data as one downloadable JSON file, reachable as Export Data in Settings under Account. See [export chats and files](https://whizi.io/docs/export-chats-and-files).

What is not available as a workaround: there is no public or developer API, no API key issuance and no personal access tokens, so a shared internal tool cannot be built in front of one subscription either.

#### Invoicing several people at once

There is no self-serve way to do it. Checkout is per account, and the product has no basket that buys more than one subscription.

One exception exists on the operator side and it is deliberately not a product feature: entitlements can be granted by hand for an invoice covering several seats, paid outside the app. That is a manual arrangement handled by Whizi, not a setting anyone can reach in the interface, and it does not create a team object, an admin, or any link between the accounts it grants.

If that is what you need, [support](https://whizi.io/support) is the only route to it. Everything else about billing, including where to cancel and when changes take effect, is in [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

#### Before you take Whizi into a company

Two things in the Terms of Service matter to anyone evaluating Whizi for work.

The agreement can be entered by a company. The terms are "made between you, whether personally or on behalf of an entity" and the license covers "internal business purpose". So business use is contemplated, it is simply business use of individual accounts.

Regulated work is excluded. The terms state that the Services "are not tailored to comply with industry-specific regulations (Health Insurance Portability and Accountability Act (HIPAA), Federal Information Security Management Act (FISMA), etc.), so if your interactions would be subjected to such laws, you may not use the Services", and that you may not use the Services in a way that would violate the Gramm-Leach-Bliley Act. That rules out the health, federal and financial-services cases before the team question even comes up.

For where data goes and what is retained, which is the other question procurement asks, see [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

Key points:

- Nothing in Whizi represents more than one person on an account: no seats, no admin, no invites
- What each person in a group actually needs, and what one shared login costs you
- The verbatim refusals a shared login produces, with their status codes

Checklist:

- Nothing in the data model or the routes represents more than one user on an account
- Three individual tiers only: Starter, Pro and Powerhouse
- No seat-based, enterprise, education or student pricing exists
- Billing is per account and tier is read from that user's own subscription
- More than 3 device or session keys in 24 hours returns `Account sharing limit reached. Please use your own Whizi account.`
- One generation at a time per account, so two people on one login block each other
- Read-only share links and file exports are the supported ways to pass work around
- Multi-seat invoices are a manual operator arrangement, not a self-serve feature

FAQ:

Q: My colleague has no subscription. Can I add them to mine?
A: There is nothing to add them to. Their tier is read from the subscription on their own account, so an account with no plan gets a 7 message lifetime allowance of its own and nothing more, however much anyone else is paying. Past that the reply is `Your free messages are used up. Start a subscription to keep chatting.`, HTTP 402 `free_limit_reached`, and a route that requires a paid plan answers `Choose a subscription to use Whizi.`, HTTP 402 `subscription_required`. Why an account that is paying can still hit that second refusal is covered in [subscription not active](https://whizi.io/docs/subscription-not-active).

Q: I already subscribe. Can I move my plan onto a colleague's account instead?
A: No. The plan stays with the account it was bought on, because tier is derived in exactly one place, from the subscription attached to the requesting user, and nothing in the data model represents more than one user on an account. There is no reassignment and nothing to reassign it to. The working version is that they subscribe from their own account and you cancel yours, from Settings under Subscription. When cancellation takes effect is in [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

Q: We shared one login and now we want to separate. What happens to the chat history?
A: It stays with the account, not with whoever typed it, so changing the password locks the other people out of that history rather than dividing it. Nothing splits an account's chats across two accounts. `GET /api/account/export` hands you the account's data as one JSON file, from Export Data in Settings under Account, but that file is not an import format: Whizi's import accepts exactly two sources, ChatGPT and Claude, so a Whizi export cannot be loaded into a second Whizi account.

Q: Does the 3 device limit count my own phone and laptop as two?
A: Yes. The guard counts distinct device or session keys seen on one account in the last 24 hours, not people, and it fires once more than 3 have been seen: `Account sharing limit reached. Please use your own Whizi account.`, HTTP 403 `account_sharing_limit`. One person on a phone, a laptop and a tablet sits inside that. The same person plus a colleague on two more devices does not. Signing in on several of your own devices is supported, with the separate rule that only one generation can run per account at a time.

### Whizi troubleshooting: fixing common problems

URL: https://whizi.io/docs/troubleshooting/
Updated: August 2026

Quick answer: Most Whizi problems come down to a plan, a setting or a file rather than a fault. A locked model means the plan you are on. An empty PDF means no text layer, so upload the pages as images. An out of date answer usually means web search is off in that chat.

A model is locked, an upload came back empty, the answer is out of date, credits ran out early. What each symptom means and what to do about it.

#### Symptom, cause and fix

The symptoms below account for most support questions. Each one has its own section further down.

| Symptom | Cause | Fix |
| --- | --- | --- |
| A model shows as locked | It needs a higher plan than the one you are on | Upgrade, which takes effect immediately, or pick a model at your tier |
| An uploaded PDF produced nothing useful | The PDF has no text layer, so browser-side extraction returned nothing | Upload the pages as `.png` or `.jpg` instead |
| The answer is out of date or invented | Web search is switched off in that conversation | Turn the search toggle on in the composer and ask again |
| Credits ran out before the month did | An expensive default model doing routine work | Set a 1 credit model as the default and switch up deliberately |
| A conversation will not continue on its model | A downgrade, or a trial on a higher plan ending | Pick a model at your current tier and continue in the same thread |
| Billing did not stop after cancelling | The subscription was bought inside the iOS or Android app | Cancel in Apple or Google subscription settings |

Questions about price, plans, credits and privacy rather than a symptom are answered on [the Whizi FAQ](https://whizi.io/docs/whizi-faq). If no symptom here matches yours, [how to contact Whizi support](https://whizi.io/docs/contact-whizi-support) sets out which channel to use and what to include.

#### A model shows as locked

The model requires a higher plan than the one you are on. Access is decided on the server rather than in the interface, so this is a real entitlement rather than a display bug.

The three most common cases: Claude of any kind on Starter, which includes no Anthropic model at all; Claude Opus or GPT-5.6 Sol on Pro, both of which are Powerhouse; and any newly added model, because the fallback tier for anything not explicitly listed is Powerhouse rather than Pro.

The fix is to upgrade, which takes effect immediately, or to pick a model at your tier. The full mapping of model to plan is in [the model reference](https://whizi.io/docs/models).

#### An uploaded PDF produced nothing useful

Almost always because the PDF has no text layer. Text is extracted in your browser before anything is sent, so a PDF that is a photograph or a scan of a page contains no extractable text and the model receives an empty document.

The fix: upload the pages as images (`.png`, `.jpg`) instead. Vision-capable models read a page image directly, which works on scans, photographs of documents and handwriting.

A second, subtler version of this: the text extracted fine but the layout was lost, so a complex table came through flattened. For tables specifically, upload the `.csv` or `.xlsx` rather than the PDF containing them. The model gets far cleaner structure and the answers get noticeably more reliable. See [supported file types](https://whizi.io/docs/supported-file-types).

#### The answer is out of date or invented

Check whether web search is on for that conversation. It is on by default, but the setting is stored per chat, so a chat you once switched it off in stays off. Without it, the model answers from training data with a cutoff months in the past, confidently and without flagging the staleness. This is the most common way an AI answer goes quietly wrong.

Turn the search toggle on in the composer and ask again. If the answer changes, the first one was stale. With the toggle on the model still decides per question whether to search, so if a sourceless answer looks doubtful, ask again and say you want it checked against current sources. If you are asking about something genuinely obscure, search matters for a second reason: thin training coverage is precisely the condition under which models invent plausible detail.

For anything you intend to act on, ask a second model the same question in the same conversation. Where two models agree on a specific figure and a source, confidence is reasonable. Where they disagree, at least one is wrong and you would not have known from reading either alone. See [using web search](https://whizi.io/docs/web-search), and if the toggle is already on but no results ever come back, [web search not working](https://whizi.io/docs/web-search-not-working) lists the error strings behind that.

#### Credits ran out before the month did

Look at what your default model costs. The single biggest driver is defaulting to an expensive model for routine work: at 10 credits per message, a 2,000 credit Pro allowance is gone in 200 messages, while the same allowance on a 1 credit model is 2,000 messages.

Roughly half the catalogue costs 1 credit and those models handle lookups, rewrites, summaries and formatting perfectly well. Set one of them as your default and switch up deliberately for the messages that decide something.

Two other quiet drains: side-by-side comparison costs two messages rather than one, since both models answer and each bills at its own rate. And the 20 credit rung, Claude Opus 5, GPT-5.6 Sol and Sonar Pro, is worth twenty fast-tier messages each. The credit cost of every model is shown in the picker before you send. Full detail in [how credits work](https://whizi.io/docs/credits).

#### A conversation will not continue on the model it started on

This happens after a downgrade, or when a trial on a higher plan ends. Downgrading changes which models you can send new messages to, not what you can read, so the conversation is intact and fully readable but that particular model is no longer available to you.

Pick a model at your current tier and continue in the same thread. Context carries across the switch, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### Billing did not stop after I cancelled

Check where you subscribed. A subscription bought inside the iOS or Android app is billed by Apple or Google, and it can only be cancelled in their subscription settings. Cancelling on the website does not touch it, deleting the app does not touch it, and deleting your Whizi account does not touch it.

If you subscribed on whizi.io, cancellation happens in account settings and takes effect at the end of the current paid term rather than immediately, so one further charge is not expected but continued access is. See [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

Key points:

- The five symptoms that account for most support questions
- Why a scanned PDF returns nothing, and the one line fix
- What to do when credits run out before the month does

Checklist:

- Locked model means the plan, not a bug: check the tier mapping
- Empty PDF extraction means no text layer: upload pages as images
- Upload tables as .csv or .xlsx rather than inside a PDF
- Stale answer usually means search is off in that chat
- Cross-check anything important against a second model
- Credits running out usually means an expensive default model
- Store subscriptions can only be cancelled with Apple or Google

FAQ:

Q: Why is a model locked when I am already paying?
A: Because it needs a higher tier than your current plan. The three usual cases are any Claude model on Starter, which includes no Anthropic model, Claude Opus or GPT-5.6 Sol on Pro, which are both Powerhouse, and recently added models, which default to Powerhouse until they are explicitly promoted. Upgrading applies immediately.

Q: Why did my PDF upload not work?
A: The most likely reason is that it is a scan or a photograph with no text layer, so browser-side extraction returned nothing and the model received an empty document. Upload the pages as .png or .jpg instead, since vision-capable models can read a page image directly. This also works for handwriting.

Q: Why does the AI keep giving me old information?
A: Web search is probably off in that conversation, so the model is answering from training data with a cutoff months ago, and it will not flag that itself. Search ships on, but the setting is per chat, so an older conversation can still be carrying an off. Switch the toggle on in the composer and ask again. If the answer changes materially, the original was stale rather than wrong in any deeper sense.

Q: I ran out of credits in the second week. What should I change?
A: Almost certainly your default model. A 10 credit model burns a 2,000 credit Pro allowance in 200 messages, while a 1 credit model gives you 2,000. Set a 1 credit model as the default for routine work and reach for the expensive tier deliberately. If you genuinely need frontier models constantly, Powerhouse at 8,000 credits is the tier that fits.

### How to upload documents to AI in Whizi

URL: https://whizi.io/docs/upload-documents/
Updated: August 2026

Quick answer: Attach the file in the composer, by dragging it in or clicking the paperclip, then ask your question. Whizi accepts PDFs, Word documents, spreadsheets, images and code or plain text files. What the model receives differs by type: extracted text for documents, cell values for spreadsheets, and the image itself for pictures.

Upload PDFs, Word docs, spreadsheets, images, and code to Whizi and chat with them using Claude, GPT, or Gemini.

#### Supported file types, and what actually happens to each

Attaching a file is one action, but what the model receives differs by type, and knowing which is which explains most of the surprising results.

| Type | Extensions | What the model works with |
| --- | --- | --- |
| Text PDFs | .pdf | The extracted text, in reading order, including tables as text |
| Scanned PDFs | .pdf | Text recovered by OCR, which is very good but not perfect |
| Word documents | .docx | The document text; comments and tracked changes may not carry through |
| Spreadsheets | .xlsx, .csv | Cell values as text, not live formulas |
| Images | .jpg, .png, .webp | The image itself, read by a multimodal model |
| Code and plain text | .py, .js, .ts, .go, .md, .txt and similar | The file contents verbatim |

Two consequences worth internalising. First, a spreadsheet arrives as values rather than as a working workbook, so the model can read what your formula produced but cannot re-execute the formula. Second, a scanned PDF has passed through OCR, which means an unusual font, a handwritten annotation, or a low resolution scan can produce a misread character in a number. Verify figures from scanned documents against the original.

#### Prepare the file before you ask anything

Five minutes here saves a confusing conversation later.

- **Name the file for what it is.** The model sees the filename, and `2025-q3-vendor-contract-acme.pdf` gives it useful context that `document(3).pdf` does not. It also matters enormously once several files are in the thread.
- **Send the source, not the presentation version.** For spreadsheets especially, the version with merged cells, subtotals inside the data, and notes in the margin causes far more errors than a large plain file does.
- **Split enormous documents by purpose, not by size.** If you only need the methodology and results sections, uploading those is better than uploading 400 pages and hoping attention lands in the right place.
- **Remove personal data you do not need.** Names, emails, and identifiers are rarely required for the analysis, and stripping them is faster than deciding whether you were permitted to upload them.
- **Check that a PDF is actually text.** Try selecting text in your PDF reader. If you cannot, it is a scan and OCR is involved, which changes how much you should trust extracted numbers.

#### Choosing the model for the file

This choice matters more than the prompt for document work.

- **Gemini** for anything long. It carries the largest context window of the models in Whizi, which means a 200 page report, a full contract set, or a dozen papers can be present at once rather than being processed in pieces. Cross-document questions only work properly when everything is genuinely in context.
- **Claude** for nuance. Legal language, policy documents, anything where the meaning depends on qualification and hedging. It is also the best at telling you what a document does not say, which is frequently the actual question.
- **GPT** for extraction. Invoices, forms, structured data, anything where you want a strict table or JSON back and consistent field names across a hundred rows.
- **Any multimodal model** for images, charts, screenshots, and diagrams. Ask it to describe what it sees before asking it to interpret, which catches misreadings early.

You can switch models mid-thread without re-uploading, which is the point of doing this in one workspace. Read with Gemini, extract with GPT, and have Claude write the summary, all against the same attached file. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### Working with several documents at once

Single document questions are useful. Multi document questions are where this stops being a convenience and starts being a capability you did not previously have.

**Prompt: compare**

`Compare contract-a.pdf and contract-b.pdf. Return a table of every clause that differs: clause, what A says, what B says, and which is more favourable to us and why. Ignore differences in formatting and numbering. List clauses present in one and absent from the other separately.`

**Prompt: consolidate**

`Across all attached files, build one table answering [question]. One row per document, with the document name in the first column. Where a document does not address it, write NOT ADDRESSED. Do not merge documents that use different definitions without flagging it.`

**Prompt: find the contradiction**

`Where do these documents contradict each other? Quote both passages side by side and name the file each came from. Distinguish genuine contradictions from differences in scope or date.`

Always reference files by name in your prompt once there are more than one. "In the vendor agreement, what is the notice period?" is unambiguous; "what is the notice period?" makes the model choose, and it will not tell you which it chose.

#### When the answer looks wrong

A short diagnostic sequence that resolves most problems.

**The model says it cannot see the file.** Confirm the attachment actually uploaded, then reference it explicitly: "In the attached report.pdf". For images especially, saying "look at the attached image" ensures the multimodal path is used rather than the model answering from the text of your question alone.

**The answer is about the wrong document.** Name the file in every question once the thread has several.

**The numbers are subtly wrong on a scanned document.** This is OCR, not the model. Ask it to quote the surrounding line verbatim so you can see what it read, then check the original.

**It missed something you know is in there.** Ask a locating question first: `Quote the passage that discusses [topic] and give its page or section.` If it cannot find it, the extraction is the problem, not the reasoning. Try a different model or upload the relevant pages separately.

**It is summarising instead of answering.** Ask for the evidence rather than the conclusion: `Quote the exact text that supports your answer.` This also catches the case where the answer was inferred from general knowledge rather than from your document, which is the most important failure to catch.

That last check deserves emphasis. A model asked about a well known contract type, a common regulation, or a famous paper can answer plausibly from training data without consulting your file at all. Requiring a quote forces the answer to be anchored to the document actually in front of it.

#### Privacy and what to think about before uploading

Files are stored in your workspace, are not used to train models, and can be deleted at any time. Each provider's data handling policy is available for review before you enable that model.

The judgment that remains yours is whether the document is one you are permitted to process with a third party service at all. The categories worth pausing on: personal data belonging to other people, anything covered by a specific confidentiality clause, material non public information, and regulated data such as health or financial records subject to particular handling rules.

The practical middle path for most people is to redact rather than abstain. Names replaced with roles, figures indexed rather than absolute, and identifiers removed will preserve everything the analysis needs while removing most of what makes a document sensitive.

Key points:

- What each file type can and cannot do once uploaded, including the limits of scanned PDFs
- How to work with several documents at once, which is where most of the value is
- The preparation steps that turn a bad answer into a good one before you ask anything

Checklist:

- Name files descriptively, since the model sees the filename
- Check whether a PDF is text or a scan before trusting extracted numbers
- Upload the source spreadsheet, not the formatted presentation version
- Pick Gemini for long files, Claude for nuance, GPT for extraction
- Reference files by name once more than one is attached
- Ask for a supporting quote so answers stay anchored to the document
- Strip personal data you do not need for the analysis

Steps:

1. Open a chat: Start a new thread or open an existing one.
2. Attach the file: Drag the file into the composer or click the paperclip icon.
3. Pick the right model: Use Gemini for long documents, Claude for nuance, GPT for structured extraction.
4. Ask your question: Reference the file by name if you have multiple, for example "In report.pdf, summarize section 3."
5. Save or export: Copy the answer into your draft or keep the conversation for later.

FAQ:

Q: Are my uploads private?
A: Files are stored securely in your workspace, are not used to train models, and can be deleted whenever you want. Each provider’s data handling policy is available for review before you enable that model. What remains your call is whether a given document is one you are permitted to send to a third party service, which is usually governed by your own contracts or internal policy rather than by ours.

Q: What is the maximum file size?
A: It depends on your plan and the model, and in practice the binding constraint is the model context window rather than a hard file size. Gemini in Whizi handles the largest documents, which is why it is the default recommendation for long contracts, reports, and multi-document sets. For anything beyond that, upload the relevant sections rather than the whole file, since targeted context usually produces a better answer anyway.

Q: Can I delete a file after chatting?
A: Yes, files can be removed from your workspace at any time. Note that removing a file does not retroactively erase what was already discussed in the conversation, so if the content is sensitive enough to warrant deleting the file, delete the conversation as well.

Q: Does it work with scanned PDFs?
A: Yes. Scanned documents go through OCR automatically on most plans, and you can also ask the model to OCR and then answer. Accuracy is high on clean scans and drops with low resolution, unusual fonts, handwriting, and complex table layouts. Because a misread character in a number is invisible in a fluent answer, verify any figure from a scanned document against the original.

Q: Can I upload several documents to the same conversation?
A: Yes, and it is where most of the value is. Comparing two contracts clause by clause, consolidating findings across a set of reports, or finding contradictions between documents are all things that require everything to be in context at once. Reference files by name in your prompts once there are more than one, otherwise the model picks for you without saying which it chose.

### How to use Projects in Whizi

URL: https://whizi.io/docs/use-projects/
Updated: August 2026

Quick answer: A project in Whizi is a folder of conversations that also carries pinned files and its own instructions, and both are folded into the model prompt on every message. Projects are available to any signed-in account, including free. Each account can hold 50 projects, each with up to 10 pinned files of 10 MB.

Projects are folders of conversations that carry pinned files and their own instructions. How to create one, what fits inside, and how that context reaches the model.

#### The short answer

A project in Whizi is a folder of conversations that also carries pinned files and its own custom instructions, and on every message in that project the instructions and the pinned file text are loaded and folded into the model prompt. Projects are available to any signed-in account, including the free tier. The per-project caps are abuse guards rather than a paid gate. How that lines up against the feature people usually arrive from is set out in [ChatGPT Projects alternative](https://whizi.io/resources/chatgpt-projects-alternative).

The caps published in the product:

| Limit | Value |
| --- | --- |
| Projects per account | 50 |
| Project name | 120 characters |
| Project instructions | 32,000 characters |
| Pinned files per project | 10 |
| Size of one pinned file | 10 MB |
| Extracted text used from one pinned file | 32,000 characters |
| Whole project prompt block | 120,000 characters |
| Projects a chat can belong to | 1 |

Two behaviors matter more than the numbers. Pinned file bytes are stored until you delete them, unlike chat attachments which expire. And project context is rebuilt into the prompt on every turn rather than once at the start, which is what makes a project reliable and also what makes it expensive in context.

#### What lives inside a project

A project carries three things worth understanding.

**Conversations.** A project is first a folder. Chats can be moved into it and back out of it, and a chat belongs to at most one project at a time, which is what makes the sidebar able to show a clean "everything else" view of the chats that are in no project. The sidebar lists projects above the chat list. Sidebar search matches chat titles and cuts across projects, so a chat stays findable by name whichever folder it is sitting in. Renaming and deleting are covered in [organize your conversations](https://whizi.io/docs/organize-conversations), and what the search box will and will not match is in [search your chat history](https://whizi.io/docs/search-your-chat-history).

**Instructions.** A project carries custom instructions of up to 32,000 characters, applied to every conversation in it. This is worth stating plainly: per-project instructions are the only user-authored instructions in Whizi. There is no account-level custom instructions setting. Settings offers Subscription, Cancel, Reset password, Export data, Import chats, Default model, Dark mode, Language, and support and legal links, and nothing that sets a global instruction. If you want the same standing brief on every chat about one piece of work, a project is the mechanism.

**Pinned files.** Up to 10 files, each up to 10 MB, whose text rides along with every turn. These are not chat attachments and they do not behave like chat attachments, which is the next section.

#### Pinned files, and why images are not allowed

Pinned files accept PDF, plain text, Markdown, CSV, JSON, `.docx` and `.xlsx`. Images are deliberately excluded.

The exclusion is the clearest signal of what a pinned file actually is. A pinned file is not handed to the model as a file. Its text is extracted and rides along as prompt text on every turn in the project, so a format with no text to extract has nothing to contribute. An image attached to a single chat message is passed to a vision-capable model as an image and works fine; the same image pinned to a project would have to be re-sent on every turn to be of any use, which is exactly what the product declines to do.

That also explains the retention difference. Chat attachments and generated media are configured to expire after up to 30 days. Pinned file bytes are stored until you delete them, because a pinned file has to be available for every future turn in the project rather than for one conversation. The full storage picture is in [where Whizi stores your data](https://whizi.io/docs/where-whizi-stores-data).

If a pinned file extracts to nothing, Whizi says so beside the file rather than letting you assume it worked: `No readable text found in this file, so the model cannot see it.` A scanned PDF with no text layer is the usual cause. The type, size and signature check failures on project files are covered in [why a file upload failed](https://whizi.io/docs/file-upload-failed).

#### How project context reaches the model

On every message, the project instructions and the pinned file text are loaded and folded into the model prompt. Not on the first message of a chat, not once per chat. Every message.

Two caps trim what gets folded in. Text extracted from one pinned file contributes at most 32,000 characters, and the whole project prompt block is capped at 120,000 characters. A project holding ten large PDFs is therefore not sending you ten whole PDFs on every turn, and the parts past the caps are simply not there. What the rebuilt block costs a long conversation is worked through in [when the conversation is too long](https://whizi.io/docs/context-too-long).

What this does not change is what a turn costs. A credit charge is the rung of the model that answered, so a message in a project costs the same credits as the same message on the same model outside one. Pinned files spend context, not credits.

What carries across a model change inside a project chat is covered in [switch models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### The errors a project produces

Project management actions fail with their own strings, and they are worth recognising because each one names a different operation.

| Message | When it appears |
| --- | --- |
| `Could not create the project. Please try again.` | Project creation failed |
| `Could not save the project. Please try again.` | Saving the name or the instructions failed |
| `Could not delete the project. Please try again.` | Project deletion failed |
| `Could not move the chat. Please try again.` | Moving a chat into or out of a project failed |
| `Project not found.` | HTTP 404, a project id that does not resolve for your account |
| `Project file not found.` | HTTP 404, a file id that does not resolve inside the project |

Two rate limits are specific to projects, on top of the account-wide ceilings. Creating or editing a project is limited to 20 per minute, and uploading a project file to 10 per minute. Both are per minute, so the fix is to wait rather than to retry immediately. The wider picture is in [rate limits in Whizi](https://whizi.io/docs/rate-limits).

Upload failures on a pinned file are a different set of strings, including the 10 file ceiling and the 10 MB cap, and they live in [why a file upload failed](https://whizi.io/docs/file-upload-failed).

#### What a project is not

Four honest limits, because they are the questions people arrive with.

**A project is not shared.** There is no team, organization, workspace or seat concept in Whizi, and a project belongs to the account that created it. There is no member list on a project and no way to hand one to a colleague. See [team accounts in Whizi](https://whizi.io/docs/team-accounts).

**A project is not an account-level setting.** Instructions are set per project. A chat outside every project has no instructions at all, and there is nothing that applies a standing instruction across the whole account.

**A project is not a second home for a chat.** A chat belongs to at most one project, so a conversation that is relevant to two pieces of work has to pick one.

**A project is not a prompt library.** Instructions are one block of standing context for that project, not a set of saved, named prompts you pick from per message.

What a project does travel through is your data export. `GET /api/account/export` returns your account data as one downloadable JSON file, and its contents include projects and their file metadata. Read that precisely: the metadata, not the pinned bytes. See [export chats and files](https://whizi.io/docs/export-chats-and-files).

#### Getting the most out of one

A few things follow directly from how the mechanism works.

- **Pin the stable material, attach the volatile material.** A style guide, a schema, a contract or a spec is worth pinning because every turn benefits. A one-off document you want summarised belongs in a single message as a chat attachment.
- **Unpin what you are not asking about.** Ten pinned files consume the project block on every turn whether the current question touches them or not.
- **Choose the format before you pin, not after.** The choice costs more on a pinned file than on a chat attachment, because the extracted text is re-sent on every turn rather than read once. What each format extracts to is in [supported file types](https://whizi.io/docs/supported-file-types).

On a free account, projects, pinned files and instructions are all available. The constraint is elsewhere: a free account has a lifetime allowance of 7 messages, which never resets, so a project set up on free is really set up ahead of a subscription rather than something you can work inside for long.

Key points:

- The project caps in one table: 50 projects, 10 pinned files, 32,000 characters of instructions
- Why images cannot be pinned, and what that tells you about how pinned files are sent
- What the model actually receives on each turn in a project, and the two caps that trim it

Checklist:

- A project is a folder of chats plus pinned files plus its own instructions
- Available on any signed-in account, including free
- Up to 50 projects, each with up to 10 pinned files of 10 MB
- Instructions can be up to 32,000 characters
- Pinned files accept PDF, text, Markdown, CSV, JSON, .docx and .xlsx, never images
- Instructions and pinned file text are folded into the prompt on every message
- One file contributes at most 32,000 characters, the whole block at most 120,000
- Pinned file bytes are kept until deleted, unlike attachments which expire
- A chat can sit in at most one project

Steps:

1. Open the project list: Projects are listed in the sidebar, above the chat list.
2. Create the project and name it: The create dialog takes a name of up to 120 characters.
3. Write the project instructions: Up to 32,000 characters, applied to every conversation in that project.
4. Pin the files the work depends on: Up to 10 files, each up to 10 MB. PDF, text, Markdown, CSV, JSON, .docx and .xlsx are accepted.
5. Move chats in: A chat can be moved into a project, or back out of it. A chat belongs to at most one project.

FAQ:

Q: Do I need a paid plan to use Projects in Whizi?
A: No. Projects are available to any signed-in account, including the free tier. The real constraint on free is the message allowance rather than the feature: a free account gets a lifetime allowance of 7 messages, which never resets, so you can build the project and pin its files but you will not get far chatting inside it.

Q: How many files can I pin to a project?
A: Ten, each up to 10 MB. Accepted formats are PDF, plain text, Markdown, CSV, JSON, .docx and .xlsx. Images cannot be pinned, because a pinned file is sent as extracted prompt text on every turn and an image has no text to extract. An eleventh file is refused before the upload starts with `This project already holds 10 files. Remove one first.`

Q: Are pinned files sent to the model every time?
A: Yes, every turn. Ten pinned files are re-sent whether the current question touches them or not, rather than being sent once at the start of a chat, so unpinning what you are not asking about is a real lever. Two caps trim what is sent: 32,000 characters of extracted text from any single file, and 120,000 characters for the whole project block.

Q: Does a project cost extra credits?
A: No. A turn is charged at the rung of the model that answered it, so the same message on the same model costs the same whether it sits in a project or not. What pinned files consume is context rather than credits.

Q: Can I set custom instructions for my whole account instead of per project?
A: No. There is no account-level custom instructions setting anywhere in Whizi, and per-project instructions are the only user-authored instructions in the product. A project is the mechanism: one standing block of up to 32,000 characters, applied to every conversation you file in it.

Q: Can two people work in the same project?
A: No, and the nearest workaround runs per conversation rather than per project: a chat can be published behind an unguessable public link, and that public view is read-only and stripped of user ids, emails and internal fields. A colleague can therefore read one transcript that came out of the project, which is a different thing from working in the project. See [team accounts in Whizi](https://whizi.io/docs/team-accounts).

Q: Do pinned files expire like uploads do?
A: No. There is no expiry window on a pinned project file: the bytes are stored until you delete them. Uploaded chat attachments and generated media are the ones with a clock on them, configured to expire after up to 30 days. The full storage picture is in [where Whizi stores your data](https://whizi.io/docs/where-whizi-stores-data).

### Can you use Whizi offline?

URL: https://whizi.io/docs/use-whizi-offline/
Updated: August 2026

Quick answer: No. Whizi has no offline mode. Every answer is a network round trip to OpenRouter and a routed model provider, and your conversations, plan and allowances live on the server rather than the device. The mobile app may keep a local cache of up to 40 recent conversations, which is a cache rather than an offline mode.

No. Every answer is generated by a model provider reached over the network, and your conversations, plan and allowances live on the server. What is on your device, what is not, and what to do before you lose signal.

#### The short answer

No. Whizi has no offline mode. To generate a response Whizi sends the information reasonably needed for the request to OpenRouter and a routed model provider, or to another provider used by the selected feature, so an answer is always a network round trip to a company other than Whizi. Hosting, synchronized chats, media, account state, usage limits, reports, and security and operational processing all run on Cloudflare Workers, D1, R2 and Durable Objects, which means the conversation you are typing into lives on the server too.

The model list is server-side for the same reason. The picker is served by the Whizi backend at `GET /api/models`, and the model lists hardcoded into the web and mobile clients are offline fallbacks only, used when that call cannot be reached. A fallback list is a list, not a model: seeing a name in the picker when the network is gone does not mean anything behind it can answer.

| What needs a connection | What happens on your device |
| --- | --- |
| Every model reply, on every plan | The theme you chose, saved to local storage and restored on your next visit |
| Loading your chat list and any conversation | Parsing a ChatGPT or Claude export file, which the browser reads before anything is posted |
| Sending a message, editing one, switching versions | Building the .xlsx or .docx from a file card in an answer |
| Web search, voice, image generation | The mobile app cache of recent conversations, described below |
| Uploading a file, and viewing stored media through its signed URL | |

#### What the mobile app keeps on the device

The privacy policy is specific about this. The mobile app may keep a local cache of up to 40 recent conversations on a signed-in or guest device, and that cache is not a separate cloud backup. How it is cleared, and where everything else on the account sits, is in [where Whizi stores data](https://whizi.io/docs/where-whizi-stores-data).

A cache of recent conversations is not an offline reading mode, and it changes nothing about where the work happens. Conversations and messages are kept on the account until you delete them or delete the account, and a new reply still has to reach a provider over the network.

Your plan is resolved on the server too. Billing is per individual account, with tier derived in exactly one place from that account's own subscription. See [Whizi on mobile](https://whizi.io/docs/whizi-on-mobile) for how the two platforms share one account.

#### A connection that drops in the middle of a reply

Generation runs in a per-conversation Durable Object on the server, so closing the tab mid-stream does not kill the answer and the reply is still persisted. A reopened tab or a second device can reattach to a generation already in flight through `GET /api/conversations/:id/generate/active`. Come back, reload the conversation, and read what was saved.

There is one narrow window at the very start where that does not apply. If the client disconnects before the upstream connect completes, the server records `Generation aborted before it started.` and answers HTTP 408 with an empty body, so nothing reached a model.

Stopping is server authoritative rather than client side: the coordinator cancels the provider read, saves the partial answer, releases the lock and refunds an undelivered turn. The full set of mid-stream strings and what each one means is in [a message will not send](https://whizi.io/docs/message-not-sending).

#### The strings you see when the call cannot complete

Whizi surfaces the failure of whichever call was in flight rather than one generic message, so the string you get depends on what you were doing. Each is quoted verbatim, with the condition it is documented for.

| String | Documented condition | Where it is covered |
| --- | --- | --- |
| `Failed to send message.` | The website fallback toast when a send fails and the backend supplied no message | [message not sending](https://whizi.io/docs/message-not-sending) |
| `Could not load your chats. Please refresh to try again.` | The toast shown when the chat list fails to load | [message not sending](https://whizi.io/docs/message-not-sending) |
| `Failed to retrieve session token.` | Thrown whenever the token fetch returns nothing before a chat list, chat load, delete, rename or upload call | [login and account access](https://whizi.io/docs/login-and-account-access) |
| `The import could not be finished. Check your connection and try again.` | Any other upload failure during an import run | [chat import errors](https://whizi.io/docs/chat-import-errors) |
| `Something went wrong. Please close and try again.` | The share dialog generic error, shown only when the failure carries no message of its own, such as a dropped connection or an HTML error page | [share a conversation](https://whizi.io/docs/share-a-conversation) |
| `The realtime voice connection failed.` | Shown on a websocket error event | [voice mode not working](https://whizi.io/docs/voice-mode-not-working) |
| `The realtime voice connection was closed.` | Shown when the websocket closes after setup completed | [voice mode not working](https://whizi.io/docs/voice-mode-not-working) |
| `Connection error` | The voice overlay state label for a failed voice session | [voice mode not working](https://whizi.io/docs/voice-mode-not-working) |
| `That didn't send. Check your connection and try again.` | The feedback fallback for every other failure | [troubleshooting](https://whizi.io/docs/troubleshooting) |

Voice deserves one clarification, because its architecture reads like independence and is not. The worker mints a short-lived token and the browser opens a WebSocket straight to Google, so no audio ever passes through Whizi backend servers. That is a privacy property, not an offline one: the socket to Google is still a socket, and the session token that opens it is minted by the worker.

#### What to do before you lose signal

None of these give you a working model without a connection. They are about not being stuck without your content.

**Take your data with you.** `GET /api/account/export` returns the account data as one downloadable JSON file, reachable as Export Data in Settings under Account. The caveat that matters offline: media is exported as metadata plus a signed URL rather than inlined bytes, so the JSON keeps the record while the image itself still needs fetching. What the account holds is laid out in [where Whizi stores data](https://whizi.io/docs/where-whizi-stores-data).

**Get an answer out as a file while you still have signal.** A chat answer can emit a fenced block that the website renders as a file card and turns into a real .xlsx or .docx in the browser. The build is local, the answer that produced it was not.

**Import when you land, not on the way.** Picking the export file is local work and the upload is not, though a connection that drops during one costs nothing, since every conversation is keyed by account plus source plus source id and lands once. The failure strings are in [chat import errors](https://whizi.io/docs/chat-import-errors).

Key points:

- Why no part of producing an answer can happen on your device
- What the mobile app caches on the device, and why a cache is not an offline mode
- What survives a connection that drops in the middle of a reply

Checklist:

- There is no offline mode and no on-device answer path in Whizi
- Every reply is a network request to a routed model provider
- The model picker is served by the backend, and the built-in lists are fallbacks only
- The mobile app may cache up to 40 recent conversations, and that cache is not a separate cloud backup
- Your tier is derived on the server from the subscription on your own account
- A reply survives a dropped client: generation continues in a Durable Object
- Export your account as JSON from Settings under Account before you travel
- Media in an export is a signed URL, so it still needs fetching later

FAQ:

Q: Does Whizi work on a plane with no wifi?
A: No. Every reply is generated by a model provider reached over the network, so with no connection there is no path to a model and no path to your chat list.

Q: My phone lost signal mid-generation. Can I start a new one on my laptop right away?
A: Not necessarily, because only one generation can run per account at a time. A second attempt while the per-user lock is still held is refused with "Generation already in progress. Please stop it or wait for it to finish." The lock is not permanent: it becomes stealable once the holder is judged orphaned or stalled, and expired locks are swept on a schedule.

Q: Can I scroll an old conversation on my phone in airplane mode?
A: Only within what the mobile app has cached, and images are a separate matter even there. Generated images are stored server side and served back through signed URLs, so a picture in an old thread is fetched at the moment you look at it rather than held on the device.

Q: Does a guest account on mobile change any of this?
A: No. The local cache the privacy policy documents covers a signed-in or guest device alike. The guest credential itself is stored on the device, but Whizi stores a one-way derived identifier so it can recognize the guest profile, synchronize data, verify subscription access and later link that data to an account, which are all server-side operations. Account export works for guest sessions too, since guests hold real chats and memory they may want out before the device is gone.

Q: How long do the media links inside an export keep working?
A: Uploaded attachments and generated media are configured to expire after up to 30 days, and may be deleted sooner by you or during account deletion. Once an object is past its retention window the request answers "Media has expired.", HTTP 404 media_expired, instead of returning the bytes. Fetch anything you want to keep while you still have a connection.

### Using Claude in Whizi: models, credit costs and plans

URL: https://whizi.io/docs/using-claude-in-whizi/
Updated: August 2026

Quick answer: Claude models are included in Whizi, so there is no Anthropic subscription to buy. Pro is the lowest paid plan that reaches Claude, unlocking the Sonnet class and Claude Haiku. The Opus class and Claude Fable 5 require Powerhouse. Starter includes no Claude model. Claude Haiku costs 4 credits per message, the Sonnet class 10.

Which Claude models Whizi carries, what each one costs in credits per message, which plan unlocks it, and how to use Claude without an Anthropic subscription.

#### The short answer

Claude models are included in Whizi, and there is no Anthropic subscription to buy. Pro is the lowest paid plan that reaches Claude: it unlocks the Sonnet class and Claude Haiku. The Opus class and Claude Fable 5 require Powerhouse. Starter includes no Claude model at all.

| Plan | Price | Claude models included | Credits per month |
| --- | --- | --- | --- |
| Free | no charge | any text model, inside a lifetime allowance of 7 messages | 0 |
| Starter | $15.99/month, or $10.99/mo billed annually at $131.88 | none | 400 |
| Pro | $29.99/month, or $19.99/mo billed annually at $239.88 | Claude Sonnet 5, Sonnet 4.6, Sonnet 4.5, Haiku 4.5 | 2,000 |
| Powerhouse | $49.99/month, or $34.99/mo billed annually at $419.88 | the rest of the catalogue, including the Opus class and Claude Fable 5 | 8,000 |

A free account is not model gated. The chat path returns before the model gate for the free tier, so a prospect can send a Claude message inside the 7 message lifetime allowance. That allowance never resets.

#### Every Claude model, with credit cost and plan

A credit is one message on the house model, Whizi AI. A Claude message spends the number of credits on its own row, and nothing about message length, conversation length or answer length changes it.

The cost column below is what one standard answer costs to run at provider list rates. A standard answer is 1,000 input tokens plus 500 output tokens, and the rates were read from OpenRouter on 20 August 2026.

| Model | Plan | Credits per message | Model context window | Cost of one standard answer |
| --- | --- | --- | --- | --- |
| Claude Haiku 4.5 | Pro | 4 | 200K | $0.0035 |
| Claude Sonnet 5 | Pro | 10 | 1M | $0.007 |
| Claude Sonnet 4.6 | Pro | 10 | 1M | $0.0105 |
| Claude Sonnet 4.5 | Pro | 10 | 1M | $0.0105 |
| Claude Opus 5 | Powerhouse | 20 | 1M | $0.0175 |
| Claude Opus 4.8 | Powerhouse | 20 | 1M | $0.0175 |
| Claude Fable 5 | Powerhouse | 50 | 1M | $0.035 |

Six further Claude rows carry a credit cost but are not in the priced comparison index, so there is no per-answer figure to publish for them.

| Model | Plan | Credits per message |
| --- | --- | --- |
| Claude Sonnet 4 | Powerhouse | 10 |
| Claude Opus 4.7 | Powerhouse | 20 |
| Claude Opus 4.6 | Powerhouse | 20 |
| Claude Opus 4.5 | Powerhouse | 20 |
| Claude Opus 4.1 | Powerhouse | 80 |
| Claude Opus 4 | Powerhouse | 80 |

Every Claude model Whizi prices at 4 credits or more per message appears in those two tables. The 80 credit pair sits on the top rung in use, and that rung is Anthropic only: two Claude ids are priced there and nothing else in the catalogue is.

The Pro list names three Sonnet releases, so the oldest Sonnet resolves to Powerhouse despite carrying the same 10 credit rung as the three Pro can open.

#### What an allowance buys in Claude messages

Credits are one shared allowance, so these figures are what you get if you spend the whole period on a single Claude model and nothing else. The Haiku, Sonnet and current Opus arithmetic is already published in the [credits reference](https://whizi.io/docs/credits). What follows is the part that page does not price out: the two dearest Claude rungs, and the weekly allowance.

| Plan and period | Credits | Claude Fable 5 at 50 | the 80 credit Opus rows |
| --- | --- | --- | --- |
| Starter, monthly | 400 | not included | not included |
| Pro, monthly | 2,000 | not included | not included |
| Powerhouse, monthly | 8,000 | 160 messages | 100 messages |

The monthly rows are the ones that decide a Claude budget today. Weekly billing is sold on mobile only, and the production value of the credit allowlist is the web client, so a weekly buyer is metered in messages rather than credits: 100 messages a week on Starter, 400 on Pro, 2,000 on Powerhouse. Each tier does set a weekly credit number explicitly rather than inheriting a fallback share of the monthly one, at Starter 100, Pro 500 and Powerhouse 2,000, but no weekly buyer is charged against those numbers while the allowlist reads `web`.

A 20 credit Opus turn on a balance of 3 is refused whole rather than part charged, because charging 3 for it would make the same model cost different amounts depending on when you sent it. A generation that fails or is aborted is refunded by removing its usage row. The rounding and flooring rules behind that are in the [credits reference](https://whizi.io/docs/credits).

#### Using Claude without an Anthropic subscription

Whizi bills you, not Anthropic. There is no Anthropic account to connect and no Anthropic key to paste. Claude sits in the same model picker as the GPT, Gemini, Grok, DeepSeek, Llama, Mistral, Qwen and Kimi families, and the same credit allowance pays for all of them.

What that does not include is worth being straight about:

- **No developer API.** Whizi has no public API, so Claude in Whizi is a chat surface rather than something you can call from your own code.
- **No team accounts.** There is no seat-based or shared workspace plan.
- **A Whizi turn budget, not the full window.** A Claude turn is capped at 40,000 input tokens and 20,000 output tokens, so the 1M window on the rows above is a limit set by the model itself, not by what one Whizi turn can spend. Every Claude row priced in the index above carries a 200K window or larger, so none of them takes the reduced proportional budget that applies below a 93,000 token window.
- **No Claude image, video or music generation.** The image section is Whizi Image, Nano Banana, Flux and Stable Diffusion. Video is one row, Whizi Video AI, and audio is one row, Whizi Music AI. No Anthropic model appears in any of them.
- **The agent personas do not run on Claude.** No Anthropic model backs the Canadian CPA, study mentor, writing copilot or math tutor.

The Anthropic fast identifiers are excluded from the catalogue on purpose. They are the same base weights at a different latency tier for two to five times the money, so those rows carry no credit rung and cannot be selected at any tier.

If you are moving over, Claude is one of the two chat histories Whizi imports, alongside ChatGPT. The export is parsed in your browser. Thinking blocks, tool calls and attachments are dropped, so only the text of each turn is imported.

#### Switching into and out of Claude mid conversation

A conversation is not pinned to one model. Each turn is charged at the credit cost of whichever model answered it, so drafting on a 1 credit model and finishing on Claude costs 1 credit per draft turn and 10 or 20 for the answer you keep.

One behavior to know when you switch with a file in the thread: only the most recent user message carrying attachments has its attachments forwarded to the provider. If you attached a document six turns ago and then switch to Claude, re-attach it on the turn you want Claude to read it.

Auto is the other route into Claude. Auto has no credit cost of its own, because a turn costs whatever the model Auto picked costs. The rung Auto reaches for code is Claude Sonnet 5 at 10 credits, and difficulty is sticky inside a thread, so a follow-up to a code answer cannot drop back off that Claude row onto something cheaper. The published ranges are 1 to 1 on free and Starter, 1 to 10 on Pro, and 1 to 20 on Powerhouse.

More on the mechanics in [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### When Claude will not open

Three refusals cover the plan and allowance gates, and each carries an exact message.

Selecting a Claude model above your tier returns HTTP 403 with the code `tier_upgrade_required` and the message "Upgrade your plan to use this model." That is the Opus class on a Pro plan, or any Claude model at all on Starter. See [model unavailable](https://whizi.io/docs/model-unavailable).

Running out of allowance returns HTTP 429 with the code `message_limit_reached` and the message "Your monthly message limit has been reached." Weekly plans say "weekly" instead. See [out of credits](https://whizi.io/docs/out-of-credits).

A free account that has spent its 7 lifetime messages gets HTTP 402 with the code `free_limit_reached` and the message "Your free messages are used up. Start a subscription to keep chatting."

The gate is server side, and it is decided by the model identifier alone rather than by the conversation it sits in. That is why the oldest Sonnet refuses on Pro while the three Sonnet releases named on the Pro list open at the same 10 credit rung. How the resolution order works is covered in the [models list](https://whizi.io/docs/models).

#### How the Claude credit costs were set

A rung is derived from a reference turn of 3,000 input and 800 output tokens, divided by the cost of that same turn on the 1 credit anchor model. Models that fall between rungs round up, because the cheap direction is the one that costs money. The legal rungs are 1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 20, 25, 30, 40, 50, 60, 80 and 100, and a model may not be priced anywhere else on the scale.

The Claude Sonnet class is one of the rows set by owner policy rather than arithmetic: it is charged 10 credits against a cost-true rung of 12.

One more thing the ladder guards against. A model identifier with no rung is charged 100 credits, the top of the scale, so an unpriced arrival surfaces as a support ticket rather than quietly running below cost. A build test fails if any catalogue model reaches production without a rung.

Credit pricing is enabled per client, and the production value of the allowlist is the web client. Off the allowlist the rungs on this page stop applying to what you are charged: an 80 credit Claude row costs one message against the older message allowance, exactly like a 1 credit row. The full scale across all families is in the [credits reference](https://whizi.io/docs/credits), and the per-plan catalogue is in the [models list](https://whizi.io/docs/models).

Key points:

- Every Claude model at 4 credits or more, with its credit cost and the plan that unlocks it
- What one Claude answer costs to run, priced on the same 100 row comparison index as the other families
- What a Whizi subscription does and does not cover on the Claude rows: no developer API, no team accounts, and a flat per turn token budget

Checklist:

- Pro is the lowest paid plan that reaches any Claude model
- The Sonnet class and Claude Haiku are the Pro set
- The Opus class and Claude Fable 5 require Powerhouse
- Claude Haiku is 4 credits, the Sonnet class is 10, Claude Opus 5 is 20, Claude Fable 5 is 50
- The 80 credit rung is the top rung in use, and only two Claude rows sit on it
- A free account can send a Claude message inside a lifetime allowance of 7 messages
- The flat turn budget is 40,000 input and 20,000 output tokens, and every Claude row in the price table carries a 200K window or larger
- A turn that does not fit the remaining balance is refused whole, not part charged

FAQ:

Q: Can I use Claude without an Anthropic subscription?
A: Yes. Pro at $29.99/month reaches the Sonnet class and Claude Haiku, and Powerhouse at $49.99/month adds the Opus class and Claude Fable 5. The one thing no Whizi plan reaches is the Anthropic fast latency tier: those identifiers carry no credit rung and are not selectable at any tier, so paying up does not put them in the picker.

Q: How many credits does a Claude message cost in Whizi?
A: Claude Haiku is 4 credits, the Sonnet class is 10, the current Opus class is 20, Claude Fable 5 is 50, and the oldest Opus rows are 80. No Claude row sits at 5, 6, 8, 12, 15 or 25 credits. The rungs for every other family are in the [credits reference](https://whizi.io/docs/credits).

Q: Which Whizi plan do I need for Claude Opus?
A: Powerhouse, at $49.99/month or $34.99/mo billed annually at $419.88. Two more Claude rows sit above the Pro line with the Opus class: Claude Fable 5 at 50 credits, and the oldest Sonnet, which costs the same 10 credits as the Sonnet releases Pro can open but is not named on the Pro list.

Q: Does Claude in Whizi have the full 1M context window?
A: Not in one turn. The number to hold on to is the turn budget: 40,000 input tokens and 20,000 output tokens, the flat allocation every catalogue model gets. One thing that surprises people comparing windows: OpenRouter counts input plus the requested maximum output against a window, not input alone, so the headline figure is not what a single request has free to spend.

Q: Can I import my Claude chat history into Whizi?
A: Yes, within per request caps: 25 conversations at a time, and 400 messages per conversation, with the remainder sent as append slices onto the chat the first slice created. A raw .json export is accepted up to 300 MB, past which you are told to use the .zip export instead, which has no size ceiling of its own. Imports are keyed by conversation, so re-importing the same export or clicking Import twice lands the same chats once. If an import fails, see [chat import errors](https://whizi.io/docs/chat-import-errors).

Q: What happens when a free account runs out of Claude messages?
A: Chat stops until you subscribe, and the refusal is the `free_limit_reached` one covered above. The part worth knowing before you register: a guest profile's usage rows are carried into the real account when the guest is claimed, so spending the allowance as a guest and then creating an account does not hand you a fresh one.

### Using DeepSeek in Whizi: models, credit costs and plan access

URL: https://whizi.io/docs/using-deepseek-in-whizi/
Updated: August 2026

Quick answer: Whizi carries DeepSeek models on Pro and Powerhouse, and Starter reaches none. Three DeepSeek identifiers are reachable on Pro, and any DeepSeek identifier outside that set resolves to Powerhouse. The DeepSeek rows the Whizi cost index prices cost 1 or 2 credits per message. A free account can try any text model inside 7 lifetime messages.

The DeepSeek models Whizi carries, what each one costs per message in credits, the plan that unlocks it, and the context window and per turn token budget behind it.

#### The short answer

Starter reaches no DeepSeek model. Three DeepSeek identifiers are reachable on Pro, and any DeepSeek identifier outside that set resolves to Powerhouse. A free account is not model gated at all: it can try any text model, DeepSeek included, inside a lifetime allowance of 7 messages.

The DeepSeek rows carried in the Whizi cost index, with what each costs to answer once and what it costs you in credits:

| Model | Context window | Credits per message | USD per standard answer |
| --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 | 1,310,720 tokens | 1 | $0.00028 |
| DeepSeek V3.2 | 164K tokens | 1 | $0.000469 |
| DeepSeek V3.1 | 164K tokens | 1 | $0.000725 |
| DeepSeek R1 | 64K tokens | 2 | $0.00195 |
| DeepSeek V4 Pro 0813 | 1M tokens | 2 | $0.00297 |

The USD column is one standard answer of 1,000 input and 500 output tokens, at OpenRouter rates read on 20 August 2026. Per thousand answers those rows run from $0.28 to $2.97.

#### Where the DeepSeek rows sit against the rest of the index

The cost index prices 100 rows, not the catalogue. Against that index, the DeepSeek rows land as follows.

| Index reference point | Model | USD per standard answer |
| --- | --- | --- |
| Cheapest priced row overall | Ling-3.0-flash, inclusionAI | $0.000053 |
| Cheapest DeepSeek row | DeepSeek V4 Flash 0731 | $0.00028 |
| Median priced row | Kimi K2 Thinking | $0.00185 |
| Priciest DeepSeek row | DeepSeek V4 Pro 0813 | $0.00297 |
| Priciest priced row overall | Claude Opus 4.7 (Fast) | $0.105 |

The spread between the cheapest and the priciest priced row in the index is roughly 2000x, and every DeepSeek row sits in the lower part of that range. The largest context window anywhere in the price index belongs to a DeepSeek V4 Flash row at 1,310,720 tokens. 47 of the 100 priced rows carry a window of 1,000,000 tokens or more.

#### What a DeepSeek message costs in credits

A model credit cost is a fixed integer per message. Nothing about the length of your message, the length of the answer, or the length of the conversation enters the calculation, so a one line question on a DeepSeek row costs exactly what a long one does. Every DeepSeek row priced in the index sits on the first or second rung of the credit ladder. The ladder itself is in the [credits reference](https://whizi.io/docs/credits).

What each plan allowance buys in DeepSeek messages:

| Plan | Monthly credits | DeepSeek rows reachable | Messages on a 1 credit row | Messages on a 2 credit row |
| --- | --- | --- | --- | --- |
| Free | 0 credits, 7 message lifetime allowance | Any text model, inside those 7 messages | 7 in total, across all models | 7 in total, across all models |
| Starter | 400 | None | Not available | Not available |
| Pro | 2,000 | Three identifiers | 2,000 | 1,000 |
| Powerhouse | 8,000 | All of them | 8,000 | 4,000 |

That credit arithmetic is what a credit metered platform sees. Credit pricing is gated per platform and the production allowlist today is the web client. A platform that is not on the allowlist is charged one per turn and spends against the older message allowance instead, which is 400 messages a monthly period on Starter, 800 on Pro and 5,000 on Powerhouse. Message limits and credit limits coexist, and which one applies is decided per request.

Weekly billing, which is offered on mobile only, carries its own allowances: 100 credits on Starter, 500 on Pro and 2,000 on Powerhouse per weekly period. There is no grant event and no top up job behind any of this. Usage is summed per period key, so the arrival of a new period is itself the reset.

#### Which plan opens which DeepSeek model

Model access is resolved on the server by identifier, checked against the Starter set, then the agent set, then the Pro set, and otherwise returned as Powerhouse.

**Starter includes no DeepSeek model.** Starter reaches exactly four picker entries: Auto, the house model Whizi AI, one base OpenAI model and one fast Gemini model. Starter is $15.99/month, or $10.99/mo billed annually at $131.88. Pro is the lowest plan that reaches any DeepSeek model.

**Pro opens three DeepSeek identifiers.** Two are V3 generation rows and one is an R1 reasoning snapshot, written out by identifier in the table below. What the DeepSeek promotion into Pro selected for is the V3 generation plus a single reasoning row, not the family as a whole. Pro is $29.99/month, and it is a strict superset of Starter.

**Everything else in the family needs Powerhouse.** Any DeepSeek identifier that is not one of those three requires Powerhouse, which is $49.99/month. The V4 generation rows in the table above are outside the Pro list, so the newest DeepSeek rows and the largest DeepSeek window are the ones that need the top plan.

| Plan | DeepSeek identifiers it opens |
| --- | --- |
| Starter | None |
| Pro | `deepseek deepseek-v3.2`, `deepseek deepseek-chat-v3.1`, `deepseek deepseek-r1-0528` |
| Powerhouse | Every other DeepSeek identifier in the catalogue |

A free account is the one case with no model gate. The chat path returns for the free tier before the model gate is reached, because the funnel deliberately lets prospects try any text model inside the 7 message cap.

#### The DeepSeek row where the token budget is clamped

Within the DeepSeek rows, only one sits under the clamp threshold. Any model whose context window sits below 93,000 tokens gets a smaller, proportional per turn budget instead of the flat one, and the R1 row at 64K is the DeepSeek row under that line. Every other DeepSeek row in the table above sits far above it.

On a 64,000 token window the clamp caps output at 40 percent of the window, which is 25,600 tokens, and holds back a 1,000 token safety margin. If you are pasting long documents into a reasoning conversation, that is the row where you will feel it first.

The unclamped rows all share one allowance. A catalogue model gets a flat budget of 40,000 input tokens and 20,000 output tokens per turn unless its window is under 93,000 tokens, so the 1,310,720 token window on the V4 Flash row and the 164K windows on the V3 rows draw the same per turn ceiling. A context window is conversation headroom, not a bigger per turn budget.

#### What this does not include

No DeepSeek row generates media. The image, video and audio sections of the picker are hand curated lists, and the rows in them are Whizi Image, Nano Banana, Flux and Stable Diffusion for images, Whizi Video AI for video, and Whizi Music AI for audio. None of them is a DeepSeek model.

No DeepSeek row is an agent persona. The four personas plus their three legacy aliases all run on the same base model and are priced separately from the catalogue, and personas are a Pro feature rather than a model tier.

Auto never picks DeepSeek. Auto carries six rungs: a quick rung on the house model at 1 credit, a rewrite rung and an attachment rung on a fast Gemini model at 2, a long form rung on a mid GPT model at 4, a code rung on Claude Sonnet at 10, and a reasoning rung on a top GPT model at 20. No DeepSeek row is on that ladder, so reaching DeepSeek means selecting it in the picker.

There is no per model capability flag to publish. Web search is a request level mode and attachments are handled per request, so neither is a property of a DeepSeek row, and no source records a knowledge cutoff per model.

#### How the DeepSeek list changes over time

Two policies shape what gets added. Nothing older than a 2025 release is added to the catalogue any more, an owner rule from 19 August 2026. And deep research identifiers are excluded on principle, because they bill a whole multi search run rather than a turn.

The picker contents are served live by the Whizi backend, so the app list is always the current one. Older DeepSeek generations stay listed alongside the newer ones because a model never leaves the catalogue once it has been offered: a conversation carries the model it was created with, so removing a row would break the threads already on it.

Key points:

- The DeepSeek rows carried in the Whizi cost index, with credit cost and context window
- Why a DeepSeek message costs the same whether it is short or long
- Starter reaches four picker entries and none of them is a DeepSeek model

Checklist:

- Starter includes no DeepSeek model
- Pro opens three DeepSeek identifiers, Powerhouse opens the rest
- The priced DeepSeek rows cost 1 or 2 credits per message
- Credit cost is fixed per model, not per message length
- A free account can try any text model inside 7 lifetime messages
- Catalogue models run on a flat 40,000 in, 20,000 out per turn budget unless their window is under 93,000 tokens
- DeepSeek R1 at 64K is clamped below that flat budget

FAQ:

Q: How many credits does a DeepSeek message cost?
A: 1 or 2 credits on the rows the cost index prices. The consequence worth planning around is the refusal rule: a turn that does not fit in the remaining balance is refused whole rather than part charged, so 1 credit left is not enough to send a 2 credit DeepSeek turn. Running out returns HTTP 429 with the code `message_limit_reached`.

Q: Can I use DeepSeek on the Starter plan?
A: No. The gate runs on the server by model identifier rather than in the interface, so a locked row cannot be reached by editing a request: selecting one returns HTTP 403 with the code `tier_upgrade_required` and the message "Upgrade your plan to use this model."

Q: Which DeepSeek model has the largest context window in Whizi?
A: The DeepSeek V4 Flash row, at 1,310,720 tokens, which is the largest window of any row in the Whizi price index. It resolves to Powerhouse, so the largest window in the index is not one of the rows a Pro subscription opens.

Q: Does a DeepSeek reasoning answer cost more than a short one?
A: No. What a long answer costs you is room, not credits: the R1 row carries the smallest window in the family at 64K, which puts it under the 93,000 token clamp threshold, so its per turn output is capped at 25,600 tokens with a 1,000 token margin held back while the V3 and V4 rows draw the full flat budget. Moving that thread onto a V3 or V4 row is a per message decision, and each message is charged at the rung of whichever row answered it.

Q: Can DeepSeek read a PDF in Whizi?
A: In the web app, PDF, Word and spreadsheet text is extracted in your browser and sent as text to whichever model you selected, so a DeepSeek row receives text rather than a binary file. That extracted text spends the same per turn input budget as anything else you type, so what bounds a long document is the per turn ceiling rather than the size of the file you dropped in. The accepted formats and how each one is handled are in [supported file types](https://whizi.io/docs/supported-file-types).

### Using Gemini in Whizi: models, credit costs and plan gates

URL: https://whizi.io/docs/using-gemini-in-whizi/
Updated: August 2026

Quick answer: Whizi carries Google Gemini models inside the same subscription as GPT and Claude, so no separate Google AI subscription is needed. One Gemini Flash row is on Starter at 1 credit per message, seven further Google identifiers are on Pro, and every remaining Gemini identifier resolves to Powerhouse. Priced Gemini rows cost 1 to 10 credits.

How Google Gemini works in Whizi: what the priced rows cost in credits, the plan that unlocks them, and how to reach Gemini without a Google AI subscription.

#### The short answer

Whizi carries Google Gemini models inside the same subscription as GPT, Claude and the rest of a 280+ model catalogue, so no separate Google AI subscription is needed. One Gemini Flash row is on Starter at 1 credit per message, seven further Google identifiers are on Pro, and everything else in the family resolves to Powerhouse.

What a Gemini message costs, at a glance:

| Plan | Gemini rows it opens | Credits per message | Monthly credits |
| --- | --- | --- | --- |
| Free | Any text model, inside a 7 message lifetime allowance | not metered in credits | 0 |
| Starter | Gemini 3 Flash | 1 | 400 |
| Pro | Seven Google identifiers, listed below | 1 to 10 | 2,000 |
| Powerhouse | Every remaining Gemini identifier in the catalogue | varies by row | 8,000 |

Starter is $15.99/month, or $10.99/mo billed annually at $131.88. Pro is $29.99/month, or $19.99/mo billed annually at $239.88. Powerhouse is $49.99/month, or $34.99/mo billed annually at $419.88. New accounts start with a 7 day trial for $0.99.

#### The priced Gemini rows, with credit cost and plan

The Gemini rows below are the ones this page covers, ordered cheapest first. The credit column is what a single message on that row spends from your monthly allowance. The last two columns come from the Whizi Cost Index, which prices a "standard answer" of 1,000 input tokens plus 500 output tokens so that list rates from different providers become comparable. Those prices were read from OpenRouter on 20 August 2026.

| Model | Credits per message | Plan required | Context window | Cost per standard answer |
| --- | --- | --- | --- | --- |
| Gemini 3 Flash (`gemini-3-flash-preview`) | 1 | Starter | not in the price index | not in the price index |
| Gemini 2.5 Flash Lite | 1 | Pro | 1M | $0.0003 |
| Gemini 3.1 Flash Lite | 1 | Powerhouse | 1M | $0.001 |
| Gemini 3.7 Flash | 2 | Pro | 1M | $0.001313 |
| Gemini 2.5 Flash | 2 | Pro | 1M | $0.00155 |
| Gemini 3.5 Flash Lite | 2 | Powerhouse | 1M | $0.00155 |
| Gemini 3.6 Flash | 3 | Pro | 1M | $0.002625 |
| Gemini 3.5 Flash | 8 | Pro | 1M | $0.006 |
| Gemini 2.5 Pro | 8 | Pro | not in the price index | not in the price index |
| Gemini 3.1 Pro Preview | 10 | Pro | 1M | $0.008 |
| Gemini 3 Pro Image Preview | 10 | Powerhouse | not in the price index | not in the price index |

One caveat on how to read that table. The Cost Index prices 100 rows for comparison, not the whole catalogue, so a blank price cell means the row was not in that sample and not that the model is free. How the credit figure itself is arrived at is set out in the [credits reference](https://whizi.io/docs/credits).

One row is priced by owner policy rather than by arithmetic. The Starter Gemini Flash identifier is charged 1 credit against a cost-true rung of 2, which is what allows a 400 credit Starter allowance to be 400 Gemini messages rather than 200.

The median row in the price index answers for $0.00185. Most of the Gemini Flash rows above sit at or under that figure, and the one priced Pro-class Gemini row answers for $0.008 against $0.105 for the priciest row in the index.

#### What each plan opens, by identifier

The gate reads the model identifier and returns a minimum tier, which is why the Starter and Pro sides below can be written out identifier by identifier at all. The [models reference](https://whizi.io/docs/models) covers the enforcement model; what follows is only the Google half of it.

**Starter** reaches four identifiers in total, one of which is Google: `auto`, `whizi-ai`, `openai gpt-5.6-luna`, and `google gemini-3-flash-preview`. So on Starter, Google is exactly one fast Gemini row and nothing else.

**Pro** folds in the whole Starter set and adds 37 further identifiers, of which seven are Google: `gemini-3.7-flash`, `gemini-3.6-flash`, `gemini-3.5-flash`, `gemini-3.1-pro-preview`, `gemini-2.5-pro`, `gemini-2.5-flash`, and `gemini-2.5-flash-lite`. That is the current Flash tier, the Flash Lite row, and the Pro-class rows.

**Powerhouse** is the fallback. Any identifier that is not in the Starter set, not an agent persona, and not in the explicit Pro list resolves to Powerhouse, and that includes every Gemini identifier the catalogue gains until it is deliberately promoted into the Pro list. That is the safe direction rather than an oversight.

If you select a Gemini row above your tier, the request is refused server-side with HTTP 403, code `tier_upgrade_required`, and the message "Upgrade your plan to use this model." See [why a model shows as unavailable](https://whizi.io/docs/model-unavailable).

#### Using Gemini without a Google AI subscription

Whizi pays the providers and bills you once, so a Whizi plan is the only subscription involved. There is no Google account to connect and no Google AI key to paste.

Before paying anything, a free Whizi account gets a lifetime allowance of 7 messages that never resets. That allowance is deliberately not model-gated: the funnel lets prospects try any text model inside the cap, so the Gemini rows are reachable on a free account even though the free credit limit and free message limit are both 0. When the allowance is gone the reply is HTTP 402, code `free_limit_reached`, with the message "Your free messages are used up. Start a subscription to keep chatting."

After that, what a Gemini allowance actually buys:

| Plan | Monthly credits | Messages on the 1 credit Gemini row | Messages on a 2 credit Gemini row | Messages on the 10 credit Gemini Pro row |
| --- | --- | --- | --- | --- |
| Starter | 400 | 400 | not included | not included |
| Pro | 2,000 | 2,000 | 1,000 | 200 |
| Powerhouse | 8,000 | 8,000 | 4,000 | 800 |

Credits do not roll over: the balance sums only the rows carrying the current period key, so last period's unspent credits are unreachable. The [credits reference](https://whizi.io/docs/credits) has the mechanism.

#### Where Gemini runs without you picking it

**Auto routing.** Two of Auto's six rungs are Gemini Flash rows. Any turn carrying an attachment is routed to Gemini Flash at 2 credits, on the reasoning that an image needs vision rather than a large context window, and the rewrite rung is the same Gemini Flash row at 2 credits. Auto is a router, not a model, so it carries no rung of its own: a turn costs whatever the picked model costs. Starter cannot reach that 2 credit Gemini Flash row, so a Starter account that sends an attachment is answered on the house model at 1 credit instead. The per-tier ranges are free 1 to 1, Starter 1 to 1, Pro 1 to 10, Powerhouse 1 to 20.

**Voice mode.** Whizi voice runs on the Gemini Live API. The worker mints a short-lived session token and the browser opens a WebSocket straight to Google, so no audio passes through Whizi's own backend. Sessions are capped at 15 minutes and the token is minted with a 16 minute lifetime, one minute of slack past the cap. Replies use a single warm preset voice. Voice is metered in minutes rather than credits: Starter 10 minutes per month, Pro 80, Powerhouse 500, and free accounts 0. The weekly allowance is the monthly figure divided by four and rounded up, which gives Starter 3 minutes, Pro 20 and Powerhouse 125. See [voice mode will not start](https://whizi.io/docs/voice-mode-not-working).

#### Two Gemini facts inside the picker

**The Recommended row.** The top section of the model picker is the house model, Auto, and one current row each from GPT, Claude and Gemini. That section is constrained to models reachable at Pro or below, so the top of the picker is never Powerhouse-locked.

**Image generation is separate.** The Google rows in the Image Generation section of the picker carry no per-message credit rung: they are priced in the Cost Index but absent from the credit table, as are the other standalone image generators in that section. The Gemini image preview row in the model table above is not one of them and does carry a rung, at 10 credits. Media generators skip the per-model gate and are tiered by picker section instead, against a monthly image count: free 0, Starter 0, Pro 100, Powerhouse 500. Generated media is retained for 30 days.

#### What a 1M context window does and does not buy

Almost every Gemini row in the table above carries a 1,000,000 token context window, and that is the provider's figure, not the per-turn budget Whizi sends. What the large window buys on a Gemini row is that it clears the 93,000 token line by a wide margin, so it always gets the flat per-turn budget rather than the smaller proportional one Whizi applies to short-window models. It does not mean a single message can carry a million tokens. The budget itself is set out in [when a conversation is too long](https://whizi.io/docs/context-too-long).

Two details that change the arithmetic. OpenRouter counts input plus requested maximum output against a model's window, not input alone. And Whizi's token estimator can undercount a real tokenizer by up to 1.66x on JSON, so an allowance of 1.8x is applied to input before a request goes out. Both push in the conservative direction.

Web search is a per-request toggle rather than a per-model capability flag. How the toggle works is covered in the [web search reference](https://whizi.io/docs/web-search), and [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation) covers moving a thread from Gemini to another family.

Key points:

- The credit cost, plan gate, context window and per-answer cost of the priced Gemini rows
- What Starter and Pro each open on the Google side, and why everything else lands on Powerhouse
- Where Gemini runs inside Whizi outside the model picker, including voice mode

Checklist:

- A Whizi plan is the only subscription needed, with no Google account to connect
- Starter opens one Gemini Flash row at 1 credit per message
- Pro is the tier that reaches the Pro-class Gemini rows
- Every other Gemini identifier resolves to Powerhouse by default
- Auto routes any turn carrying an attachment to a Gemini Flash row at 2 credits
- Voice mode runs on the Gemini Live API and is metered in minutes, not credits
- The Google rows in the Image Generation section carry no credit rung and draw on the monthly image count
- A free account can try any text model inside a 7 message lifetime allowance

FAQ:

Q: Can I use Gemini without a Google AI subscription?
A: Yes. A Whizi subscription is the only plan involved. Starter at $15.99/month opens a Gemini Flash row at 1 credit per message, and Pro adds the Flash tier, the Flash Lite row and the Pro-class Gemini rows.

Q: How many credits does a Gemini message cost in Whizi?
A: From 1 to 10, depending on the row. The Starter Gemini Flash identifier is 1 credit, the Flash and Flash Lite rows run from 1 to 8, and the Pro-class rows sit at 8 and 10. The per-row figures are in the table above.

Q: Which Whizi plan do I need for Gemini Pro?
A: Pro. The Pro-class Gemini rows sit in Pro's Google set, alongside the Flash tier and the Flash Lite row, and the exact identifiers are listed above. Starter reaches one Gemini Flash row and no Pro-class Gemini model, so a Starter account that selects one gets HTTP 403 with the message "Upgrade your plan to use this model."

Q: Do I get the full 1 million token context window on Gemini in Whizi?
A: No. The window is the provider's figure, and it is 1,000,000 tokens on almost every Gemini row in the table above, but what Whizi sends per turn is its own budget: a flat 40,000 input tokens and 20,000 output tokens. What happens once a thread outgrows that is in [when a conversation is too long](https://whizi.io/docs/context-too-long).

Q: Is Whizi voice mode running on Gemini?
A: Yes. Voice runs on the Gemini Live API, metered in minutes rather than credits: 10 per month on Starter, 80 on Pro, 500 on Powerhouse, and 0 on a free account.

Q: Does Whizi list which Gemini models can read images or browse the web?
A: No, and there is no list to publish, because the catalogue carries no per-model vision or search capability field. Web search is a per-request toggle rather than a per-model flag, and per-model knowledge cutoff dates are not recorded either. Auto routing settles the vision question in one direction on its own: any turn carrying an attachment is routed to a Gemini Flash row.

### Using GLM in Whizi: one row on Pro, the rest on Powerhouse

URL: https://whizi.io/docs/using-glm-in-whizi/
Updated: August 2026

Quick answer: GLM is in the Whizi catalogue, with Z.ai rows priced across the GLM 4 and GLM 5 generations. One row, GLM 5, opens on Pro at 2 credits per message, and every other GLM row needs Powerhouse. Starter includes no GLM model. A GLM message costs between 1 and 5 credits.

The GLM models Whizi prices, the credit cost of each, the plan that unlocks each one, and why the cheapest GLM row in credits is not the cheapest one to run.

#### The short answer

Yes, GLM is in the Whizi catalogue. Z.ai rows are priced across the GLM 4 and GLM 5 generations, one row opens on the Pro plan at 2 credits per message, and every other GLM row needs Powerhouse.

| Model | Context window | Credits per message | Plan required |
| --- | --- | --- | --- |
| GLM 4.6 | 205K | 1 | Powerhouse |
| GLM 4.7 | 205K | 2 | Powerhouse |
| GLM 5 | 205K | 2 | Pro |
| GLM 5.1 | not in the price index | 3 | Powerhouse |
| GLM 5.2 | 1M | 3 | Powerhouse |
| GLM 5.3 | 1M | 5 | Powerhouse |
| GLM 5 Turbo | not in the price index | 5 | Powerhouse |
| GLM 5V Turbo | not in the price index | 5 | Powerhouse |

A GLM identifier not listed here is still charged a rung and still requires Powerhouse.

Free accounts are not model-gated. The chat path returns for the free tier before the model gate runs, so a prospect can try any text model in the catalogue, GLM included, inside a lifetime allowance of 7 messages that never resets.

#### The one row on Pro is not the cheapest row

Pro resolves to exactly one Z.ai identifier, GLM 5, at 2 credits per message. The curation rule the Pro list is built on is set out in the [model reference](https://whizi.io/docs/models), and on this family it lands on a single row, which is the whole reason the rest of the generation sits a tier up.

That single row is not the cheap one. The cheapest GLM row in credits charges 1 credit per message and requires Powerhouse, while the Pro row charges twice that. Credit cost and plan tier are two separate decisions in Whizi, and neither one predicts the other.

**Starter does not include any GLM model.** Starter reaches exactly four picker entries: Auto, the house model, one OpenAI model and one Google model. Every one of them is a 1x or 2x row, which is why the Starter credit allowance and the Starter message allowance are the same number.

**Powerhouse covers the rest, by name in one case and by fallback in every other.** The frontier reasoning row in the GLM 5 generation is named explicitly among the models held back for Powerhouse, alongside a Grok row, the Qwen thinking row, the Kimi thinking row and others. Every remaining GLM row lands on the top tier because any identifier that is not in the Starter set and not in the explicit Pro list resolves there.

Gating is enforced on the server, not in the interface. Selecting a model above your tier returns HTTP 403 with the code `tier_upgrade_required` and the message "Upgrade your plan to use this model." If you hit that on a GLM row, see [why a model shows as unavailable](https://whizi.io/docs/model-unavailable).

#### The GLM 4 pair, where the rung and the running cost disagree

In the GLM 4 pair the two columns disagree. The row that is cheaper per standard answer carries the higher rung, and its sibling carries the lower one. A rung is measured on a reference turn of 3,000 input and 800 output tokens, the index figure on a 1,000 plus 500 standard answer, and two rows whose input and output rates are weighted differently can swap places between those two mixes. The rung method is in the [credits reference](https://whizi.io/docs/credits).

| Model | Cost per standard answer | Cost per thousand answers | Credits per message |
| --- | --- | --- | --- |
| GLM 4.7 | $0.001275 | $1.275 | 2 |
| GLM 4.6 | $0.0015 | $1.50 | 1 |
| GLM 5 | $0.00156 | $1.56 | 2 |
| GLM 5.2 | $0.002484 | $2.484 | 3 |
| GLM 5.3 | $0.0036 | $3.60 | 5 |

The Whizi cost index prices 100 OpenRouter rows on one identical unit so providers can be compared: a standard answer of 1,000 input tokens plus 500 output tokens. Source prices were fetched from OpenRouter on 2026-08-20, across 29 providers.

For scale: the median priced row in the index is $0.00185 per standard answer, the cheapest is $0.000053 and the dearest is $0.105, a spread of roughly 2000x. Three of the five priced GLM rows sit below that median, and two sit above it.

Rows marked "not in the price index" in the table at the top carry a credit charge in the product but are not among the 100 rows the cost index prices, so there is no per answer figure to publish for them.

#### What a GLM message spends from your allowance

A credit is one message on the house model, and credit cost is fixed per message by the row you picked rather than by length. The method is in the [credits reference](https://whizi.io/docs/credits). What it means on this family: the GLM rows listed at the top of this page span a single band from 1 credit to 5, so the dearest of them costs exactly five of the cheapest, and the whole family sits inside the bottom five rungs of an eighteen rung ladder.

| Plan | Price | Credits per month | Messages on the Pro GLM row (2 credits) | Messages on the dearest GLM row (5 credits) |
| --- | --- | --- | --- | --- |
| Starter | $15.99/month, or $10.99/mo billed annually at $131.88 | 400 | no GLM row on this plan | no GLM row on this plan |
| Pro | $29.99/month, or $19.99/mo billed annually at $239.88 | 2,000 | 1,000 | not on this plan |
| Powerhouse | $49.99/month, or $34.99/mo billed annually at $419.88 | 8,000 | 4,000 | 1,600 |

On Powerhouse the 1 credit GLM row buys 8,000 messages from an 8,000 credit allowance, and the 3 credit rows buy 2,666. Credits do not roll over: usage is summed against a period key, so the arrival of a new period is itself the reset.

A GLM turn that will not fit the remaining balance is rejected outright instead of being charged in part. That is deliberate, because charging 3 credits for a 5 credit row would make one model cost different amounts depending on when it was sent.

Credit pricing is gated per platform by an allowlist whose production value today is the web client. Off that allowlist a GLM row shows a flat 1x in the picker, and every turn costs one against the older message allowance rather than a rung against the credit allowance.

#### Why the 205K rows and the 1M rows are sized the same

A GLM row with a 1M window and a GLM row with a 205K window get identical room inside Whizi. Both sit above the 93,000 token threshold, and every catalogue model above that threshold gets one flat budget of 40,000 input tokens and 20,000 output tokens per turn, which is true of every GLM row the cost index prices. Models below the threshold get a smaller proportional budget instead, described in [when a conversation is too long](https://whizi.io/docs/context-too-long).

So moving up to the larger window does not buy a longer prompt. What it buys is headroom on the provider side, since OpenRouter counts input plus requested maximum output against the window rather than input alone.

#### Moving a thread onto GLM and off again

Auto will not take you into GLM. Auto has no rung of its own, and its ladder is six fixed routes across five model rows: the house model for short questions, a Google fast row for rewrites and for attachments, an OpenAI row for long form, an Anthropic row for code, and an OpenAI reasoning row for multi step reasoning. No GLM row is on that ladder, so a turn only lands on GLM when you pick it yourself.

Every model row served to the web client carries its cost signal, so the credit price of a GLM row is visible before you send rather than after.

A GLM thread you opened months ago still opens today, because a conversation keeps the identifier it was created with even after the picker has moved on to newer rows.

Whichever way you move, the charge follows the model that actually answered the turn, not the model the conversation started on.

#### Where GLM does not appear

**Media generation does not use GLM.** The image generation section is four fixed rows, video generation is a single row, and audio generation is a single row, all of them house or partner generators, so no GLM row is involved in generating an image, a video or music.

**There is no per model vision flag to quote.** Whizi does not track which models can read images. Attachments are handled per request rather than per model, and on the web client the text of a PDF, Word or spreadsheet file is extracted in the browser before anything is sent. See [supported file types](https://whizi.io/docs/supported-file-types) for what that means in practice.

**Web search is a request level toggle, not a model capability.** Search runs in one of three modes, and turning it on affects the request rather than selecting a different model, so there is no published list of GLM rows that can or cannot browse.

**Agent personas are not GLM.** The four agent personas are a Pro feature rather than a model tier, and every one of them runs on the same base model, so choosing a persona never routes a turn to Z.ai.

Key points:

- Which single GLM row opens on Pro, and why it is not the cheapest GLM row in credits
- The credit rung on every GLM row Whizi prices, and the plan each one needs
- Where the GLM 4 pair inverts, with the cheaper row to run carrying the higher rung

Checklist:

- Whizi prices GLM models from Z.ai across the GLM 4 and GLM 5 generations
- GLM 5 is the only GLM row reachable on Pro, at 2 credits per message
- Every other GLM row requires Powerhouse
- Starter includes no GLM model at all
- Free accounts can try any text model, GLM included, within a lifetime allowance of 7 messages
- Credit cost is fixed per message and does not vary with length
- Every GLM row the cost index prices runs on the flat 40,000 input and 20,000 output token budget
- Auto never routes a turn to a GLM row

FAQ:

Q: Is GLM included in Whizi?
A: Yes. GLM rows sit in the All models section of the picker, under the Z.ai vendor logo rather than the generic fallback icon that most smaller labs get. They are not in the Recommended section, which is five fixed rows and carries no GLM entry, so browse All models rather than the top of the picker to find them.

Q: Which Whizi plan do I need to use GLM?
A: Pro reaches one GLM row, GLM 5. Everything else in the family is Powerhouse. If you send on a row your tier does not include, the worker answers HTTP 403 with a JSON body shaped `{"error": {"code": "tier_upgrade_required", "message": "Upgrade your plan to use this model."}}`, and the website re-wraps that into its older code and message pair before the chat page reads it, so the sentence you see on screen is the backend string verbatim rather than a client rewrite of it.

Q: How many credits does a GLM message cost?
A: Between 1 and 5 credits depending on the row. The GLM 4 generation is 1 or 2 credits, GLM 5 is 2, two further rows in the GLM 5 generation sit at 3 credits, and the dearest GLM rows are 5.

Q: Can I use GLM on the free tier?
A: Yes, within a hard cap. Free accounts get a credit allowance of 0 and a message allowance of 0, but they do get a lifetime allowance of 7 messages that never resets, and the free path deliberately skips the model gate so a prospect can try any text model inside that cap. When it is used up the app returns the code free_limit_reached and the message "Your free messages are used up. Start a subscription to keep chatting."

Q: What is the context window on GLM in Whizi?
A: Before any budget is checked, the measured input is scaled by 1.8x, because the token estimator can undercount a real tokenizer by as much as 1.66x on content such as JSON. A GLM prompt heavy with structured data is therefore sized larger than it literally measures. The priced GLM rows carry 205K or 1M token windows depending on the row, and neither number is what one turn sends: the per turn budget is in the section on why the two window sizes come out the same.

### Using GPT in Whizi: models, credit costs and plans

URL: https://whizi.io/docs/using-gpt-in-whizi/
Updated: August 2026

Quick answer: GPT models are part of the Whizi catalogue and are reached with a Whizi subscription, not a ChatGPT one. The cheapest GPT row costs 1 credit per message and is included on Starter. Pro adds nine OpenAI entries, and anything outside those lists, including the Codex rows and the o-series, requires Powerhouse.

Which GPT models Whizi carries, what they cost in credits, which plan unlocks each one, and how to use GPT without a separate ChatGPT subscription.

#### The short answer

GPT models are part of the Whizi catalogue and are reached with a Whizi subscription, not a ChatGPT one. The cheapest GPT row costs 1 credit per message and is included on Starter. The rest sit on Pro or Powerhouse, and the most expensive GPT row in the catalogue costs 25 credits per message.

The default assistant, Whizi AI, is a house persona that runs on the base GPT model, so every paid plan reaches GPT from the first message. One credit is defined as one Whizi AI message, which means one credit is also one message on that base GPT model.

| Plan | Price | Monthly credits | GPT models reachable |
| --- | --- | --- | --- |
| Free | no subscription | 0 credits, 7 message lifetime allowance | any text model in the catalogue, inside that allowance |
| Starter | $15.99/month, or $10.99/mo billed annually at $131.88 | 400 | GPT-5.6 Luna, plus Whizi AI which runs on it |
| Pro | $29.99/month, or $19.99/mo billed annually at $239.88 | 2,000 | GPT-5.6 Terra, GPT-5.5, GPT-5.2, GPT-5 Mini, GPT-4.1, GPT-4.1 Mini, GPT-4o, GPT-4o Mini, GPT-OSS-120B |
| Powerhouse | $49.99/month, or $34.99/mo billed annually at $419.88 | 8,000 | everything above, plus every other GPT row in the catalogue |

#### What each GPT model costs

A credit cost is a fixed integer per model, and the mechanics of the scale are in [how credits work](https://whizi.io/docs/credits). The OpenAI rows in the catalogue run from 1 credit to 25, so the cheapest GPT answer and the dearest differ by a factor of twenty five on the same bill.

The dollar column is what one answer costs at provider list rates, measured on a standard answer of 1,000 input tokens plus 500 output tokens, using rates read on 2026-08-20. It is there so the credit column can be checked against something real, not because you are billed in dollars per message.

| Model | Context window | Credits per message | One standard answer | Plan |
| --- | --- | --- | --- | --- |
| GPT-5.6 Luna | 1M | 1 | $0.0008 | Starter |
| GPT-5.4 Nano | 400K | 2 | $0.000825 | Powerhouse |
| GPT-5.4 Mini | 400K | 4 | $0.003 | Powerhouse |
| GPT-5.6 Terra | 1M | 4 | $0.008 | Pro |
| GPT-5.4 | 1M | 15 | $0.01 | Powerhouse |
| GPT-5.6 Sol | 1M | 20 | $0.01 | Powerhouse |
| GPT-5.5 | 1M | 20 | $0.02 | Pro |
| GPT Chat Latest | 400K | 25 | $0.02 | Powerhouse |

Further GPT rows carry a credit cost but are not in that price set:

| Model | Credits per message | Plan |
| --- | --- | --- |
| GPT-4o | 1 | Pro |
| GPT-4.1 Mini | 2 | Pro |
| GPT-5 Mini | 2 | Pro |
| GPT-4.1 | 8 | Pro |
| GPT-5 | 8 | Powerhouse |
| GPT-5.1 | 8 | Powerhouse |
| GPT-5.1 Codex | 8 | Powerhouse |
| GPT-5.1 Codex Max | 8 | Powerhouse |
| GPT-5.2 | 10 | Pro |
| GPT-5.2 Chat | 10 | Powerhouse |
| GPT-5.2 Codex | 10 | Powerhouse |
| GPT-5.3 Codex | 12 | Powerhouse |

The OpenAI reasoning series sits on Powerhouse:

| Model | Credits per message | Plan |
| --- | --- | --- |
| o3-mini | 4 | Powerhouse |
| o4-mini | 4 | Powerhouse |
| o3-mini-high | 5 | Powerhouse |
| o4-mini-high | 5 | Powerhouse |
| o3 | 8 | Powerhouse |

Three OpenAI rows are charged below what the cost formula returns, by owner policy rather than arithmetic. The rung scale itself, and the rule for a model that lands between two rungs, are in [how credits work](https://whizi.io/docs/credits).

| Model | Charged | What the formula gives |
| --- | --- | --- |
| GPT-5.6 Terra | 4 credits | 10 |
| GPT-5.5 | 20 credits | 25 |
| GPT-5.6 Sol | 20 credits | 25 |

#### Which plan unlocks which GPT model

Access is resolved on the server by model identifier. On Starter the gate falls on every GPT row above the base 1 credit model. On Pro it falls on the OpenAI identifiers outside Pro's nine, which is where the Codex coding rows, the o-series and the top GPT row sit. In both cases the refusal is HTTP 403 with the code tier_upgrade_required and the message "Upgrade your plan to use this model."

**Starter** reaches exactly four picker entries: Auto, Whizi AI, the base GPT model, and one fast Gemini model. That is the whole list, so on Starter every GPT message costs 1 credit and nothing above the base row is available.

**Pro** adds nine OpenAI entries, the ones named in the short answer table above. The set pairs each included flagship with its mini variant, and adds the open weight GPT-OSS row. Pro stops there: the Powerhouse trophy GPT row, the Codex coding rows and the o-series are not in the set. Pro is a strict superset of Starter, so the Starter GPT row is included as well.

**Powerhouse** is the fallback tier, not Pro. Any model identifier that is not in the Starter set and not in the Pro list resolves to Powerhouse, which is why the coding specialist GPT rows and the o-series sit there. It also means a GPT row the catalogue gains arrives on Powerhouse and stays there until it is deliberately promoted, so the nine names in the Pro column are the whole of what Pro reaches on the OpenAI side, not a sample of it.

The dearest GPT row is held back on Powerhouse deliberately rather than by the fallback rule: it is the trophy the top tier sells. The picker shows the consequence. Its Recommended section is limited to models a Pro subscriber can already reach, so the OpenAI slot there is the Pro tier GPT row instead.

#### What an allowance buys in GPT messages

A GPT turn either fits in the remaining balance or it is refused outright: a 25 credit GPT turn on a balance of 20 credits is rejected, not discounted down to what is left.

| Plan | Monthly credits | At 1 credit | At 4 credits | At 20 credits |
| --- | --- | --- | --- | --- |
| Starter | 400 | 400 GPT messages | not included | not included |
| Pro | 2,000 | 2,000 | 500 | 100 |
| Powerhouse | 8,000 | 8,000 | 2,000 | 400 |

Weekly billing, offered on mobile only, carries its own numbers: Starter 100 credits, Pro 500, Powerhouse 2,000 per week.

Every model a Starter account can reach costs 1 credit, so on that tier 400 credits is 400 messages.

Credit pricing is enabled per platform. The web app is on the allowlist today and is charged the real multipliers against the credit allowance. A client that is not on the allowlist is charged one per turn against the older message allowance instead, which is 400 monthly on Starter, 800 on Pro and 5,000 on Powerhouse. A client outside the allowlist also draws no credit badges in the picker, because the cost field is absent rather than pinned to 1.

Running out returns HTTP 429 with the code message_limit_reached and the message "Your monthly message limit has been reached." Weekly plans say weekly. What to do about it is in [running out of credits](https://whizi.io/docs/out-of-credits).

#### Using GPT without a ChatGPT subscription

A free Whizi account gets a lifetime allowance of 7 messages that never resets, and the free tier is deliberately not model gated. Any text model in the catalogue can be tried inside that allowance, GPT included.

When the free allowance is spent, the response is HTTP 402 with the code free_limit_reached and the message "Your free messages are used up. Start a subscription to keep chatting."

After that, a paid plan is what unlocks the ongoing allowance. New accounts start with a 7 day trial charged at the intro rate of $0.99, after which the selected plan renews unless it is canceled.

Existing ChatGPT history can come with you. Whizi imports ChatGPT and Claude exports, and the export file itself never reaches the server: the browser parses it and posts normalized conversations. One request carries up to 25 conversations and 400 messages per conversation, and anything larger is split into follow up slices automatically. A raw .json export is accepted up to 300 MB, past which the .zip export is the route. Imports are idempotent, so clicking Import twice or resuming after a dropped connection lands the same chats once.

Two things do not survive a ChatGPT import: only the branch that was on screen is imported, so regenerated replies and discarded drafts are not replayed, and non transcript content is skipped, which covers hidden reasoning, analysis tool code and output, browsing scratchpads, custom instruction blocks and function calls. If an import stalls or reports skipped conversations, see [chat import errors](https://whizi.io/docs/chat-import-errors).

#### Moving between GPT and the other families

A credit cost is looked up per model per turn, so the price of the next answer is set by whichever model answers it. Moving a conversation from GPT to another family changes what the next turn costs and leaves the turns already sent alone. A conversation also carries the model it was created with, which is why a model is never removed from the catalogue once it has been offered.

Auto is the other way to reach GPT. It is a router row, not a model, and it resolves to a concrete model before anything is charged, so the credit charge, the tier gate and the stored history all record the model that actually ran. Its ladder is six rungs. The cheapest lands on the house model at 1 credit, and two land on OpenAI rows: long form at 4 credits and multi step reasoning at 20 credits.

Auto never routes above your plan. That behavior, and the other reasons a row you expect is missing from the picker or refuses to send, are in [model unavailable](https://whizi.io/docs/model-unavailable).

#### What a GPT model in Whizi does not do

The context window column above is the provider window, not your per turn budget. Most catalogue models get a flat budget of 40,000 input tokens and 20,000 output tokens per turn, and a model whose window sits below 93,000 tokens gets a smaller proportional budget instead, with output capped at 40 percent of the window and a 1,000 token safety margin held back. Every GPT window in the first table above is 400K or 1M, far above that threshold, so those rows all take the flat budget: a million token window and a 400K one hand you the same 40,000 tokens per turn. A conversation that outgrows the budget is covered in [context too long](https://whizi.io/docs/context-too-long).

Web search is not a GPT capability flag. It is a per request toggle with three modes, applied to whichever model the turn is on, so there is no list of GPT models that can or cannot browse. With the toggle on, the request first carries a cheap probe of about 50 tokens, and only a turn where the model actually calls it is re-issued with real search. Native provider search injects a fixed preamble of roughly 4,400 input tokens, which is why the probe exists.

Sending speed is not a GPT setting either. Chat is capped at 10 messages per minute and 60 per hour per user, and those two ceilings are identical on every tier, so no GPT row and no upgrade moves them. The full set of windows is in [rate limits](https://whizi.io/docs/rate-limits).

Key points:

- The credit cost and plan gate for the GPT rows Whizi prices, with context windows for the eight in the cost index
- What an allowance buys across the 1, 4 and 20 credit OpenAI rows, and why Starter is the flat case
- The free allowance that lets you send GPT messages before you pay anything

Checklist:

- GPT is reached with a Whizi subscription, not a ChatGPT one
- The base GPT row costs 1 credit per message and is included on Starter
- Pro adds nine OpenAI entries, including the 4 credit and 20 credit rows
- Anything not in the Starter or Pro list needs Powerhouse, including the coding specialist rows
- A free account can try any text model, GPT included, inside a 7 message lifetime allowance
- Credit cost is fixed per model per message, whatever the message length
- Most models get 40,000 input and 20,000 output tokens per turn, and a model with a window under 93,000 tokens gets a smaller proportional budget

FAQ:

Q: What happens to my 7 free messages when I subscribe?
A: The free allowance is a lifetime one, so it never resets and a subscription does not top it up. It is replaced by a plan allowance that is summed per billing period, and on a monthly plan the period is the UTC calendar month, so the reset is the month boundary and unspent credits do not carry into the next one. Registering does not reset the free count either, because usage from a guest session is carried into the account when it is claimed.

Q: How many GPT messages do I get per month?
A: A real month is a mix of rows rather than one of them, so the arithmetic is subtraction. On Pro, 50 turns on a 20 credit GPT row spend 1,000 of the 2,000 credits and leave 1,000, which is another 250 turns on the 4 credit row or 1,000 on a 1 credit row. On Powerhouse the same 50 expensive turns leave 7,000 credits.

Q: Why does one GPT model cost 1 credit and another 25?
A: Because the rung tracks what the model costs to run. One credit is one message on the house model, and every other rung is derived from live provider per token pricing measured on the same reference turn. Inside the OpenAI family that puts the cheapest rows at 1 credit and the dearest at 25, and the tables above name which row is which. The full scale is in [how credits work](https://whizi.io/docs/credits).

Q: Can I switch from GPT to Claude in the same conversation?
A: Yes, and the thing that catches people out is files rather than price. Only the most recent user message carrying attachments has its attachments forwarded to the provider, so a document you attached six turns ago is not resent when Claude answers. Re-attach it on the turn you want Claude to read it. On the plan side, Anthropic models start on Pro, where Sonnet is the ceiling, and the Opus class requires Powerhouse.

### Using Grok in Whizi: models, credit costs and plan gates

URL: https://whizi.io/docs/using-grok-in-whizi/
Updated: August 2026

Quick answer: Whizi carries Grok models from xAI, and what you can open is decided by your Whizi plan alone. Exactly one Grok model is reachable on Pro, at 6 credits per message. Every other Grok identifier requires Powerhouse, and Starter includes no Grok model. The Grok rows documented here cost 3 to 6 credits.

The Grok models in the Whizi price index, what each costs in credits per message, which plan unlocks it, and how to reach Grok through a Whizi subscription.

#### The short answer

Whizi carries Grok models from xAI inside its 280+ model catalogue, and what you can open is decided by your Whizi plan alone. Exactly one Grok model is reachable on Pro. Every other Grok identifier in the catalogue requires Powerhouse, and Starter includes no Grok model at all.

| Plan | Grok access | Credits per month |
| --- | --- | --- |
| Free | Any text model, inside a lifetime cap of 7 messages | 0 |
| Starter | None | 400 |
| Pro | One Grok model, named in the table below | 2,000 |
| Powerhouse | Every Grok identifier in the catalogue | 8,000 |

Free accounts are deliberately not model gated: the chat path returns for the free tier before the model gate is reached, so a prospect can try any text model, Grok included, inside a lifetime allowance of 7 messages that never resets.

#### The Grok rows, with credits and plan

Credits are what you spend from your monthly allowance. The cost per standard answer is provider list pricing, shown so the rungs can be checked against something outside Whizi. The Grok rows in the price index are:

| Model | Credits per message | Plan required | Context window | Cost per standard answer |
| --- | --- | --- | --- | --- |
| Grok Build 0.1 | 3 | Powerhouse | 256K | $0.002 |
| Grok 4.3 | 4 | Powerhouse | 1M | $0.0025 |
| Grok 4.5 | 6 | Powerhouse | 500K | $0.005 |
| Grok 4.6 | 6 | Pro | 500K | $0.005 |

Two xAI identifiers on the 4 credit rung carry no row in the price index, so there is no context window or per answer price to publish for them. Neither is on the Pro list, so both require Powerhouse.

| Model identifier | Credits per message |
| --- | --- |
| `x-ai grok-4.20` | 4 |
| `x-ai grok-4.20-multi-agent` | 4 |

A standard answer in the Whizi cost index is 1,000 input tokens plus 500 output tokens, priced from OpenRouter list rates read on 20 August 2026. It exists to make providers comparable at one fixed shape of turn. It is not what you are billed: the bill is the subscription, and the meter inside it is credits.

Access resolves by identifier, with no separate provider step in the product: the gate checks the Starter set, then the agent set, then the Pro list, and returns Powerhouse for anything else. That fallback is why a Grok release added to the catalogue tomorrow is Powerhouse only until someone deliberately promotes it into the Pro list.

Selecting a model above your tier returns HTTP 403 with the code `tier_upgrade_required` and the message "Upgrade your plan to use this model."

#### What each plan buys in Grok messages

One credit is one message on the Whizi house model, and the credit cost of a model is a fixed integer per message. At 6 credits, the top Grok rows spend six times a house model message on every turn, whether the turn is a one line lookup or a full document. The rule behind that is in the [credits reference](https://whizi.io/docs/credits).

| Plan | Price | Credits per month | Messages on a 6 credit Grok model | Messages on a 3 credit Grok model |
| --- | --- | --- | --- | --- |
| Starter | $15.99/month, or $10.99/mo billed annually at $131.88 | 400 | No Grok model on this tier | No Grok model on this tier |
| Pro | $29.99/month, or $19.99/mo billed annually at $239.88 | 2,000 | 333 | Not reachable on Pro |
| Powerhouse | $49.99/month, or $34.99/mo billed annually at $419.88 | 8,000 | 1,333 | 2,666 |

Those are whole messages, and the remainder is not spendable: a turn that does not fit in the remaining balance is refused whole rather than part charged. What that looks like, and what to do about it, is in [out of credits](https://whizi.io/docs/out-of-credits).

Credits do not roll over. The balance sums only the usage rows carrying the current period key, and for a monthly plan that key is the UTC calendar month, so unspent credits from the last period are unreachable. Running out returns HTTP 429 with the code `message_limit_reached`.

Weekly billing is offered on mobile only. Credit metering is currently live on web, so the weekly credit figures in code are not what a weekly buyer is metered on today.

#### Why the Grok rungs sit where they do

Rungs come off a fixed ladder derived from live provider per token pricing, set out in full in the [credits reference](https://whizi.io/docs/credits). What that arithmetic returns for Grok is a 6: the Grok row on the Pro list sits at its cost true rung, so it is sold with no subsidy. At the anchor, one credit is worth $0.00156 of provider spend.

The 6 credit rung carries eight model identifiers today and two of them are Grok. Not every row on the ladder is priced at cost. The search grounded rows are priced above their token rate because a provider charges a flat fee per search that does not scale with tokens: one search is worth 3.2 credits on its own.

For outside scale: the median priced row in the index answers once for $0.00185. The dearest Grok rows sit at $0.005, a little under three times the median row and about one twentieth of the priciest row at $0.105.

A rung is not permanent. The ladder stopped at 20 until 19 August 2026, when a provider price sweep brought in models that cost more than that. Everything through 20 was left unchanged, so no price already on the ladder moved.

#### Auto does not route to Grok

Auto is a router row in the picker rather than a model. It reads the message and answers on whichever model suits it, and it chooses from six fixed rungs: quick runs on the house model at 1 credit, rewrite and attachment on a Gemini Flash model at 2, longform on a GPT model at 4, code on a Claude Sonnet model at 10, and reasoning on a GPT model at 20. None of the six is a Grok model, so selecting Auto never produces a Grok answer. To use Grok, pick it in the model picker.

Auto also carries no rung of its own. A turn costs whatever the model Auto picked costs, and the clients show a range for your tier instead of a single number: 1 to 1 credits on Starter, 1 to 10 on Pro, and 1 to 20 on Powerhouse. The range is published rather than a single figure because the endpoint that serves it is unauthenticated and does not know who is asking.

Auto also never routes above your plan. It walks back down the ladder until it finds a row your subscription can open, so a Starter account stays on the house model at 1 credit no matter what it asks.

Moving a conversation into or out of Grok is a per message decision, because the charge is per message on the model that answered. A conversation carries the model it was created with, which is why a model never leaves the catalogue once it has been offered.

#### What Whizi does not record about Grok

**There is no per model vision flag.** Whizi does not track which models can read an image, so there is no supported list to publish for Grok specifically. In the mobile app an attachment is forwarded to whichever model the turn is using; on the web the composer extracts document text in the browser before sending. The attachment rung in Auto routes to a Gemini Flash model on the reasoning that an image needs vision rather than a large context window.

**There is no per model web search flag.** Search is a request level toggle with three modes, off, probe and native, applied to the model you selected rather than stored as a property of that model. With search on, a request first carries a cheap probe tool of about 50 tokens, and only a turn where the model actually calls it is re-issued with real search.

**A context window is not a per turn budget.** A catalogue model with a context window of 93,000 tokens or more gets a flat budget of 40,000 input tokens and 20,000 output tokens per turn, and models with smaller windows get a proportional budget instead. Every Grok row above clears that threshold, so the spread from 256K to 1M makes no difference to a single turn: the widest window and the narrowest one are sent the same amount of text.

The full catalogue by plan is in the [model list](https://whizi.io/docs/models), and the complete credit scale is in the [credits reference](https://whizi.io/docs/credits).

Key points:

- The Grok rows in the price index with credit cost, plan gate and context window
- What a Pro allowance and a Powerhouse allowance buy in Grok messages
- Why a context window is not a per turn budget: a window of 93,000 tokens or more gets a flat 40,000 in and 20,000 out

Checklist:

- Starter includes no Grok model
- Pro reaches exactly one Grok model, at 6 credits per message
- Every other Grok identifier in the catalogue requires Powerhouse
- A free account can try any text model, Grok included, inside 7 lifetime messages
- The Grok rows documented here cost 3 to 6 credits per message
- Auto has six fixed rungs and none of them is a Grok model
- Credits do not roll over between periods

FAQ:

Q: Can I use Grok without an X subscription?
A: Yes. No xAI account and no provider key is involved at any point. Whizi issues no API keys and no personal access tokens, and the upstream provider credentials are server side secrets that never leave the worker, so the only account you need is a Whizi one.

Q: Which Whizi plan do I need for Grok?
A: Pro at $29.99/month, or $19.99/mo billed annually at $239.88, reaches the single Grok model on the Pro list. Every other Grok identifier resolves to Powerhouse at $49.99/month, or $34.99/mo billed annually at $419.88, including the frontier reasoning row and the build model. Starter at $15.99/month, or $10.99/mo billed annually at $131.88, includes no Grok model: it reaches exactly four picker entries, and none of them is from xAI.

Q: Why is one Grok model on Pro when another one costs the same?
A: Because the older of the two sits in the frontier reasoning tier and the newer one does not. Both rows answer for $0.005 and both sit on the 6 credit rung, but the frontier reasoning tier is one of the groups held back for Powerhouse deliberately, alongside the coding specialists, the 400B-class open weights, the routers and the Claude Opus class. Price and plan are decided separately, so what separates the two rows is the tier they were assigned, not what they cost to run.

Q: Can Grok read PDFs and images in Whizi?
A: Whizi does not record a per model vision or multimodal flag anywhere in the catalogue or the cost index, so there is no verified list of which Grok rows can read an image. What is documented is the accepted formats and how each one is handled, in the [supported file types](https://whizi.io/docs/supported-file-types) reference.

### Using Kimi in Whizi: which models, what they cost, which plan

URL: https://whizi.io/docs/using-kimi-in-whizi/
Updated: August 2026

Quick answer: Whizi carries Kimi from Moonshot inside the same subscription as the rest of the catalogue. Seven Kimi identifiers sit on the 10 credit rung, so a Kimi message costs 10 credits on every row. Two are reachable on Pro, the rest need Powerhouse, and Starter reaches no Kimi row.

The Kimi model identifiers Whizi prices at 10 credits, the plan that unlocks each one, and the context window and per-answer cost of the Kimi rows the Cost Index prices.

#### The short answer

Whizi carries Kimi from Moonshot inside the same subscription as the rest of the 280+ model catalogue. Seven Kimi identifiers sit on the 10 credit rung, so a Kimi message costs 10 credits on every row below. Two of them are reachable on Pro, the rest need Powerhouse, and Starter reaches no Kimi row.

| Model identifier | Credits per message | Plan required |
| --- | --- | --- |
| `moonshotai kimi-k3` | 10 | Pro |
| `moonshotai kimi-k2.5` | 10 | Pro |
| `moonshotai kimi-k2-thinking` | 10 | Powerhouse |
| `moonshotai kimi-k2.7-code` | 10 | Powerhouse |
| `moonshotai kimi-k2` | 10 | Powerhouse |
| `moonshotai kimi-k2-0905` | 10 | Powerhouse |
| `moonshotai kimi-k2.6` | 10 | Powerhouse |

No catalogue model reaches production without a rung: a build test fails if one does. And a free account is not model-gated, so it can open any of these rows inside its 7 message lifetime allowance.

#### Which plan opens which Kimi row

Pick the Kimi reasoning row on a Pro subscription and what comes back is HTTP 403, code `tier_upgrade_required`, message "Upgrade your plan to use this model." Five of the seven rows behave that way on Pro, for two different reasons.

Two are held back on purpose. The reasoning row sits in the frontier reasoning tier and the coding row sits with the coding specialists, the two categories reserved so the top tier has something to sell. Both categories and their membership are listed on [the models page](https://whizi.io/docs/models).

The other three are Powerhouse by omission rather than by a rule. Access is resolved by identifier: the gate checks the Starter set, then the agent set, then the Pro list, and anything it does not find returns "powerhouse". Nothing promotes a row automatically, so the next Kimi identifier the catalogue gains lands on Powerhouse until someone adds it to the Pro list by hand. Since all seven rows charge the same 10 credits, the gate and the price are unrelated questions on this family. Of the two rows Pro does reach, K3 is the Moonshot flagship, and it is named on the Pro paywall card beside Claude Sonnet.

Tiers stack by rank, so Powerhouse opens everything Pro opens. If a Kimi row shows as locked, [when a model is unavailable](https://whizi.io/docs/model-unavailable) covers what to do next.

| Plan | Price per month | Kimi rows reachable |
| --- | --- | --- |
| Free | No paid allowance | Any text model, inside a 7 message lifetime cap |
| Starter | $15.99 | None |
| Pro | $29.99 | `moonshotai kimi-k3`, `moonshotai kimi-k2.5` |
| Powerhouse | $49.99 | All seven above |

#### Context window and cost per answer

The Whizi Cost Index prices the following Kimi rows. A standard answer there is 1,000 input tokens plus 500 output tokens, priced from list rates read on 20 August 2026, so the figures are comparable across providers rather than being a Whizi charge.

| Cost Index row | Context window | Cost per standard answer | Per thousand answers | Credits |
| --- | --- | --- | --- | --- |
| Kimi K2 0711 | 131K | $0.00172 | $1.72 | 10 |
| Kimi K2 Thinking | 262K | $0.00185 | $1.85 | 10 |
| Kimi K2.7 Code | 262K | $0.00246 | $2.46 | 10 |
| Kimi K3 | 1M | $0.0105 | $10.50 | 10 |

Kimi K2 Thinking is the median priced row of the whole index, which prices 100 OpenRouter rows across 29 providers. The index is a price comparison table and its row count is not the size of the Whizi catalogue.

A window is not a per-turn budget. The flat budget is 40,000 input tokens and 20,000 output tokens, and it applies to any catalogue model whose window is at least 93,000 tokens. Every priced Kimi row clears that line, which makes this family a clean demonstration of how little the window column decides. The four windows run from 131K to 1M, and all four rows still send the same 40,000 in and 20,000 out and still cost the same 10 credits. The 1M window on the K3 row buys room for a conversation to keep growing, not a million token prompt. Long threads are covered in [when the context is too long](https://whizi.io/docs/context-too-long).

#### Why every Kimi row costs the same 10 credits

One credit is one message on the house model. A rung is derived from a reference turn of 3,000 input tokens and 800 output tokens divided by that same turn on the 1x model, rounded up when a model falls between two rungs, because the cheap direction is the one that costs money. The legal rungs are 1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 20, 25, 30, 40, 50, 60, 80 and 100, and a model may not be priced off the ladder.

Kimi does not follow that arithmetic. The family is pinned at 10 by owner policy, against cost-true rungs of 2, 3, 4 and 12 depending on the row. So on the cheaper Kimi rows you are charged above what the row costs to run, and on the dearest rows you are charged below it. The Claude Sonnet class is pinned the same way, at 10 against a cost-true 12, which is why a Kimi turn and a Claude Sonnet turn draw the same amount from the same allowance.

An identifier with no rung at all falls back to 100 credits, the top of the ladder, chosen so an unpriced model surfaces as a support ticket rather than quietly eating margin. Every Kimi identifier in the table above carries its own rung, so none of them can land there.

The full ladder and the models on every rung are in the [credits reference](https://whizi.io/docs/credits).

#### What an allowance buys in Kimi turns

Credits are one shared allowance per billing period, and every Kimi row spends 10 of them per turn, which reduces the arithmetic on this family to a division by ten. Weekly billing is offered on mobile only, and not on every tier, so the weekly column below is the allowance a weekly period carries where that product is sold, not a plan every tier can buy.

| Plan | Credits per monthly period | Credits per weekly period | Kimi messages if spent only on Kimi |
| --- | --- | --- | --- |
| Starter | 400 | 100 | Not applicable, no Kimi access |
| Pro | 2,000 | 500 | 200 monthly, 50 weekly |
| Powerhouse | 8,000 | 2,000 | 800 monthly, 200 weekly |

Credits do not roll over: the balance sums only the rows carrying the current period key, so an unspent allowance is unreachable once the period turns over.

Credit pricing is also gated per platform, and the allowlist value in production today is the web client. On a 1 credit row the two systems come to the same thing. A Kimi turn does not: off the allowlist it is charged one message against the message allowance rather than 10 credits against the credit allowance, so a Pro account is metered at 800 messages instead of 2,000 credits.

#### Switching into and out of Kimi mid-conversation

The charge follows the model that answered: `creditsForModel` returns a fixed integer per model identifier. Sending three messages to a 1 credit model and then one to Kimi in the same thread spends 3 plus 10, not 40. Switch back and the next turn is at the cheaper rate again.

What carries across a switch is covered in [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

#### What a Kimi row will not do

**Auto will not route to it.** Auto has six fixed targets and Moonshot supplies none of them. What it does have is a code target priced at 10 credits, the same rung every Kimi row sits on, so sending a turn through Auto is not a cheaper road to Kimi pricing than picking Kimi. Kimi stays a manual pick.

**It will not generate media.** Image, video and audio generation are separate picker sections with their own hand-curated rows.

**It gets no special search behaviour.** Web search is a per-request toggle, so a Kimi turn behaves like any other text model. See [web search](https://whizi.io/docs/web-search).

**It carries no published capability flags.** Whizi records no per-model vision, audio input or knowledge cutoff, so this page publishes none for Kimi.

Key points:

- Every Kimi identifier on the 10 credit rung, with its plan gate
- Why all of them cost the same 10 credits, whatever the row costs to run
- The context window and per-answer cost of the Kimi rows the Cost Index prices

Checklist:

- Seven Kimi identifiers sit on the 10 credit rung, and each costs 10 credits per message
- Starter reaches no Kimi row
- Pro reaches the K3 flagship and one K2 variant
- The reasoning and coding rows require Powerhouse
- A free account can open any text model inside a 7 message lifetime cap
- The four Kimi rows in the Cost Index all clear the 93,000 token threshold, so they get the flat 40,000 in and 20,000 out per-turn budget
- Auto never picks a Kimi row for you

FAQ:

Q: Which Kimi models are included in Whizi?
A: Seven Kimi identifiers carry a 10 credit price: the base K2 row, a dated K2 build, the reasoning variant, two further K2 variants, the coding variant, and K3.

Q: What happens to a Kimi message when I have fewer than 10 credits left?
A: It is refused whole rather than part charged, and the app returns HTTP 429 with the code `message_limit_reached` and the message "Your monthly message limit has been reached." Because every Kimi row is 10, the refusal threshold is one number for the whole family: an allowance either has a Kimi turn left in it or it does not, and there is no per-row arithmetic to do first. A cheaper row whose rung fits the remaining balance still goes through. [What to do when credits run out](https://whizi.io/docs/out-of-credits) covers the rest.

Q: Is Kimi cheaper than Claude in Whizi?
A: Not to you: both draw 10 credits a message, so a Kimi turn and a Claude Sonnet turn cost a subscriber exactly the same. They are not the same to run. In the Cost Index the Kimi reasoning row answers a standard question for $0.00185 against $0.007 for Claude Sonnet 5, and the top Kimi row at $0.0105 matches the older Sonnet rows to the tenth of a cent.

Q: Where is Kimi in the model picker?
A: Under the Moonshot vendor logo rather than the generic fallback icon. Moonshot is one of the families the picker gives its own logo to. It is not in the Recommended section: that section is Whizi AI, Auto, one GPT row, one Claude row and one Gemini row, and it is constrained to models reachable on Pro or below. The other five sections are Agents, Image Generation, Video Generation, Audio Generation and All models.

Q: Can I try Kimi without paying?
A: Yes, seven times. A free account is not model-gated and its credit allowance is 0, so a Kimi turn on the free tier spends a message off the lifetime cap and no credits at all. The same turn takes 10 credits off a Pro subscriber, so spending the whole free allowance on this family is 70 credits of Pro usage for nothing. The cap never resets, and once it is gone the account gets HTTP 402 with the message "Your free messages are used up. Start a subscription to keep chatting."

### Using Llama in Whizi: models, credit cost and plan

URL: https://whizi.io/docs/using-llama-in-whizi/
Updated: August 2026

Quick answer: Whizi carries the Meta Llama family in its catalogue, and both priced Llama rows cost 1 credit per message. Pro is the plan that unlocks the Llama set, and Powerhouse includes it. Starter includes no Llama model. Any Llama identifier the Pro set does not name resolves to Powerhouse.

The Llama models Whizi carries, what each one costs per message in credits, the plan that unlocks them, and the per turn token budget they actually run with.

#### The short answer

Yes, Whizi carries Meta's Llama family inside its 280+ model catalogue, and both priced Llama rows sit on the cheapest rung Whizi meters: 1 credit a message. Llama sits on the Pro plan, which is $29.99/month, or $19.99/mo billed annually at $239.88. Starter does not include any Llama model.

The Llama rows Whizi prices, with the credit cost charged per message:

| Model | Context window | Cost per standard answer | Credits per message | Minimum plan |
| --- | --- | --- | --- | --- |
| Llama 4 Maverick | 1M | $0.0006 | 1 | Pro |
| Llama 3.3 70B Instruct | 131K | $0.00026 | 1 | Pro |
| Llama 4 Scout | Not carried in the Cost Index | Not carried in the Cost Index | Not published | Pro |

A standard answer is 1,000 input tokens plus 500 output tokens, priced from rates read on 2026-08-20, so the figures compare like for like across providers.

Those three identifiers are the Llama rows named in the Pro model set. Any other Llama identifier the catalogue carries falls to the Powerhouse fallback described below, so treat the table as the Pro list rather than a complete catalogue enumeration.

#### What a Llama message costs in credits

Both priced Llama rows sit on the bottom rung of the credit ladder, so a Llama turn costs 1 credit whatever you send it. The third Llama identifier in the Pro set is not carried in the Cost Index, so this page publishes no credit figure for it.

At 1 credit a message, the credit allowance is the Llama message count: on the web, Pro buys 2,000 Llama turns a month and Powerhouse buys 8,000. Starter reaches no Llama row at all.

| Plan | Credits per month (web) | Messages per month off the credit allowlist |
| --- | --- | --- |
| Starter | 400 | 400 |
| Pro | 2,000 | 800 |
| Powerhouse | 8,000 | 5,000 |

Weekly billing is offered on mobile only, and not on every tier. Where it is sold, the weekly credit allowances are 100 on Starter and 500 on Pro, and the weekly message allowances are 100 and 400.

A turn is charged at the rung of the model that answered it, so a conversation that moves between families is billed per turn at each answering model rate. Credits do not roll over: the balance sums usage inside the current period key, so the arrival of a new period is itself the reset. There is no top up path.

Credit multipliers are charged only on platforms in the credit allowlist, which today is the web app. A platform outside the allowlist sees a flat 1x, is charged one per turn, and keeps the older message allowance instead. On Starter the two allowances are the same number, 400. On Pro and Powerhouse they are not, which is the second and third columns of the table above.

The full ladder, all fourteen occupied rungs, is in the [credits reference](https://whizi.io/docs/credits).

#### The plan gate, exactly

Access is resolved server side by identifier (see the [models reference](https://whizi.io/docs/models)).

**Starter** reaches exactly four picker entries: Auto, Whizi AI, one OpenAI model and one Google model. No Llama row is among them.

**Pro** adds the Llama set on top of Starter. Read as the curation rule applied to one family, that set is the current Llama 4 rows plus the older Llama 3 workhorse, three identifiers in total.

**Powerhouse** includes everything Pro reaches, because the tier ranks compare with a greater than or equal test, so a higher tier always satisfies a lower requirement. Powerhouse is also where any catalogue identifier lands that is not named in the Starter or Pro sets, which is why an unlisted Llama identifier is Powerhouse until it is deliberately promoted.

Selecting a model above your tier returns HTTP 403 with the code `tier_upgrade_required` and the message "Upgrade your plan to use this model." That symptom is covered in [when a model is unavailable](https://whizi.io/docs/model-unavailable).

Free accounts are not model gated. The chat path returns for the free tier before the model gate is reached, so a free account can try any text model, limited by a lifetime allowance of 7 messages that never resets. The free credit allowance itself is 0, so those 7 messages are the whole of it.

#### Context window versus what a turn actually gets

The context window in the table above is the model published window. It is not the budget a single Whizi turn runs with.

Every catalogue model gets a flat budget of 40,000 input tokens and 20,000 output tokens per turn. A model whose window sits below 93,000 tokens gets a smaller proportional budget instead, with output capped at 40 percent of the window and a 1,000 token safety margin held back.

Because both priced Llama rows sit above the 93,000 token threshold, both run on the same flat 40,000 input and 20,000 output budget. The larger published window does not raise what a single Whizi turn sends, and the provider counts requested maximum output against the window as well as input, so the 1M row and the 131K row are sized identically here. If a thread outgrows the budget, see [when a conversation is too long](https://whizi.io/docs/context-too-long).

#### Where these rows sit in the Cost Index

The Cost Index prices 100 rows across 29 providers at one standard answer each. That number is the size of the price index, not the size of the Whizi catalogue.

| Reference point | Cost per standard answer |
| --- | --- |
| Cheapest priced row in the index | $0.000053 |
| Median priced row | $0.00185 |
| Llama 3.3 70B Instruct | $0.00026 |
| Llama 4 Maverick | $0.0006 |
| Priciest priced row in the index | $0.105 |

Both priced Llama rows land below the index median, and the spread between the cheapest and priciest priced row is roughly 2000x. 89 of the 100 priced rows carry a Whizi credit charge; the rest are priced for comparison but are not offered as their own metered row.

#### What Llama is not used for in Whizi

**No Llama row sits on the Auto ladder.** Auto has six rungs, and the models on them are the house model for short questions, a Gemini Flash row for rewrites and again for attachments, a GPT row for long form, a Claude Sonnet row for code and a second GPT row for multi step reasoning. No Llama identifier appears on any of the six, so Auto cannot reach Llama at any tier and a Llama answer is always a manual pick in the model picker.

**Attachments on Auto go elsewhere.** Auto routes any turn carrying an attachment to Gemini Flash, on the reasoning that an image needs vision rather than a large window. Select a Llama row yourself and the attachment is handled per request rather than per model. On the app, it is forwarded to whichever model the turn is on, and only the most recent user message carrying attachments is forwarded. On the web, a PDF, Word or spreadsheet file has its text extracted in the browser and the text is what reaches the model.

**Voice and media generation are not per model choices.** Media generators skip the per model gate entirely and are tiered by picker section instead, and [voice mode](https://whizi.io/docs/voice-mode) runs on its own realtime route, so picking a Llama row changes neither.

**Whizi does not publish per model vision support.** There is no vision or multimodal capability flag anywhere in the catalogue or the cost index, so Whizi cannot tell you which Llama row reads an image. Image parts are forwarded to whichever model the turn is on and the result is between you and the model.

Key points:

- The Llama rows Whizi prices, with credit cost and context window
- The plan that unlocks Llama, and what Starter does not include
- The flat 40,000 input and 20,000 output token budget a Llama turn runs with

Checklist:

- Llama is in the catalogue and both priced rows cost 1 credit per message
- Pro is the plan that unlocks the Llama set, and Powerhouse includes it
- Starter includes no Llama model
- A free account can try any text model inside a 7 message lifetime allowance
- A Llama turn runs on the flat 40,000 input and 20,000 output token budget, because both priced rows sit above the 93,000 token threshold
- No Llama row sits on any of the six Auto rungs, so Llama is always a manual pick
- Whizi publishes no per model vision flag, so which Llama row reads an image is not something it can tell you

FAQ:

Q: Which Llama models does Whizi include?
A: The Pro model set names three Llama identifiers: the two Llama 4 variants plus the Llama 3 workhorse. Two of them are priced in the Whizi Cost Index, both at 1 credit per message, with published context windows of 1M and 131K. The catalogue may carry further Llama identifiers, and anything not named in the Starter or Pro sets requires Powerhouse.

Q: How many credits does a Llama message cost?
A: One credit, which is the bottom rung of the ladder and the same as one message on the Whizi house model. Starter cannot open a Llama row at all, and a free account can, but only inside its 7 message lifetime allowance, so Pro is the first plan where that 1 credit buys a usable number of Llama turns.

Q: Do I need the top plan to use Llama?
A: No. Pro is the cheapest plan that opens a Llama row, at $29.99/month. Starter does not reach one, and Powerhouse is needed only for a Llama identifier the Pro set does not name. A free account is not model gated at all, so it can open a Llama row inside its 7 message lifetime allowance.

Q: Can I switch a conversation from Llama to another model?
A: Yes. Each turn is charged at the rung of the model that answered it, so a thread that moves between families is billed per turn at each rate rather than at the thread level. A model never leaves the catalogue once it has been offered, because a conversation carries the model it was created with, so an old Llama thread still opens later. The mechanics are in [switching models mid conversation](https://whizi.io/docs/switch-models-mid-conversation).

Q: What can I attach to a Llama turn?
A: Attachments are handled per request rather than per model, and they cost no credits of their own. The caps are 10 MB per file, 4 attachments per prompt and 20 MB in total per prompt. The accepted types are JPEG, PNG, WebP, GIF, PDF, plain text, Markdown, CSV and JSON.

Q: Does web search work with Llama?
A: Whizi does not publish a per model web search list. Search is a per request toggle with three modes, off, probe and native, rather than a per model capability flag, and the per model matrix was never enumerated, so there is no Llama specific answer to give. The toggle itself is covered in [web search](https://whizi.io/docs/web-search).

### Using Mistral in Whizi: models, credit costs and plan gates

URL: https://whizi.io/docs/using-mistral-in-whizi/
Updated: August 2026

Quick answer: Whizi carries Mistral models under a Whizi subscription rather than a Mistral one. Four Mistral entries sit on Pro, every other Mistral entry resolves to Powerhouse, and Starter reaches none. Of the rows priced individually, a message costs 1 credit on Codestral, 2 on the current Mistral Large and Medium rows, and 6 on Mixtral.

The Mistral models Whizi publishes a credit cost for, the plan that unlocks each one, and the context window and per-answer cost behind the number.

#### The short answer

Of the Mistral rows the published reference prices individually, a message costs 1 credit on Codestral, 2 on the current Mistral Large and Mistral Medium rows, and 6 on the older Mistral Large, the newer Mistral Medium and Mixtral. Codestral and the Devstral coding family need Powerhouse: Devstral is named among the coding specialists held back for the top tier, and Codestral is not written into the Pro model set, so it falls to Powerhouse under the fallback rule. Four Mistral entries sit on Pro and every other Mistral entry resolves to Powerhouse. Starter reaches no Mistral model at all, because Starter is limited to four picker entries and none of them is a Mistral one. All of it opens with a Whizi subscription rather than a Mistral one, inside a 280+ model catalogue. How that compares with subscribing to Mistral's own assistant is set out in [the Le Chat alternative guide](https://whizi.io/alternatives/le-chat).

| Model | Lowest plan that opens it |
| --- | --- |
| Mistral Small 3.2 24B Instruct | Pro |
| Mistral Large 3 2512 | Pro |
| Mistral Medium 3.1 | Pro |
| Mistral Nemo | Pro |
| Mistral Large 2407 | Powerhouse |
| Mistral Medium 3.5 | Powerhouse |
| Mixtral 8x22B Instruct | Powerhouse |
| Codestral 2508 | Powerhouse |
| The Devstral coding family | Powerhouse |

The gate is a rule, not a hand-kept list. Access is resolved by model identifier: the Starter set is checked, then the agent set, then the Pro set, and anything else returns Powerhouse. The four identifiers above are the Mistral entries written into the Pro set, so a Mistral identifier that is not one of those four is Powerhouse by that fallback. The same fallback is why a newly added model lands on the top tier until it is deliberately promoted. Auto does not reach Mistral either: its six rungs are pinned to the house model, two Gemini Flash rows, two GPT rows and a Claude Sonnet row, so a Mistral answer is always a manual pick in the picker.

#### What a Mistral message costs in credits

The priced Mistral rows sit on three separate rungs of the credit ladder: 1, 2 and 6. The dearest of them therefore costs six times the cheapest on every turn. The charge is a fixed integer per model, so the same row costs the same amount whether the message is one line or two thousand words, because nothing about message length, conversation length or answer length enters the calculation. The definition of a credit and the full ladder are in the [credits reference](https://whizi.io/docs/credits).

| Model | Credits per message |
| --- | --- |
| Codestral 2508 | 1 |
| Mistral Large 3 2512 | 2 |
| Mistral Medium 3.1 | 2 |
| Mistral Large 2407 | 6 |
| Mistral Medium 3.5 | 6 |
| Mixtral 8x22B Instruct | 6 |

Mistral Small and Mistral Nemo sit inside credit groups that the published reference does not enumerate row by row, so this page does not state a number for them. On the web app, the model picker shows the live credit cost on every row, and that is the figure the charge is taken from.

What the allowances are: Pro carries 2,000 credits per monthly billing period and Powerhouse carries 8,000. Starter carries 400 credits but reaches no Mistral model.

| Plan | Price | Credits per month | Mistral access |
| --- | --- | --- | --- |
| Starter | $15.99/month, or $10.99/mo billed annually at $131.88 | 400 | None |
| Pro | $29.99/month, or $19.99/mo billed annually at $239.88 | 2,000 | The four entries above |
| Powerhouse | $49.99/month, or $34.99/mo billed annually at $419.88 | 8,000 | The rest of the family |

A turn is floored at one credit and rounded down. A turn that does not fit in the remaining balance is refused whole rather than part charged, so a 6 credit message on a balance of 3 is rejected rather than discounted. A generation that fails or is aborted is refunded by removing the usage row.

#### Context windows and the cost of one answer

The Whizi cost index prices a standard answer, defined as 1,000 input tokens plus 500 output tokens, so list rates from different providers become comparable. The source prices were read from OpenRouter on 2026-08-20. These are the Mistral rows the index carries; the index prices 100 rows for comparison, so it is a sample of the catalogue rather than the whole of it.

| Model | Context window | One standard answer | One thousand answers | Credits |
| --- | --- | --- | --- | --- |
| Codestral 2508 | 256K | $0.00075 | $0.75 | 1 |
| Mistral Large 3 2512 | 262K | $0.00125 | $1.25 | 2 |
| Mistral Medium 3.1 | 131K | $0.0014 | $1.40 | 2 |
| Mistral Medium 3.5 | 262K | $0.00525 | $5.25 | 6 |

For scale against the rest of the index: the cheapest priced row in the whole index costs $0.000053 per standard answer and the priciest costs $0.105, a spread of roughly 2000x. The median priced row is $0.00185. Every Mistral row above sits below that median except the priciest Mistral Medium row.

A rung is not read off that table directly. It comes from a reference turn of 3,000 input and 800 output tokens divided by the same turn on the 1x anchor model, which is $0.00156 of provider spend. A model that falls between two rungs is rounded up, because the cheap direction is the one that costs money. The legal rungs are 1, 2, 3, 4, 5, 6, 8, 10, 12, 15, 20, 25, 30, 40, 50, 60, 80 and 100, and nothing may be priced off that ladder.

#### What the context window does not mean

Three of the four priced Mistral rows above publish a window of 256K or larger and the fourth publishes 131K, yet all four are sized identically inside Whizi at 40,000 input tokens and 20,000 output tokens per turn. That flat budget is what every catalogue model gets. Only a model whose window sits below 93,000 tokens is treated differently, with a smaller proportional budget, output capped at 40 percent of the window and a 1,000 token safety margin held back, and no Mistral row in the table is anywhere near that threshold.

The 40,000 is not counted at face value on any model. The token estimator can undercount a real tokenizer by up to 1.66x on content such as JSON, so a 1.8x allowance is applied to the input side across the platform before the budget is checked. The provider also counts requested maximum output against the window rather than input alone.

#### Where a Mistral row cannot be used

Mistral is one of the families carrying its own vendor logo in the picker, but its rows only ever appear in the text part of it. Image generation is Whizi Image, Nano Banana, Flux and Stable Diffusion, video generation is a single row and audio generation is a single row, and no Mistral identifier is in any of the three. Those generators skip the per model gate entirely and are tiered by picker section instead, so the Mistral plan gate never enters into them.

Whizi publishes no per model capability flag, so nothing marks a Mistral row as reading an image or reaching the web. Web search is a per request toggle with three modes, off, probe and native, and it works on a Mistral turn exactly as on any other text model. Native provider search injects a fixed preamble of roughly 4,400 input tokens, measured as the same two message prompt going from 36 tokens to 4,472 with the tool attached, which is why a cheap probe tool of about 50 tokens is tried first.

File input reaches a Mistral row differently on each platform. On the web app, PDF, Word and spreadsheet text is extracted in your browser before the turn is sent, so the row receives extracted text rather than the original file. On the app, the attachment itself is forwarded to whichever model the turn is on, and only the most recent user message carrying attachments is forwarded.

Voice is the one mode where no model choice exists for anybody. It runs on the Gemini Live API with a single preset voice and a 15 minute session cap, and the browser opens a socket straight to Google, so no audio passes through the Whizi backend. There is no route that answers a voice turn on a Mistral row.

#### If a Mistral row is locked, missing, or your balance runs out

Selecting a model above your tier returns HTTP 403 with the code `tier_upgrade_required` and the message "Upgrade your plan to use this model." Access is resolved by model identifier through the shared tier lookup.

A Mistral row the catalogue has stopped promoting is a different case from a locked one. A conversation carries the model it was created with, so an older Mistral thread still reopens on that row. What changed is intake rather than retention: nothing older than a 2025 release is added to the catalogue any more.

A free account is not model gated, so it can open any text model in the catalogue, Mistral included, inside a lifetime allowance of 7 messages. It carries a credit limit of 0 and that allowance never resets, so the seven messages are the whole of it. When the cap is spent the response is HTTP 402 with the code `free_limit_reached` and the message "Your free messages are used up. Start a subscription to keep chatting."

On a paid plan, running out of allowance returns HTTP 429 with the code `message_limit_reached` and the message "Your monthly message limit has been reached." Weekly plans say weekly instead. Credits do not roll over: usage is summed against the current period key, so the arrival of a new period is itself the reset, and last period unspent credits are unreachable.

Credit pricing is applied per platform, which is why the numbers can look different on another device. A client on the credit allowlist sees the real multipliers in the picker and is charged them against the credit allowance. A client that is not on it is charged one per turn and keeps the older message allowance instead, which is 400 messages monthly on Starter, 800 on Pro and 5,000 on Powerhouse.

Key points:

- The four Mistral entries Pro opens, and why every other one needs Powerhouse
- Credits per message for the Mistral rows the published reference prices individually
- Context window and cost of one standard answer, from the Whizi cost index

Checklist:

- Pro opens four Mistral entries, Starter opens none
- Every other Mistral entry resolves to Powerhouse through the fallback rule
- The current Mistral Large and Mistral Medium rows sit at 2 credits per message, Codestral at 1
- Mixtral, the older Mistral Large and the newer Mistral Medium sit at 6 credits
- Credit cost is fixed per model and does not vary with message length
- Every Mistral turn is capped at 40,000 input and 20,000 output tokens whatever the window says
- Auto never routes to Mistral, and voice and media generation never run on it

FAQ:

Q: Which Whizi plan do I need for Mistral?
A: Pro at $29.99/month is the entry point: it carries the current Mistral Large and Mistral Medium rows, plus Mistral Small and Mistral Nemo. Starter at $15.99/month reaches four picker entries in total and none of them is Mistral. Powerhouse covers every other Mistral identifier, including Mixtral, Codestral, the older Mistral Large, the newer Mistral Medium and the Devstral coding models, which are held back with the other coding specialists.

Q: How many credits does a Mistral message cost?
A: It depends on the row, and what the rung really costs is easier to read against an allowance. Every Mistral row priced at 6 credits sits behind Powerhouse, so that rung is only ever spent out of the 8,000 credit monthly allowance, which is 1,333 such turns. The Mistral rows Pro opens with a published rung are 2 credit rows, and Pro carries 2,000 credits a month, so that is 1,000 turns. Starter carries 400 credits and reaches no Mistral row at all.

Q: Is Codestral included in Whizi?
A: Yes, on the bottom rung: 1 credit per message, against a published context window of 256K and a cost of $0.00075 per standard answer in the Whizi cost index. It needs Powerhouse.

### Using Qwen in Whizi: which models, what they cost, which plan

URL: https://whizi.io/docs/using-qwen-in-whizi/
Updated: August 2026

Quick answer: A Qwen message in Whizi costs 1 to 8 credits, and which rows you can send to is decided by your plan. Starter includes no Qwen model at all, Pro reaches two Qwen identifiers, and every other Qwen row in the catalogue requires Powerhouse. Auto never routes to Qwen, so Qwen is a hand pick.

The Qwen models Whizi carries, the credit cost of each one, the plan that unlocks it, and the context window and per-answer price behind the number.

#### The short answer

A Qwen message costs 1 to 8 credits, and which rows you can send to is decided by your plan: Starter includes no Qwen model at all, Pro reaches two Qwen identifiers, and every other Qwen row in the catalogue requires Powerhouse. All of them sit inside the same subscription as the rest of the 280+ model catalogue.

The Qwen rows Whizi prices in its Cost Index, with what each costs to answer once:

| Model | Context window | Credits per message | Cost per standard answer | Per 1,000 answers |
| --- | --- | --- | --- | --- |
| Qwen3.7 Flash | 1M | 1 | $0.000095 | $0.095 |
| Qwen3 Coder Next | 262K | 1 | $0.00052 | $0.52 |
| Qwen3.6 Flash | 1M | 1 | $0.00075 | $0.75 |
| Qwen3.7 Plus | 1M | 1 | $0.00096 | $0.96 |
| Qwen3.8 27B | 1M | 3 | $0.00205 | $2.05 |
| Qwen3.7 Max | 1M | 5 | $0.003688 | $3.6875 |
| Qwen3.8 Max | 1M | 6 | $0.005 | $5.00 |

The dollar columns are list rates read from OpenRouter on 20 August 2026, for a standard answer of 1,000 input plus 500 output tokens. Across the Qwen rows they run from $0.095 to $5.00 per thousand answers, while the credit column moves only from 1 to 6. Credits are what Whizi charges you, so the two scales do not track each other.

#### Qwen identifiers with a credit rung

The Cost Index prices 100 rows across 29 providers for comparison, so it is a price index rather than the catalogue. The credit table is the wider list, and it carries further Qwen identifiers, including these:

| Model identifier | Credits per message |
| --- | --- |
| `qwen qwen3.8-2.4t-a95b` | 8 |
| `qwen qwen3.6-max-preview` | 6 |
| `qwen qwen3.8-max` | 6 |
| `qwen qwen3.7-max` | 5 |
| `qwen qwen3-max` | 3 |
| `qwen qwen3-max-thinking` | 3 |
| `qwen qwen3-coder` | 3 |
| `qwen qwen3.7-plus` | 1 |

Qwen is the largest family by prefix count in the catalogue. The live picker, served by `GET /api/models`, is the authoritative list of what is on offer today.

One rule governs any Qwen row not named here: a credit cost is a fixed integer per message, so nothing about message length, conversation length or answer length changes it. No Qwen row can reach you without a price either, because a build test fails if any catalogue model or agent id reaches production without a rung.

#### Which plan reaches Qwen

Model access is decided on the server by identifier, so a Qwen row your plan does not include cannot be reached by editing a request. What a locked row looks like when you send to it anyway is in [why a model shows as unavailable](https://whizi.io/docs/model-unavailable).

| Plan | Price | Qwen access | Credits per month |
| --- | --- | --- | --- |
| Free | No charge | Any text model, inside a lifetime allowance of 7 messages | 0 |
| Starter | $15.99/month, or $10.99/mo billed annually at $131.88 | None | 400 |
| Pro | $29.99/month, or $19.99/mo billed annually at $239.88 | `qwen qwen3.7-plus` and `qwen qwen3-235b-a22b-2507` | 2,000 |
| Powerhouse | $49.99/month, or $34.99/mo billed annually at $419.88 | Every other Qwen identifier in the catalogue | 8,000 |

Starter reaches exactly four picker entries: Auto, the house assistant Whizi AI, one OpenAI row and one Google row. That is the whole tier, so no Qwen model is available on it.

Pro is a curated list rather than everything cheap: one or two current flagships per family, each family's fast tier, and the high volume workhorses. Two Qwen identifiers made that list. Everything in the catalogue that is not in the Starter set and not in the Pro set resolves to Powerhouse, and that fallback is what puts the rest of Qwen on the top tier. The frontier reasoning rows and the coder family are held back for Powerhouse deliberately.

#### What a Qwen message spends

Credits are one allowance, spent per message at the rate of the model that answered. A turn is charged whole: if a 6 credit turn does not fit in the remaining balance it is refused rather than part charged, and the send comes back as HTTP 429 with the code `message_limit_reached`.

| Allowance | At 1 credit | At 3 credits | At 5 credits | At 6 credits | At 8 credits |
| --- | --- | --- | --- | --- | --- |
| Pro, 2,000 credits | 2,000 messages | 666 messages | 400 messages | 333 messages | 250 messages |
| Powerhouse, 8,000 credits | 8,000 messages | 2,666 messages | 1,600 messages | 1,333 messages | 1,000 messages |

What that means for the two Qwen identifiers Pro actually opens: `qwen qwen3.7-plus` is a 1 credit row, so a 2,000 credit month spent only on it is 2,000 Qwen messages.

Weekly billing, which is offered on mobile only, carries its own numbers: 500 credits per week on Pro and 2,000 on Powerhouse. Credits do not roll over on either cadence, because the balance sums only the rows carrying the current period key.

One thing does not follow you across a mid-thread switch onto a Qwen row: attachments. Only the most recent user message carrying attachments has them forwarded to the provider, so re-asking a Qwen row about a file further up a long thread may need the file attached again. The step by step version is in [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Credit multipliers are shown and charged per platform. The production allowlist today is the web client, which sends its platform on every call, so the badge in the picker and what the turn costs are decided by the same value. Off the allowlist a client gets no multiplier field at all, draws no credit badges, and is charged one message per turn against the older message allowance. On the 1 credit Qwen rows that is the same charge under either scheme. On the 8 credit identifier it is the difference between one message and eight credits. The full mechanics are in the [credits reference](https://whizi.io/docs/credits).

#### What the 1M window in the table does not buy you

The window column is the model's own window, not what Whizi sends in a turn. Whizi sends a flat 40,000 input tokens and 20,000 output tokens on every catalogue model, and only a model whose window sits under 93,000 tokens gets a smaller proportional budget instead. Every Qwen row in the table clears 93,000, so the 1M rows and the 262K row alike get 40,000 in and 20,000 out. Choosing between Qwen rows on window size alone changes nothing inside Whizi. What happens when a thread outgrows that budget is in [when the context is too long](https://whizi.io/docs/context-too-long).

#### Where Qwen does not appear

**Auto never routes to Qwen.** Auto is a router with no fixed credit cost of its own: a turn costs whatever it picked, and no Qwen row is among the things it can pick. Both rules that decide what it picks instead push away from the cheap end. When your plan cannot open a target, Auto does not fall to the cheapest row it can reach: a reasoning question drops to the code model, which reasons well. And difficulty is sticky inside a thread, so the previous turn's model acts as a floor and a follow-up cannot drop back to the cheapest model either. A 1 credit Qwen row is therefore never somewhere Auto arrives on its own. Qwen is a hand pick, every turn.

**No Qwen row is in the Recommended section.** The picker opens on Whizi AI, Auto and three other rows, and that section is constrained to models reachable at Pro or below so the top of the picker is never locked. Qwen rows live in the All models section.

**Image, video and music generation do not use Qwen.** Those three picker sections are hand curated and contain no Qwen row. They are also metered differently from each other: a standalone image generation spends only the image counter, which is 100 per month on Pro and 500 on Powerhouse, while standalone video and music jobs spend one credit each in addition to their own Powerhouse gate.

**The agent personas do not run on Qwen.** Every persona runs on the same house base model, and what a Pro subscription buys there is the prompt, the token budget and the reasoning depth, not a different provider.

Key points:

- The Qwen rows in the Whizi Cost Index, with credits per message, context window and cost per standard answer
- Which plan unlocks which Qwen model, and why Starter reaches none of them
- What a Qwen turn actually spends against a 2,000 or 8,000 credit allowance

Checklist:

- Qwen rows are priced from 1 to 8 credits per message
- Starter includes no Qwen model at all
- Pro reaches two Qwen identifiers, Powerhouse reaches the rest
- A free account can try any text model inside a lifetime allowance of 7 messages
- A credit cost is fixed per model, and message length never changes it
- Every Qwen row in the table clears the 93,000 token threshold, so all of them get 40,000 input and 20,000 output tokens per turn
- Auto cannot pick a Qwen row, so Qwen is a hand pick on every turn

FAQ:

Q: Does turning on web search change what a Qwen message costs?
A: Not in credits. The charge is a fixed integer per model identifier, so a Qwen turn with search on draws the same rung as one with search off. What it changes is the token side. Web search is a per-request toggle rather than a per-model capability, and a request with it on first carries a cheap probe tool of about 50 tokens, with only a turn where the model actually calls the tool re-issued with real search. Native provider search injects a fixed preamble of roughly 4,400 input tokens. An ordinary search request is capped at 3 results per search and 3 results in total.

Q: Why is the Qwen model I want greyed out in my picker?
A: Because the gate is resolved on the server by model identifier, and your plan does not include that one. Starter reaches no Qwen row at all, Pro reaches two Qwen identifiers, and every other Qwen row resolves to Powerhouse. Sending to a locked row anyway returns HTTP 403 with the code `tier_upgrade_required`. A newly added Qwen identifier is Powerhouse-only until it is explicitly promoted into the Pro list, so a row can be locked even though a similar one next to it is not.

Q: My balance is down to a couple of credits. Can I still send to Qwen?
A: Yes, if you pick a row whose rung fits what is left. Four of the seven Qwen rows the Cost Index prices sit at 1 credit, so a single remaining credit still buys a Qwen message on the same day a 6 credit row is being turned away. The refusal carries the message "Your monthly message limit has been reached.", and weekly plans say "weekly" instead. Nothing carries forward: credits do not roll over, because the balance sums only the rows in the current period, and the arrival of a new period is itself the reset.

Q: Can I try Qwen on a free account?
A: Yes, within limits. A free account is deliberately not model gated, so it can try any text model in the catalogue, but it carries a lifetime allowance of 7 messages that never resets and 0 credits. Once it is spent, sending returns HTTP 402 with the message "Your free messages are used up. Start a subscription to keep chatting." New accounts can also start a 7 day trial for $0.99.

### AI voice conversations in Whizi: how voice mode works

URL: https://whizi.io/docs/voice-mode/
Updated: August 2026

Quick answer: Voice mode in Whizi is a real-time spoken conversation: you talk, the model answers out loud, and you can interrupt it. Every paid plan includes voice minutes, 10 per month on Starter, 80 on Pro and 500 on Powerhouse. Minutes run on their own meter and spend no credits, and every session produces a transcript in your chat history.

Real-time spoken conversation with an AI model, how the monthly voice minutes on each plan are counted, and what voice is genuinely better at than typing.

#### The short answer

Voice mode is a real-time spoken conversation: you talk, the model answers out loud, and you can interrupt it. Every paid plan includes voice minutes: 10 per month on Starter, 80 on Pro, and 500 on Powerhouse.

The session produces a transcript that lands in your chat history, so a spoken conversation is searchable and re-readable afterwards exactly like a typed one. Voice minutes are counted against their own monthly meter and do not spend credits.

#### Starting and controlling a session

Open voice mode from the chat, and the session connects and begins listening. There is no push-to-talk step: you just start talking.

The controls during a session are deliberately minimal. You can mute the microphone, which stops it hearing you without ending the session, and you can end the session entirely. The state indicator shows whether Whizi is currently listening or speaking, which matters more than it sounds, because knowing whether you have been heard is most of what makes a voice interface tolerable.

Because the conversation is real-time rather than turn-based, you can cut in while the model is still speaking, the way you would with a person. That matters more than it sounds. Without it a long wrong answer has to be waited out before you can redirect it, which is the single most frustrating thing about voice interfaces that lack it.

The transcript can be opened in the chat at any point, including mid-session. That is the escape hatch when something needs to be precise: read what was actually captured rather than assuming. Speech recognition on names, numbers and technical terms is the weakest part of any voice interface, so a mid-session look at the transcript is the fastest way to catch a mishearing before it shapes the next answer.

#### How the minutes are counted

Minutes count session time, and the allowance resets monthly along with everything else on the account. It is shared across web and mobile, so time spent talking on the phone comes out of the same allowance as time spent on the desktop.

The allowance follows the plan: 10 minutes a month on Starter, 80 on Pro, and 500 on Powerhouse. Powerhouse works out to a little over eight hours per month, or roughly twenty minutes on every working day. Voice conversations are naturally short, the format suits five minute exchanges rather than hour-long sessions, so the higher tiers are rarely a binding constraint.

Muting the microphone does not stop the clock. Ending the session does. If you are stepping away, end it rather than muting.

Voice minutes are their own allowance and draw on nothing else. Credits, image generations, voice minutes and video generations are four separate meters that do not share a pool, so a long spoken conversation does not reduce the number of Claude or GPT messages you can send that month. It also means an account that has run its credits down still has its voice minutes, and an account that has talked its voice allowance flat can still type.

The transcript a session leaves behind is an ordinary chat, so it costs nothing further to keep and stays searchable alongside everything you typed.

#### What voice is actually better at

Voice is not a faster version of typing, and using it that way is disappointing. Three cases where it genuinely wins:

**Thinking out loud.** Talking through a problem you have not structured yet is much easier spoken than written, because you do not have to commit to a sentence before you know where it ends. The transcript afterwards is usually a better description of the problem than anything you would have typed.

**Hands busy or eyes elsewhere.** Cooking, driving, walking, or working through something physical. This is the obvious one and it is real.

**Practising a conversation.** Interview answers, a difficult message to a client, a pitch. Saying it out loud and hearing a response surfaces the awkward phrasing that reads fine on a page.

Where it loses: anything involving code, exact numbers, names that are hard to spell, or long structured output you will want to copy. Type those.

| What you are doing | Spoken or typed | Why |
| --- | --- | --- |
| Thinking through a problem you have not structured yet | Spoken | You do not have to commit to a sentence before you know where it ends |
| Cooking, driving, walking, or working through something physical | Spoken | Hands are busy and eyes are elsewhere, and the transcript is in your history afterwards |
| Practising interview answers, a difficult message, or a pitch | Spoken | Saying it out loud surfaces the awkward phrasing that reads fine on a page |
| Code, exact numbers, hard-to-spell names, long output you will copy | Typed | Speech recognition is weakest on names, numbers and technical terms |

Key points:

- What voice mode includes, and which plan it requires
- How the minute allowance is counted and shared across devices
- The three things voice is genuinely better at than typing

Checklist:

- Every paid plan includes voice: 10 minutes on Starter, 80 on Pro, 500 on Powerhouse
- Minutes are shared across web and mobile
- Voice does not spend credits, it has its own meter
- Muting does not stop the clock, ending the session does
- The transcript lands in chat history and stays searchable
- Use typing for code, exact numbers and anything you will copy

FAQ:

Q: Which plan includes voice mode?
A: Every paid plan includes voice mode. The allowance is what changes: Starter includes 10 voice minutes per month, Pro includes 80, and Powerhouse includes 500. The minutes reset monthly and are shared across web and mobile.

Q: Do voice minutes come out of my credits?
A: No. Voice has its own monthly meter of minutes and spends no credits, so a long spoken conversation does not reduce the number of Claude or GPT messages you can send. Credits, image generations, voice minutes and video generations are four separate allowances that do not share a pool.

Q: Can I see what I said in a voice session?
A: Yes. Every session produces a transcript that goes into your chat history, and you can open it during the session as well as after it. This is worth doing whenever precision matters, because speech recognition on names, numbers and technical terms is the weakest part of any voice interface.

Q: Can I interrupt the model while it is speaking?
A: Yes, the conversation is real-time rather than turn-based, so you can cut in the way you would with a person. This matters more than it sounds: without it, a long wrong answer has to be waited out, which is the single most frustrating thing about voice interfaces that lack it.

### Whizi voice mode not working: every error and its fix

URL: https://whizi.io/docs/voice-mode-not-working/
Updated: August 2026

Quick answer: Voice mode fails for one of four reasons: the account is not signed in or has no voice minutes left, the browser blocked the microphone, the realtime connection to the voice service failed, or the session hit one of its own limits. The message Whizi shows maps to exactly one condition.

Whizi mints a session token and your browser talks straight to the voice provider, so no audio passes through Whizi servers. That split decides which failures you can fix and which you cannot.

#### The short answer

Voice mode fails for one of four reasons: the account is not signed in or has no voice minutes left, the browser blocked the microphone, the realtime connection to the voice service failed, or the session hit one of its own limits. The message Whizi shows tells you which, and each message maps to exactly one condition in the code.

Find the text you are looking at in this table, then read the matching section below for the detail.

| What you see | Where it comes from | What to do |
| --- | --- | --- |
| `Sign in to Whizi to use voice mode.` | No session token when the voice overlay opens | Sign in, then reopen voice mode |
| `Upgrade your subscription to talk with Whizi in voice mode.` | A free account starting a voice session | Subscribe: free accounts get 0 voice minutes |
| `Your monthly voice minutes have been used up.` | The period voice reservation was refused | Wait for the period to reset, or move up a tier |
| `Microphone access is blocked. Allow it in your browser to use voice mode.` | The microphone request was rejected by the browser | Allow the microphone for whizi.io and retry |
| `This browser does not support realtime voice.` | No `getUserMedia` or no `AudioContext` | Use a browser that has both |
| `Realtime voice is not configured yet. Please try again later.` | The voice provider key is not set on the server | Nothing you can do: retry later |
| `The realtime voice connection failed.` | A websocket error during the session | Retry, and check the network |
| `Voice sessions are limited to 15 minutes. Start a new session to continue.` | The session timer reached the 15 minute cap | Start a new session |

The Whizi server only mints a short-lived session token, and the browser then opens a websocket straight to the voice provider. No audio passes through the Whizi backend. That is why microphone failures are entirely local to your browser, while the strings in the connection section come from the token mint on the server or from the provider upstream of it.

#### Voice will not start: the account gates

Three refusals happen before any audio is captured, and none of them are bugs.

**`Sign in to Whizi to use voice mode.`** The voice overlay throws this when no token is available. A signed-out or expired session produces it. Sign in and reopen voice mode.

**`Upgrade your subscription to talk with Whizi in voice mode.`** This is HTTP 402 with code `subscription_required`, returned when a free account tries to start a voice session. Free accounts have an allowance of 0 voice minutes, so there is no partial access to fall back on. The cheapest plan with any voice allowance is Starter at $15.99/month, which includes 10 realtime voice minutes per month.

**`Your monthly voice minutes have been used up.`** and **`Your weekly voice minutes have been used up.`** These are HTTP 429 with code `voice_limit_reached`, thrown when the period voice reservation is refused. The noun is "weekly" on a weekly billing cycle and "monthly" otherwise, so which one you see tells you which cycle your plan is on.

The allowances, per monthly billing period:

| Plan | Voice minutes per month |
| --- | --- |
| Free | 0 |
| Starter | 10 |
| Pro | 80 |
| Powerhouse | 500 |

Weekly billing periods take a quarter of the monthly allowance unless the plan names its own weekly figure.

#### The microphone will not start

These four messages all come from the browser side of voice mode, before or during capture. Because audio never reaches Whizi servers, nothing about your account or plan affects them.

**`This browser does not support realtime voice.`** Shown when `getUserMedia` or `AudioContext` is unavailable. This is a capability check, not a permission problem, so no amount of granting permissions will clear it. Try a different browser.

**`Microphone access is blocked. Allow it in your browser to use voice mode.`** Shown when the microphone request rejects with `NotAllowedError` or `SecurityError`. Grant microphone access to the site and start the session again.

**`The microphone could not be started.`** The catch-all for any other microphone failure. If the permission is granted and the browser is capable, this is the message that points at the device itself or at another application holding the input.

**`Voice mode could not start.`** It is the fallback used when the start sequence throws with no message of its own, so it carries no detail to act on.

#### The connection fails or drops

Starting a session has two steps that can fail independently: the server mints a token, then the browser opens a websocket. The strings tell you which step broke.

Failures while minting the token:

| Message | Status and code | Trigger |
| --- | --- | --- |
| `Realtime voice is not configured yet. Please try again later.` | 503 `voice_live_unavailable` | The provider API key is not set on the worker |
| `Realtime voice is temporarily rate-limited. Please retry shortly.` | 429 `provider_rate_limited` | The upstream token mint answered 429 |
| `Whizi could not start a realtime voice session.` | 502 `voice_live_unavailable` | Any other upstream failure on the token mint |
| `Whizi received an invalid realtime voice token.` | 502 `voice_live_unavailable` | The mint succeeded but returned no token name |
| `Realtime voice is not available in this build yet.` | 503 `voice_not_configured` | The shared API base is missing, or the voice domain is not enabled for this build |
| `Whizi received an invalid realtime voice session.` | 502 `invalid_voice_live_response` | The session payload failed validation, including a websocket URL that does not start with the expected provider host |
| `Request failed ({status}).` | Varies | The client fallback when a voice error body cannot be read |

The first four rows come from the token mint on the Whizi worker or from the provider upstream of it. No client setting, browser change, or plan change clears them. Retry shortly, and if the same string persists, contact support with the exact message and the time.

Failures on the websocket itself:

| Message | Trigger |
| --- | --- |
| `Whizi could not reach the realtime voice service.` | The websocket closed before setup completed |
| `The realtime voice connection was closed.` | The websocket closed after setup completed |
| `The realtime voice connection failed.` | A websocket error event fired |

The overlay labels a failed session with the state `Connection error`. A close before setup points at reachability (a network that blocks the websocket, for instance), and a close after setup points at an interruption mid-session.

Retry loops have their own ceiling. Starting or ending a voice session is limited to 6 requests per minute each, so hammering the start button swaps the problem you had for a 429. See [rate limits](https://whizi.io/docs/rate-limits) for the other windows a retry pays into.

#### The session ended on its own

Four endings are deliberate, and each has its own message.

**`Voice sessions are limited to 15 minutes. Start a new session to continue.`** A single voice session is capped at 15 minutes (900 seconds) regardless of how much of your allowance is left. This message fires when the timer reaches that full maximum. Starting a new session is the intended path, not a workaround.

**`This session used the last of your plan's voice minutes for this period.`** The start call reserves seconds and returns how long this particular session may run, and that figure shrinks when your remaining minutes are low. When the server shortened the session below the 15 minute maximum, this message is what you get at the end instead of the cap message.

**`Whizi ended the session after a couple of quiet minutes. Tap retry when you want to keep talking.`** The idle timeout elapsed with no speech. This is the message to look for if your symptom is "it never heard me": a session that captures no speech at all ends this way rather than sitting open. If you see this while you were definitely talking, the problem is upstream of the timeout, in the microphone section above.

**`The realtime voice session is ending. Start a new session to continue.`** The upstream service sent a `goAway` frame, which is the provider asking the connection to wind down. Start a new session.

The session token is minted with a 16 minute lifetime, one minute of slack past the 15 minute session cap. A session cannot outlive its token, and the cap is always what you hit first.

#### Reading the error yourself

If you are looking at a network trace or reporting a problem, the `code` in the error body is the more useful half of it. These are the codes voice mode can produce: `subscription_required`, `voice_limit_reached`, `voice_live_unavailable`, `provider_rate_limited`, `voice_not_configured`, `invalid_voice_live_response` and `rate_limit_reached`.

For what happens to what you said, the training statement covers voice explicitly: Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models, and does not sell that content as training data. The full detail is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

Key points:

- No audio passes through Whizi servers, so microphone failures are local to your browser
- The server side failures all come from one place: minting the session token
- Why a session ends at 15 minutes, or sooner, or after a quiet stretch

Checklist:

- Free accounts get 0 voice minutes: the 402 upgrade message is the gate, not a bug
- Allowances per month: Starter 10, Pro 80, Powerhouse 500 minutes
- "Weekly" in a limit message means the account is on a weekly billing cycle
- Microphone errors are entirely browser side: no audio passes through Whizi servers
- `This browser does not support realtime voice.` is a capability check, not a permission one
- Token mint failures (503 and 502) are server side: retry later or contact support
- A session is capped at 15 minutes regardless of remaining allowance
- A session with no speech at all is ended by the idle timeout
- Starting a session is rate limited to 6 per minute, so retry loops backfire

FAQ:

Q: Whizi says my microphone is blocked but it works in other apps. What now?
A: That exact string, `Microphone access is blocked. Allow it in your browser to use voice mode.`, is shown when the microphone request rejects with `NotAllowedError` or `SecurityError`. Grant microphone access to the site and start the session again. If what you see instead is `The microphone could not be started.`, permission was not the refusal: that message is the catch-all for any other microphone failure.

Q: Voice mode says it is not configured. Is that my account?
A: No. `Realtime voice is not configured yet. Please try again later.` is HTTP 503, returned when the voice provider API key is not set on the server, and `Realtime voice is not available in this build yet.` is the client-side equivalent when the voice domain is not enabled for that build. Neither is fixable from your side. Retry later, and if it persists, contact support with the exact string.

### AI chat with web search: how Whizi gives models live sources

URL: https://whizi.io/docs/web-search/
Updated: August 2026

Quick answer: Web search in Whizi is on by default, so the model you selected answers from current sources instead of only from its training data. It costs no credits and works with every model in the catalogue. The toggle is permission rather than a command: Whizi searches on the questions that need the web and skips the rest.

Search is on by default and any model can answer from current sources. How Whizi decides when to search, what happens when you paste a link, and when to switch it off.

#### The short answer

Web search is on by default. The model you selected answers from current sources instead of only from its training data, it costs no credits, and it works with every model in the catalogue.

The toggle in the composer is permission rather than a command. With it on, Whizi asks the model whether this particular question needs the web before it spends a search, so a request for a poem does not trigger one and a question about this week does. Switching it off tells the model it has no way to look anything up, which is worth doing when you want an answer drawn only from the conversation.

Messages that could not possibly need a page skip even that question. A greeting, a thank you, a yes, a "keep going" and anything else that asks nothing of the world is recognised before the model is called, so those turns carry no search machinery at all.

That last part is the useful bit. Most models have no retrieval of their own, so a question about this month is normally answered from a training set that ended months ago. The toggle puts a search layer in front of whichever model you picked, which means you can ask Claude, or [a Llama model](https://whizi.io/docs/using-llama-in-whizi), or a small fast model, a question about today and get a sourced answer.

#### When it changes the answer

Three categories where the difference between search on and search off is not subtle.

**Anything with a date in it.** Prices, model releases, product launches, company news, version numbers. Without search, a model answers confidently from whatever was true when it was trained, and gives you no signal that the information is stale. This is the single most common way AI answers go quietly wrong.

**Anything obscure or specific.** A small company, a niche tool, a local regulation, a particular error message. Training data covers these thinly or not at all, and thin coverage is exactly the condition under which models invent plausible detail.

**Anything you intend to act on.** Not because search makes the model smarter, but because it produces sources you can open. A claim with a link you can check is a different object from a claim without one, even when both turn out to be correct.

#### When to leave it off

Search is not free of cost to answer quality, which is why the toggle exists at all.

For reasoning, writing and analysis of something you supplied, search adds nothing and can actively distract. If you have uploaded a document and want the model to reason about that document, pulling in web results introduces material that is not yours and was not asked about. Same for creative writing, code review of your own code, and working through a problem step by step.

The rough rule: switch it on when the answer depends on facts about the world, switch it off when the answer depends on thinking about something already in the conversation.

| What you are asking about | Leave search on | Why |
| --- | --- | --- |
| Prices, releases, launches, company news, version numbers | On | Without it the model answers from training data and gives no signal that the information is stale |
| A small company, a niche tool, a local regulation, a particular error message | On | Training data covers these thinly, and thin coverage is when models invent plausible detail |
| Anything you intend to act on | On | It produces sources you can open, which is a different object from a claim without one |
| Reasoning about a document you uploaded | Off | Retrieved pages introduce material that is not yours and was not asked about |
| Creative writing, or code review of your own code | Off | The answer depends on thinking about something already in the conversation |

#### Paste a link and it gets read

Put a URL in a message and Whizi opens the page before the model answers, then hands it the text of that page. It works whether search is on or off, because pasting a link is you telling Whizi what to read rather than asking it to go looking.

Up to three links per message are read. Scripts, styling and navigation are stripped out, so the model gets the article rather than the cookie banner and the menu.

Some pages cannot be read: sites that block automated readers, pages behind a login or a paywall, and pages that build themselves in the browser with JavaScript. When that happens the model is told the page could not be opened and says so, which is the point. The old failure was a confident summary of a page nothing had actually seen. The rest of the failure surface, including a toggle that is on but returns nothing, is in [web search not working](https://whizi.io/docs/web-search-not-working).

#### Reading the sources properly

A sourced answer is more checkable, not automatically more correct. Two failure modes survive retrieval and are worth watching for.

The first is a source that exists but does not support the claim. The model retrieved a page, and the page is real, but the specific number or statement attributed to it is not in it. Opening one or two cited links on any answer you plan to rely on catches this quickly.

The second is a source that is itself wrong or out of date. Search does not evaluate credibility, so a stale blog post ranks like an authoritative one. For prices and specifications especially, a vendor page beats a roundup article, and both beat a forum comment.

If a claim matters, the fastest verification is to ask a second model the same question with search on, in the same conversation. Where two independent retrievals agree on a specific number and a source, confidence is reasonable. Where they diverge, that is your finding.

Key points:

- Web search works with any model, not only the ones that ship with it
- The three question types where it changes the answer completely
- What happens when you paste a URL into the composer

Checklist:

- The toggle is in the composer, next to attach, and it is on by default
- Web search costs no credits
- It works with every model, including ones with no native retrieval
- Whizi searches only on the questions that need it, not on every message
- Paste a URL and the page is read for you, search on or off
- Switch it off for reasoning about a document you uploaded
- Open at least one cited link on any answer that matters
- Cross-check important claims by asking a second model with search on

Steps:

1. Open a chat: Start a new conversation or continue an existing one.
2. Check the search toggle: The web search control sits in the composer, next to the attach button. It is on by default, and green when it is active for this conversation.
3. Ask a question with a time element: Anything about current prices, releases, or recent events. Whizi searches when the question needs it and skips the search when it does not.
4. Or paste a link: Whizi opens the page and answers from what is on it, whether search is on or off.
5. Check the sources: Read what the model cited before you rely on the answer.

FAQ:

Q: Can Claude search the web in Whizi?
A: Yes. Web search applies to whichever model you selected, so Claude Sonnet 5 or Claude Opus 5 answers from current sources. This is one of the practical advantages of a multi-model workspace: you are not limited to whichever models happen to ship with their own retrieval.

Q: Does web search use extra credits?
A: No. A message costs the same whether search is on or off, because the credit rate is set by the model alone. The only thing that changes the cost of a message is which model answers it.

Q: Why did the model give me outdated information?
A: Either search is switched off in that conversation, or the model judged the question answerable without it and answered from training data. Check the toggle first. If it is on and the answer carries no sources, ask again and say you want it looked up; if the answer changes, the first one was stale.

Q: Should I leave web search on all the time?
A: For most work, yes, and it is on by default: Whizi only searches when a question needs the web, so leaving it on costs nothing on the messages that do not. Switch it off for a conversation about material already in the chat, such as a document you uploaded, where retrieved pages pull the model away from your own file.

### Web search not working in Whizi: why it did not run

URL: https://whizi.io/docs/web-search-not-working/
Updated: August 2026

Quick answer: In most cases no search was ever issued. With the toggle on, Whizi first sends the model a cheap probe tool of about 50 tokens, and a real web search only runs on a turn where the model calls that tool. If it judged it could answer already, you get an unsourced answer with the toggle still green.

The search toggle is on but no search happened, or the results came back thin. What Whizi actually does on a search turn, the exact error strings, and the fix for each.

#### The short answer

In most cases no search was ever issued. With the toggle on, Whizi first sends the model a cheap probe tool of about 50 tokens, and a real web search only runs on a turn where the model actually calls that tool. If the model judged it could answer from what it already knew, no search happens and you get an unsourced answer with the toggle still showing green.

The causes, in the order worth checking:

| What you saw | Most likely cause | What to do |
| --- | --- | --- |
| Toggle on, answer has no sources | The model never called the probe tool, so no search was issued | Nothing on your side forces a search: the model decides per turn |
| Toggle looks off in this chat | The toggle is stored per conversation, not per account, and search only ships on for new chats | Switch it on again in this conversation |
| You pasted a link and got a summary that does not match the page | The page could not be opened, or was read and misread | A page Whizi cannot open is reported as unread in the answer. Sites that block automated readers, logins, paywalls and JavaScript-rendered pages are the usual causes |
| Only two or three sources came back | An ordinary search is capped at 3 results | Nothing to change: 3 per search and 3 in total is the cap |
| An error mentioning a boolean | The request body was malformed | Reload the page and resend |
| The whole message errored | A model provider failure | Retry, then contact support if it repeats |

#### Why the toggle being on does not mean a search ran

A request runs in one of three search modes: off, probe, or native.

**Off mode** is both the toggle being off and a message that cannot need the web. Greetings, acknowledgements and "keep going" are recognised before the model is called, so those turns carry no search tool and cost nothing extra. Everything else with the toggle on goes to probe mode.

**Probe mode** is the cheap stand-in. The model is handed an ordinary function it can call, costing roughly 50 tokens, and only a message where it actually calls that function gets re-issued with real search attached. On messages that never needed the web, probe mode is about 46 times cheaper than native, and on the ones that do need it, it costs about 0.8 percent extra.

**Native mode** runs the provider real web search tool directly. It is not free: the tool definition injects a fixed preamble of roughly 4,400 input tokens. The same two-message prompt measured 36 tokens without the tool and 4,472 with it.

The toggle permits a search. The model decides whether to run one, which is why a question like "what is a good CRM" can come back with no sources while "what did this company announce this week" comes back searched, in the same conversation, with the toggle untouched.

Web search is a per-request user toggle rather than a per-model capability flag, so there is no published list of models that can browse and models that cannot.

#### The exact error strings and what each one means

Whizi clients render the backend `error.message` directly, so the string in front of you is the one the route emitted. If you have one of these in front of you, this is what it is telling you.

| Error string | HTTP and code | Trigger |
| --- | --- | --- |
| `webSearchEnabled must be a boolean.` | 400, `invalid_generate_request` | The website send route received a `webSearchEnabled` field that was not true or false |
| `webSearch must be a boolean.` | 400, `invalid_chat_request` | The app chat route received a `webSearch` field that was not true or false |
| `Web searches failed. You have not been charged.` | Written onto a failed run | Fewer than the minimum number of searches completed on a Deep Research run |

The two boolean errors report a malformed request. The value the client sent for the search flag was not a boolean, and nothing about your account, plan or model caused it. Reload the page and send the message again. If a reload does not clear it, contact support.

`Web searches failed. You have not been charged.` is written onto a failed Deep Research run when fewer than the minimum number of searches complete. The credits are not taken. Try the request again later.

#### The results came back, but there were only three

An ordinary search request is capped at 3 results per search and 3 results in total, at low search context size. A thorough looking question and a trivial one both come back with at most three pages behind them.

One route behaves differently. The Canadian CPA agent is the only route that forces web search on every turn, capped at 7 searches, and its searches are restricted to an allowed domain list and an approximate Canadian user location. On that agent, results from a narrow set of sites are the allowed domain list at work.

If the answer is out of date rather than unsourced, or you want to check a figure against more than one retrieval, that is covered in [using web search](https://whizi.io/docs/web-search) and in [the troubleshooting hub](https://whizi.io/docs/troubleshooting).

#### When the whole message fails

A search turn can also die for reasons that have nothing to do with search. These strings come from the model provider layer and mean the generation itself failed.

| Error string | HTTP and code | Meaning |
| --- | --- | --- |
| `The model is temporarily rate-limited. Please retry shortly.` | 429, `provider_rate_limited` | The upstream provider refused the connection, not Whizi |
| `The model provider rejected the request.` | 502, `provider_error` | The provider error body could not be read |
| `The model provider returned an empty stream.` | 502 | The upstream response was fine but carried no body |
| `The model stream was interrupted.` | In-stream error event | The stream failed part way through |
| `Too many requests. Please wait and try again.` | 429, `rate_limit_reached` | You hit a Whizi rate window |

Chat messages are limited to 10 per minute and 60 per hour per user, flat across every plan, with a shared ceiling of 30 requests per minute across all routes. A refused request costs nothing: rate-limit refusals are unwound, so a 429 never spends a message off your allowance.

None of these have a setting on your side that prevents them. Wait a moment and send again. If the same model keeps failing across several minutes, switch to a different model to keep working and contact support so it can be looked at on your account.

Key points:

- Why a green search toggle does not guarantee a search actually ran
- The two verbatim error strings that mean the request was malformed
- The three result cap that makes a working search look like a broken one

Checklist:

- A green toggle permits a search, and the model decides whether to run one
- Probe mode costs about 50 tokens, and only a turn where the model calls it gets real search
- An ordinary search returns at most 3 results
- The two boolean errors are malformed requests: reload and resend
- Search is a per-request toggle, not a per-model capability flag
- Provider errors and 429s are transient: wait, retry, then contact support

FAQ:

Q: I turned on web search in Whizi and it did not search. Why?
A: Because the model decided it did not need to. With the toggle on, Whizi sends a cheap probe tool of about 50 tokens alongside your message, and a real search is only issued on a turn where the model actually calls that tool. The design keeps the cost down, since native provider search injects roughly 4,400 input tokens of preamble on every request.

Q: What does "webSearchEnabled must be a boolean." mean?
A: It is an HTTP 400 with the code `invalid_generate_request`, returned when the send request carries a `webSearchEnabled` value that is not true or false. It reports a malformed request, so nothing about your account, plan or model caused it. Reload the page and send again, and contact support if a reload does not clear it. The app chat route has the equivalent string `webSearch must be a boolean.` with the code `invalid_chat_request`.

Q: What does "Web searches failed. You have not been charged." mean?
A: It is written onto a failed Deep Research run when fewer than the minimum number of searches complete. The wording is literal: the credits are not taken. There is no configuration change that prevents it, so the only useful response is to run the request again later.

### Where does Whizi store my data? Servers, retention windows, and what a lapsed plan keeps

URL: https://whizi.io/docs/where-whizi-stores-data/
Updated: August 2026

Quick answer: Whizi stores your account, conversations, files and generated media on Cloudflare, using Workers, D1, R2 and Durable Objects. Sign-in is Clerk, website payments are Stripe, and image generation is fal.ai. Processing may happen in countries other than your own, including Canada and the United States. Attachments and generated media are configured to expire after up to 30 days.

Which service holds each kind of Whizi data, which countries processing may happen in, how long every category is kept, and what survives when a subscription ends.

#### The short answer

Whizi stores your account, conversations, files and generated media on Cloudflare, using Workers, D1, R2 and Durable Objects. Authentication and account management are provided by Clerk, website payment processing is handled by Stripe, and image generation jobs and their output are handled by fal.ai.

Whizi and its service providers may process information in countries other than your own, including Canada and the United States. Where required, Whizi uses appropriate contractual or legal safeguards for international transfers.

| Data | Where it lives | How long |
| --- | --- | --- |
| Conversations and messages | Cloudflare, on the stack that runs synchronized chats | Retained while the account remains active, unless you delete the content or the account |
| Uploaded attachments | Cloudflare R2 object storage | Configured to expire after up to 30 days |
| Generated media | Cloudflare R2 object storage | Configured to expire after up to 30 days |
| Pinned project files | Cloudflare, with the project | Stored until deleted, these do not expire |
| Saved memory | Cloudflare, with the account | Kept until deleted |
| Entitlement and usage state | Cloudflare, with the account | Kept while the account exists |
| Sign-in identity | Clerk | Removed by account deletion |
| Card details | Stripe, for website payments | Whizi does not store full payment-card numbers |
| Recent chats on a phone | A local cache on the device, up to 40 conversations | Cleared by the in-app deletion flow, by clearing app storage, or by uninstalling |
| Deleted database rows | Cloudflare D1 point-in-time recovery history | Up to 30 days before aging out |

#### The service providers named in the policy

The privacy policy names the providers Whizi uses to operate the service. The ones that touch stored content or account state:

- **Clerk** for authentication and account management.
- **Cloudflare Workers, D1, R2, and Durable Objects** for hosting, synchronized chats, media, account state, usage limits, reports, and security and operational processing.
- **OpenRouter and routed model providers** for AI prompts, responses, attachments, search context, transcription, speech, and related model processing.
- **fal.ai** for image-generation prompts, jobs, and generated media.
- **Stripe** for website payment processing.
- **Expo** for optional transactional push-notification delivery.

Two of these are visible in the product when something goes wrong. `Media could not be stored.` is raised when the R2 write returns no object, which is the storage layer failing rather than the model, and `Invalid or expired session token.` is HTTP 401 `unauthorized` when token verification against the Clerk JWKS throws.

#### Retention, category by category

The policy commits to a ceiling rather than a lifetime. Uploaded attachments and generated media "are configured to expire after up to 30 days and may be deleted sooner by you or during account deletion". Whizi's shared media constant is what currently sets that ceiling, at 30 days, so 30 days is the longest an object is configured to survive rather than the age it is guaranteed to reach.

The observable consequence is a specific refusal rather than a broken image. An object past its retention window returns `Media has expired.`, HTTP 404 `media_expired`, while `Media not found.` is HTTP 404 `media_not_found` when the media id does not resolve for the caller at all. Generated images are served through signed URLs, so a link inside an old conversation was never a permanent address.

Two retention facts that are easy to miss:

- **Pinned project files do not expire.** Their bytes are stored until deleted, unlike chat attachments, because a pinned file has to ride along with every turn in the project.
- **A phone keeps its own copy.** The mobile app may keep a local cache of up to 40 recent conversations on a signed-in or guest device. That cache is not a separate cloud backup, and it is cleared by a successful in-app deletion flow, by clearing app storage, or by uninstalling.

Security and operational logs carry no published number of days. They are kept only as long as reasonably needed to operate and secure the service, investigate incidents, prevent fraud or abuse, resolve disputes, or comply with law.

#### What providers may cache after they answer

OpenRouter and routed AI or media providers may temporarily cache portions of a prompt or response to support multi-turn conversations, reduce latency, provide prompt caching, prevent abuse, or operate their services. Cache duration, retention, and any provider-controlled use vary by provider and endpoint.

Whizi does not represent that every provider follows the same retention period or a universal zero-data-retention setting, and does not claim it can instantly remove a provider cache that has already been created. Which provider receives a given message is covered in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

#### What a lapsed subscription keeps

Cancelling and deleting are two different actions, and only one of them removes data. Cancelling stops billing and ends access at the end of the paid term, while the account and its contents remain. Account information and synchronized chat history are generally retained while the account remains active, unless you delete particular content or delete the account, which is what makes resubscribing later a continuation rather than a fresh start.

What changes is the tier, not the storage: an account that syncs down to the free tier keeps its account record, chats, media and saved memory exactly where they were. The limits and messages that come with that tier are covered in [subscription not active](https://whizi.io/docs/subscription-not-active).

One retention rule works in your favour here: Whizi does not keep deleted chat content or uploaded or generated media merely because a payment record must be retained. Payment processors may separately retain transaction, receipt, tax, fraud, chargeback or dispute records for periods required by law or their compliance obligations, and Whizi may retain limited account, transaction, security, tax, fraud-prevention or dispute records when legally or operationally required.

The timing rule for cancelling itself, and where to do it, is in [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

#### Getting a copy, or getting it all removed

Both paths are covered on neighboring pages: [export chats and files](https://whizi.io/docs/export-chats-and-files) carries the export contents, and [what Whizi does with your data](https://whizi.io/docs/data-and-privacy) carries the account deletion section, with the item-by-item removal list and what survives deletion.

Key points:

- Which named service holds each kind of data, and the countries processing may happen in
- A retention window per category, from uploaded files to the local cache on a phone
- What a lapsed subscription keeps, and what it does not

Checklist:

- Storage runs on Cloudflare Workers, D1, R2 and Durable Objects
- Sign-in is Clerk, website payments are Stripe, image generation is fal.ai
- Processing may occur in countries other than your own, including Canada and the United States
- Attachments and generated media are configured to expire after up to 30 days
- Pinned project files are stored until deleted and do not expire
- A phone may cache up to 40 recent conversations locally, which is not a cloud backup
- Cancelling keeps the account and its contents; only deletion removes them

FAQ:

Q: Where are Whizi conversations physically stored?
A: On Cloudflare. The privacy policy names Cloudflare Workers, D1, R2 and Durable Objects as the providers used for hosting, synchronized chats, media, account state, usage limits, reports, and security and operational processing. Attachments and generated media sit in R2 object storage specifically, which is why a failed write surfaces as `Media could not be stored.`

Q: Can I choose which country my Whizi data is stored in?
A: The privacy policy makes no residency commitment. It states that Whizi and its service providers may process information in countries other than your own, including Canada and the United States, and that where required Whizi uses appropriate contractual or legal safeguards for international transfers.

Q: An image in an older Whizi chat will not load. What happened?
A: Its retention window most likely passed. Generated images are served through signed URLs rather than permanent addresses, so a link sitting inside an older conversation can stop resolving, and the request answers `Media has expired.`, HTTP 404 `media_expired`, instead of returning the bytes.

Q: If I cancel my subscription, do I lose my chats?
A: No. Cancelling stops billing and ends access at the end of the paid term, while the account and its contents remain, so resubscribing later picks up where you left off. Account information and synchronized chat history are generally retained while the account remains active. Removing the data is a separate action, account deletion, and Whizi does not keep deleted chat content or media merely because a payment record must be retained.

Q: Where is my data stored if I never created an account?
A: In the same place, under an anonymous device-bound guest profile. Creating an account is optional in the mobile app, and continuing without one needs no name or email address. The guest credential is stored on the device, and Whizi stores a one-way derived identifier so it can recognize the guest profile, synchronize data, verify subscription access, and later link that data to an account.

### Does Whizi have a developer API?

URL: https://whizi.io/docs/whizi-api/
Updated: August 2026

Quick answer: No. Whizi has no public or developer API, no API key issuance and no personal access tokens, on any plan including Powerhouse. Every backend route authenticates as an end user session rather than a third-party integration, so there is no credential to create, rotate or revoke. The terms prohibit automating product or account access.

No. Whizi issues no API keys and no personal access tokens, every route authenticates as an end user session, and the terms forbid scripted product access. What to use instead.

#### The short answer

No. Whizi has no public or developer API, no API key issuance, and no personal access tokens. Every route on the Whizi backend authenticates as an end user session rather than as a third-party integration, so there is no credential to create, paste, rotate or revoke. This is true on every plan, including Powerhouse.

The upstream provider credentials Whizi uses to reach the models are server side secrets on the backend and never leave it.

| What people ask for | Status in Whizi |
| --- | --- |
| A REST or streaming chat API | Does not exist |
| An API key or secret you can generate | Does not exist |
| A personal access token | Does not exist |
| A credential that lets an outside tool act on your account | Does not exist |
| A team, organization or workspace with seats | Does not exist |
| A full export of your own account data | Exists, as Export Data in Settings, guest sessions included |

If you need to call models from your own code, Whizi is not the tool for that job. Whizi is a chat workspace: one subscription, one picker, a 280+ model catalogue. The section below covers what is supported for getting data in and out.

#### What the terms of service say about scripting it

The absence of an API is not an oversight waiting for a workaround. Driving the product with a script is prohibited in the terms, in two separate places.

Section 3, user representations, has you warrant that "you will not access the Services through automated or non-human means, whether through a bot, script or otherwise".

Section 8, prohibited activities, lists it again from the other direction. You agree not to "Engage in any automated use of the system, such as using scripts to send comments or messages, or using any data mining, robots, or similar data gathering and extraction tools." The same section separately forbids you to "Systematically retrieve data or other content from the Services to create or compile, directly or indirectly, a collection, compilation, database, or directory without written permission from us", and, standard search engine and browser usage aside, to "use, launch, develop, or distribute any automated system, including without limitation, any spider, robot, cheat utility, scraper, or offline reader that accesses the Services".

Those restrictions apply to automating the product, an account or user-generated content. The same section expressly permits search engines and artificial intelligence systems to crawl publicly accessible marketing, documentation, reference, pricing and dataset pages for search indexing, real-time retrieval, citation, model training and fine-tuning. Crawlers must follow `robots.txt`, stay out of accounts and private areas, and avoid excessive requests that disrupt the service. That public-page permission does not create an API or authorize a script to send chat messages through a Whizi subscription.

#### The errors a script hits

Because the backend only knows end user sessions, anything automated runs into session and rate limit refusals rather than an integration error. A caller with no session token, or an expired one, gets the ordinary 401 `unauthorized` refusals catalogued in [login and account access](https://whizi.io/docs/login-and-account-access). A caller that does authenticate as a browser session is then refused by the ordinary per user rate limits, which are in [rate limits and what they mean](https://whizi.io/docs/rate-limits).

One refusal is specific to scripts. More than 3 device or session keys in 24 hours is the guard behind HTTP 403 `account_sharing_limit`, so a script that starts a fresh session per run looks exactly like a shared account and gets refused on that basis. It clears on its own once the window rolls forward.

#### What to use instead

**Getting all your data out.** Export Data sits in Settings under Account and returns the whole account as one downloadable JSON file, guest sessions included. The mechanics of that file are in [exporting your chats and files](https://whizi.io/docs/export-chats-and-files).

**Getting data in.** Chat import accepts exactly two sources, ChatGPT and Claude. The export file never reaches the server: the browser reads `conversations.json` or the raw export `.zip`, parses it, and posts normalized conversations. Every imported conversation is keyed by account plus source plus source id, so re-importing the same export or clicking Import twice lands the same chats once. Import Chats sits in Settings under Account.

**Publishing an answer somewhere else.** A conversation can be published behind an unguessable public link that is read-only, stripped of user ids, emails and internal fields, and revocable at any time. See [share a conversation](https://whizi.io/docs/share-a-conversation).

**Getting structured output rather than prose.** A chat answer can produce a real `.xlsx` or `.docx` in your browser rather than a payload, which is documented in [exporting your chats and files](https://whizi.io/docs/export-chats-and-files).

**Reusable instructions across conversations.** There is no account-level custom instructions setting, but a project carries its own instructions, up to 32,000 characters, applied to every conversation in it, plus up to 10 pinned files that ride along as prompt text on every turn. That covers the "same system prompt every time" use case without a script.

The question that usually arrives alongside this one is whether Whizi has team or workspace accounts. It does not, and that has its own page: [team and workspace plans](https://whizi.io/docs/team-accounts).

Key points:

- No public API, no API key issuance and no personal access tokens, on any plan
- What the terms say about product automation and the separate permission to crawl public pages
- The supported ways to move data in and out: export, import, share links and generated files

Checklist:

- No public or developer API exists, on any plan
- No API keys and no personal access tokens are issued
- Every backend route authenticates as an end user session
- Provider credentials are server side secrets and never leave the backend
- Automating the product or an account is prohibited, while crawling public pages is expressly permitted
- A script hits session, sharing and rate limit refusals, not an integration error
- Account data export, chat import, share links and generated files are the supported routes in and out
- There are no team, organization or workspace accounts either

FAQ:

Q: Is there a Whizi API key on any plan?
A: No. No tier issues one. Starter, Pro and Powerhouse differ in what the plan opens and how much of it you get, and none of them adds a credential, an integration surface or a developer console. The tier differences themselves are in [plans and limits](https://whizi.io/docs/plans-and-limits).

Q: Can I use my Whizi subscription in my own app or script?
A: No. There is no endpoint to call and no credential to call it with, and the terms prohibit automating product or account access. They separately permit search engines and AI systems to crawl public marketing, documentation, reference, pricing and dataset pages, but that permission does not authorize a script to send chat messages through your subscription.

Q: Whizi calls model providers, so can I get at those keys?
A: No. The upstream provider credentials Whizi uses to reach the models are server side secrets on the backend and never leave it.

Q: Can I schedule the data export to run on its own?
A: There is nothing to call on a schedule. The export is a button in Settings under Account rather than a route you can hold a credential for, and the terms prohibit driving the product with a script. What the file contains is in [exporting your chats and files](https://whizi.io/docs/export-chats-and-files).

Q: Could a third-party tool connect to my Whizi account?
A: Not through any credential you can issue. Nothing in Whizi lets an outside tool hold delegated access to your account, so anything claiming to connect would be driving a browser session, which is what the terms prohibit and what the account sharing guard refuses.

Q: Is a developer API on the roadmap?
A: As of August 2026 there is no public API, no API key issuance and no personal access token issuance in the product. The terms prohibit automating product or account access while separately permitting search engines and AI systems to crawl public pages.

### Whizi FAQ: the questions people ask most

URL: https://whizi.io/docs/whizi-faq/
Updated: August 2026

Quick answer: Whizi is a multi-model AI chat workspace: one subscription and one login covering 280+ AI models from OpenAI, Anthropic, Google and others in a single chat window. There are three individual plans, at $15.99, $29.99 and $49.99 per month, and every new account starts with a 7 day trial.

One page of direct answers: what Whizi is, what it costs, which models it includes, how credits work, what it does with your data, and four things it does not do.

#### The short answer

Whizi is a multi-model AI chat workspace: one subscription and one login covering 280+ AI models from OpenAI, Anthropic, Google, xAI, DeepSeek, Meta, Mistral, Moonshot, Qwen, Z.ai and others, in a single chat window. There are three individual plans, and every new account starts with a 7 day trial for $0.99.

The nine facts most people are looking for:

| Question | Answer |
| --- | --- |
| What does it cost | $15.99, $29.99 or $49.99 per month for Starter, Pro or Powerhouse |
| Is there a free plan | No ongoing free plan. A new account gets a lifetime allowance of 7 messages |
| How many models | 280+, tiered across the three plans |
| What is a credit | One message on the house model Whizi AI. Plans get 400, 2,000 or 8,000 per month |
| Do you train on my chats | No, and Whizi does not sell that content as training data |
| Can I import from ChatGPT | Yes, and from Claude. Those are the only two sources accepted |
| Is there a mobile version | Web, iOS and Android, included in every plan on one login |
| Is there a developer API | No. No API keys and no personal access tokens, on any plan |
| Are there team plans | No. Three individual tiers only |

Everything below is the longer version of those answers, each one linking the page that owns it.

#### Price, trial, cancellation and refunds

**How much does Whizi cost?** Starter is $15.99/month, Pro is $29.99/month and Powerhouse is $49.99/month. Billed annually the rates are $10.99/mo billed annually at $131.88 for Starter, $19.99/mo billed annually at $239.88 for Pro, and $34.99/mo billed annually at $419.88 for Powerhouse. Every number, plus every quota attached to each plan, is in [plans and limits](https://whizi.io/docs/plans-and-limits).

**Is there a free plan?** There is no ongoing free tier. A new signed-in account gets a lifetime allowance of 7 messages, and that counter never resets. A free account is not model-gated, so those 7 messages can go to any text model in the catalogue rather than a cut-down free list. When they run out the request is refused with "Your free messages are used up. Start a subscription to keep chatting." The refusal and the reset rules are on [out of credits](https://whizi.io/docs/out-of-credits).

**What does the trial include, and what happens after 7 days?** New accounts get 7 days for $0.99 on the plan chosen at signup, with that plan fully unlocked. It is a paid trial rather than a free one. It converts to the selected plan unless you cancel before the 7 days are up. The trial terms sit on [plans and limits](https://whizi.io/docs/plans-and-limits).

**How do I cancel?** Cancel from account settings if you subscribed on the web, and through Apple or Google if you subscribed inside the iOS or Android app, because the store owns that billing relationship. Cancellation takes effect at the end of the current paid term, so you keep full access until then rather than losing it the moment you cancel. Cancelling during the trial keeps access for the rest of the 7 days with nothing further charged. See [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

**Can I get a refund?** Whizi publishes no refund policy for a subscription bought on whizi.io, and the Terms of Service contain no refund clause for one. The only remedy the Terms describe is cancellation, which stops the next charge rather than returning the last one, and the Cancellation clause names whizihelp@gmail.com for anyone unsatisfied. A subscription bought inside the mobile app is the store decision under the store terms, not a Whizi one. The clause is quoted in full on [refunds](https://whizi.io/docs/refunds).

#### Models and credits

**Which AI models does Whizi include?** 280+ models across OpenAI, Anthropic, Google, xAI, DeepSeek, Meta, Mistral, Moonshot, Qwen, Z.ai and other providers, all reachable from the same conversation. Starter reaches a narrow set of fast 1 credit models plus Auto, Pro adds roughly 37 more, and Powerhouse opens the full catalogue. The per-plan lists are in [models](https://whizi.io/docs/models).

**Which plan do I need for Claude?** Pro is the lowest paid plan that reaches Claude, because Starter includes no Anthropic model at all. A free account is not model-gated, so a Claude message is reachable inside the 7 message lifetime allowance. Pro reaches the Claude Sonnet class plus one Haiku, and the Claude Opus class is Powerhouse only. [Using Claude in Whizi](https://whizi.io/docs/using-claude-in-whizi) has the model by model table.

**What is a credit, and how many do I get?** One credit is one message on the house model Whizi AI, and what a message costs depends only on which model answered it, never on message length or conversation length. Starter gets 400 credits per month, Pro 2,000 and Powerhouse 8,000. Roughly half the catalogue costs 1 credit per message, the current frontier models are 20, and a handful of older Opus rows sit as high as 80, so the default model you choose decides whether the allowance lasts. The full ladder is on [credits](https://whizi.io/docs/credits).

**What spends no credits at all?** Uploading a file costs no credits, and switching models mid-thread carries no charge of its own: each turn is simply charged at the credit cost of the model that answers it. Image generations and voice minutes are counted on their own monthly meters and spend no credits.

**What does Auto do?** Auto reads your message, sorts it into one of six fixed rungs, and answers on the model pinned to that rung, never above what your plan can open. It has no credit price of its own: a turn costs whatever the model Auto picked costs, which is always 1 credit on Starter, 1 to 10 on Pro and 1 to 20 on Powerhouse. The whole ladder is on [how auto routing works](https://whizi.io/docs/how-auto-routing-works).

#### Getting your work in, and working on it

**Can I import my ChatGPT or Claude history?** Yes, and those two are exactly the sources Whizi accepts. Open Settings, choose Import Chats under Account, and pick the `conversations.json` file out of the export; the same dialog handles both vendors. The export file is never uploaded, because your browser parses it and posts only normalized conversations, and running the import twice lands the same chats once. Walkthrough on [import chats from ChatGPT](https://whizi.io/docs/import-chats-from-chatgpt).

**What is a project?** A folder of conversations that also carries pinned files and its own custom instructions, both of which are folded into the model prompt on every message in that project. Projects are available to any signed-in account, and the caps are abuse guards rather than a paid gate: 50 projects per account, 10 pinned files per project at 10 MB each, and 32,000 characters of project instructions. See [use projects](https://whizi.io/docs/use-projects).

**What files can I upload?** Four groups: images, PDFs and Word documents, Excel spreadsheets, and around forty plain text and source code formats. Starter accounts upload on the same terms as the paid tiers above them, at no credit cost. The caps that bind are per prompt rather than per plan: the web composer stages up to 6 files at 10 MB each, while a single prompt carries at most 4 attachments and 20 MB in total. The complete extension list is on [supported file types](https://whizi.io/docs/supported-file-types).

#### Voice, search, images and mobile

**Does Whizi have voice mode?** Yes. It is a real-time spoken conversation that you can interrupt, and the session produces a transcript that lands in your chat history like a typed one. Every paid plan includes voice minutes on its own monthly meter: 10 on Starter, 80 on Pro and 500 on Powerhouse. Voice minutes do not spend credits. See [voice mode](https://whizi.io/docs/voice-mode).

**Can Whizi search the web?** Yes, and it is on by default. The toggle in the composer stays on for the conversation and grants permission rather than forcing a search: a request runs in one of three modes, off, a cheap probe that offers the model an ordinary search function, and native, where the real provider search tool runs. The probe is what makes leaving it on cheap, since a message that never needed the web never pays for the full search preamble. Paste a URL and the page is read for you either way. See [web search](https://whizi.io/docs/web-search).

**Can Whizi generate images?** Yes, inside the conversation rather than in a separate tool, which means the model already has the context you have been building. Image generations are their own monthly meter: Pro gets 100 and Powerhouse gets 500, and Starter gets none, so image generation means Pro at minimum. See [generate images in chat](https://whizi.io/docs/generate-images-in-chat).

**Does Whizi work on my phone?** Yes: web, iOS and Android, all three on one login and inside the plan you already pay for, Starter included. Nothing about the account splits per device. The meters are shared, so voice minutes spent on a phone come out of the same monthly allowance as voice minutes spent at a desk. See [Whizi on mobile](https://whizi.io/docs/whizi-on-mobile).

#### Privacy and your data

**Do you train AI models on my conversations?** No. Quoting the privacy policy exactly: "Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models. Whizi does not sell this content as training data."

**Where does my message go?** Whizi is a workspace in front of models built by other companies, so a message you send to Claude reaches Anthropic and a message you send to GPT reaches OpenAI. Nothing is sent to a model you did not select. Providers may temporarily cache portions of a prompt or response under their own policies, and Whizi does not claim every provider follows the same retention period. See [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

**How long is my data kept, and can I delete it?** Uploaded attachments and generated media are configured to expire after up to 30 days. Conversations and messages are kept until you delete them or delete the account, and deleting one conversation or one uploaded file does not require deleting the account. Account deletion sits under Settings behind a two-step confirmation, with instructions at whizi.io/account-deletion. Cancelling a subscription is a different action: it stops billing and ends access at the end of the paid term while the account and its contents remain. See [where Whizi stores your data](https://whizi.io/docs/where-whizi-stores-data).

#### What Whizi does not offer

Four of the things Whizi does not offer:

| Question | Answer |
| --- | --- |
| Is there a developer API | No. No public API, no API key issuance and no personal access tokens, on any plan including Powerhouse |
| Are there team or enterprise accounts | No. No team, workspace, organization or seat concept anywhere in the product. A team of five is five accounts and five subscriptions |
| Does it work offline | No. Every model reply is a network round trip, and the conversation itself lives on the server |
| Is there a regenerate button | No. Editing your own message and sending it again replaces the reply and files the previous attempt as a version |

The reasoning and the supported alternatives for each: [does Whizi have a developer API](https://whizi.io/docs/whizi-api), [team accounts](https://whizi.io/docs/team-accounts), [use Whizi offline](https://whizi.io/docs/use-whizi-offline), and [regenerate and edit messages](https://whizi.io/docs/regenerate-and-edit-messages).

Key points:

- Price, trial, cancellation and refunds answered in one place
- What a credit is, how many each plan gets, and what spends nothing
- Four straight no answers: developer API, team plans, offline use, regenerate

Checklist:

- Three individual plans: $15.99, $29.99 and $49.99 per month
- The entry point is a 7 day trial for $0.99, not a free tier
- A free account gets 7 lifetime messages, on any text model, and they do not reset
- One credit is one Whizi AI message; plans get 400, 2,000 or 8,000 per month
- Uploads, sharing and model switching cost no credits
- Whizi does not train Whizi-owned models on your content, or sell it as training data
- No developer API, no team plans, no offline mode, no regenerate button

FAQ:

Q: Can I compare two models side by side?
A: Yes, on Powerhouse. The same prompt goes to two models in parallel and neither model sees the answer from the other, so neither answer is anchored by the other. Each model counts as one message, so a comparison costs two turns rather than one. See [compare models side by side](https://whizi.io/docs/compare-models-side-by-side).

Q: Are there limits on how fast I can send messages?
A: Yes: 10 chat messages per minute and 60 per hour per user, plus a shared ceiling of 30 requests per minute across all routes. The three figures are the same on every plan, and a refusal is not charged. See [rate limits](https://whizi.io/docs/rate-limits).

Q: Can I get my conversations back out of Whizi?
A: Yes, two different ways. Export Data sits in Settings under Account and returns the account data as one downloadable JSON file covering chats and messages, memory, projects and their file metadata, media, share links, subscription and referral profile; the status of that export is on [does Whizi have a developer API](https://whizi.io/docs/whizi-api). Separately, an answer can produce a real .xlsx or .docx file card rather than a table stuck in a chat message, which is covered on [export chats and files](https://whizi.io/docs/export-chats-and-files).

### Whizi memory: what it remembers across your chats

URL: https://whizi.io/docs/whizi-memory/
Updated: August 2026

Quick answer: Yes. Whizi remembers durable facts you share, like your name, preferences and goals, and reuses them in later conversations. The facts belong to your account rather than one chat, so they follow you into a new conversation and into whichever model you pick. They are stored as a plain list of short lines, capped at 2,000 tokens.

Whizi keeps a short list of durable facts about you and replays it into later conversations. What earns a line, what is deliberately never stored, and how to see or clear the list.

#### The short answer

Yes, Whizi remembers durable facts you share, and reuses them in later conversations. Settings describes it in one line: "Whizi remembers durable facts you share in conversations, like your name, preferences, and goals. It uses them across your chats on web and mobile."

The facts belong to your account, not to a single conversation, so they follow you into a brand new chat and into whichever model you pick. They are stored as a plain list of short bullet lines, capped at 2,000 tokens.

| Surface | What memory does there |
| --- | --- |
| Text chat | The saved list is attached to the request as its own system message before the model answers |
| Voice | The saved list is added to the voice session prompt when the session starts |
| Settings | A "What Whizi remembers" row lists the facts, and one action clears all of them |
| Data export | The export file carries the list under a memory field |
| Account deletion | Saved memory is deleted along with the rest of the account |

#### What earns a remembered fact

The test a fact has to pass is strict: it qualifies only if it would change how Whizi answers weeks from now, in an unrelated conversation. Mentioning something is not asking to be remembered, and returning the list unchanged is the normal outcome of a pass.

| Recorded | Not recorded |
| --- | --- |
| Durable facts you share about yourself: name, preferences, goals, relationships, ongoing situations | Things you merely asked about or discussed, small talk, one-off tasks, transient state, passing opinions |
| Standing preferences you explicitly ask to be remembered | A joke, an experiment, or a one-off request treated as a standing rule |
| The newest version when two facts conflict | Health or financial details, unless you explicitly asked for them to be remembered |
| Names, numbers and dates kept verbatim | Facts about Whizi itself: its name, interface, features or models, or what you believe about them |
| Facts folded into an existing line where they fit | Anything a line you already have implies |

Whizi is instructed never to store observations about the app itself, and never to store instructions about how the assistant should speak or format replies unless you explicitly asked for that as a standing preference. A stored rule of that kind would be obeyed in every future conversation as if you had just asked for it, which is exactly the failure the rule exists to prevent.

Voice gets one extra guard. A voice transcript comes from speech recognition and can contain mishearings, so a line that would read as garbled or surprising is left out rather than guessed at.

Memory is not the same thing as per-project instructions. Instructions you want applied to one body of work belong in a project, covered in [use projects](https://whizi.io/docs/use-projects).

#### When Whizi writes to memory

Not after every message. In text chat the extraction pass runs when the conversation crosses a multiple of six turns, which works out to roughly every third reply. The pass reads the last six turns plus the list you already have, and returns the rewritten list.

There is one deliberate bypass. If your message asks to remember something, using wording like "remember", "memorize", "do not forget", or asking Whizi to save or keep something in memory, a pass fires immediately rather than waiting for the next boundary. That exists because a first exchange of "my name is..." followed by "remember that" used to extract nothing at all, since a conversation's first scheduled pass only arrives at six turns.

Voice works on a different trigger: the pass runs once at the end of a session, over that session transcript.

Every pass runs in the background after your reply is already on screen, and it rewrites the whole list rather than appending to it. That is also how stored junk gets cleaned up: existing lines are held to the same bar on every pass, and a line that breaks the rules is deleted rather than kept out of caution.

The read rides alongside the rest of the request rather than in front of it, so a memory failure never blocks a reply. If the read fails, the model simply gets no memory block for that message.

#### The size ceiling

Memory is capped, which is why it stays a short list of dense lines rather than a growing transcript of everything you ever said.

| Setting | Value |
| --- | --- |
| Maximum stored memory | 2,000 tokens |
| Size a rewrite aims at | 1,500 tokens |
| Shape it aims at | About 50 lines, about 55 characters each |
| Chat turns read per pass | The last 6 |
| Transcript input cap, chat and voice | 20,000 tokens |
| Input cap on either model pass | 50,000 tokens |

When a rewrite comes back over the ceiling, a second pass shortens it: overlapping lines about the same person, thing or situation are merged, and the least useful lines are dropped. Usefulness is judged on content, never on position in the list, and long standing facts about who you are (name, family, work, standing preferences) are treated as the most valuable lines rather than the most disposable.

A positional trim survives underneath that as a last resort floor, for the case where the shortening pass fails or comes back still over budget. Memory has to come back under the ceiling because the chat system prompt is budgeted around it.

#### Seeing and clearing what Whizi remembers

On the web, open Settings. The Memory section sits at the end of Preferences, with a row labelled "What Whizi remembers" carrying a count of the stored facts. Opening the row expands the full list, with a red "Clear memory" action underneath it.

Clearing asks first. The confirmation is titled "Clear Whizi's memory?" and reads: "This permanently deletes everything Whizi remembers about you, across web and mobile. This can't be undone."

With nothing stored yet, the row is replaced by "Nothing remembered yet. Facts you share in chats will show up here."

In the app, the same thing lives in Settings under the row "What Whizi knows about me". It opens an alert titled "What Whizi remembers" with the facts as bullets, a Close button, and a destructive "Forget everything" button. A successful clear confirms with "Memory cleared" and "Whizi no longer remembers anything about you."

**There is no way to edit or delete one individual fact, and no switch that turns memory off.**

#### When memory does not load

Both memory failure strings name the memory service, not your account.

| What you see | What it means |
| --- | --- |
| "Memory is unavailable right now. Please try again later." | The Settings read failed. Your stored facts are not affected |
| "Failed to clear memory. Please try again." | The clear request failed, so the list is still there. Retry the clear |

In the app the equivalents are alerts titled "Memory unavailable" and "Memory could not be cleared". Neither one deletes anything, so a failed clear leaves your facts exactly as they were.

#### Guests, privacy and getting the data out

Memory works without an account. A guest session gets its own fact list, and when that guest creates an account the facts are merged into the account line by line, without duplicating any line, and the guest copy is deleted.

Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models, and does not sell that content as training data. The full detail is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

Memory is the one exception to picking your own model. The background passes that write and condense your fact list do not run on the model you selected for the conversation. They run on a fixed model Whizi picks for that job, and what they read is the recent turns of the conversation plus the fact list you already have. Switching your chat to a different model does not change which model performs the memory pass.

Saved memory is kept until deleted, unlike uploads and generated media which are configured to expire after up to 30 days. Where each category physically lives is set out in [where Whizi stores your data](https://whizi.io/docs/where-whizi-stores-data).

Your data export includes the list, under a memory field alongside your chats, projects, media, share links, subscription and referral profile. The account data export is covered on [the Whizi API page](https://whizi.io/docs/whizi-api). Account deletion removes saved memory along with the account itself.

Key points:

- What earns a remembered fact, and the categories Whizi refuses to store
- The 2,000 token ceiling, and what happens when memory outgrows it
- Where to see the list and clear it, on web and in the app

Checklist:

- Memory is per account, not per conversation, and is shared between web and the app
- A fact qualifies only if it would change an answer weeks later in an unrelated chat
- The extractor is instructed not to record health or financial details unless you explicitly ask
- Whizi is instructed never to store facts about the app itself or uninstructed behaviour rules, and to delete them from the list if they got in
- Text chat extracts roughly every third reply, and immediately if you ask to be remembered
- Voice extracts once, at the end of a session
- The list is capped at 2,000 tokens and is condensed first, with a positional trim as the fallback
- Settings shows the full list, and the only edit control is clearing all of it

FAQ:

Q: Does Whizi remember previous conversations?
A: It remembers durable facts from them rather than the conversations themselves. A short list of facts about you, things like your name, preferences, goals, relationships and ongoing situations, is kept on your account and attached to later requests as its own system message. Old chat text is not replayed into a new chat.

Q: How do I make Whizi forget something?
A: There is no per fact delete. Settings gives you the full list and a single Clear memory action that deletes everything at once, confirmed by a dialog and not undoable. For one wrong line, stating the correct version in a chat is usually enough, since an extraction pass keeps the newest fact when two conflict and deletes lines that no longer belong.

Q: Does memory cost credits?
A: No. Nothing in the memory code path debits credits, neither the extraction pass that updates the list nor the pass that condenses it when it overflows. Both run in the background after your reply, and neither is a message you sent. Credit costs are covered in [credits](https://whizi.io/docs/credits).

Q: Does memory carry over when I switch models?
A: Yes. The fact list is attached to the request regardless of which model you selected, so switching mid conversation keeps everything Whizi knows about you. That is the point of keeping the facts on the account rather than inside one model thread. See [switch models mid conversation](https://whizi.io/docs/switch-models-mid-conversation).

Q: Does memory work without an account?
A: Yes. The memory endpoint accepts a guest session as well as a signed-in one, and no plan or entitlement check sits in front of it, so a guest gets the same fact list a paid account gets.

Q: Is what Whizi remembers about me used to train AI models?
A: No, and the stored fact list is not an exception to that. It sits inside the same training statement as the rest of your content rather than outside it, which is set out in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy). The memory-specific part is the timing: clearing the list in Settings deletes the stored row straight away, so you do not have to delete the whole account to get rid of it.

### The Whizi mobile app: iPhone, iPad and Android

URL: https://whizi.io/docs/whizi-on-mobile/
Updated: August 2026

Quick answer: Whizi runs on the web, on iPhone and iPad, and on Android, and all of them are included in every plan including Starter. One login covers all three, and there is no separate mobile subscription to buy. Conversations, credits and quotas follow the account rather than the device, and every device reads the same model catalogue.

One subscription across web and mobile, what carries over between devices, and the differences worth knowing before you subscribe on a phone.

#### The short answer

Whizi runs on the web, on iOS and on Android, and all three are included in every plan including Starter. One login covers all of them, and there is no separate mobile subscription to buy.

The web app, the iPhone app and the Android app read the same server-side model catalogue and the same tier gate, so a model available on one is available on all three at the same plan level.

That is worth spelling out, because it is the part a store listing cannot show you. Pro reaches Claude Sonnet 5 and Powerhouse reaches Claude Opus 5, and both open on an iPhone or an Android handset exactly as they do in a browser. There is no reduced mobile catalogue and no model that is web only.

#### What is shared across devices

| Shared | Not shared |
| --- | --- |
| Your subscription and plan | The local chat cache on each device |
| Conversations and history | |
| The credit allowance | |
| Image, voice and video quotas | |
| The model catalogue and tier access | |

Usage meters are genuinely shared rather than duplicated. Voice minutes spent on the phone come out of the same 500 minute monthly allowance as voice minutes spent on the desktop, and the same is true of credits, image generations and video generations. The settings screen describes these as resetting monthly and shared across web and mobile, which is exactly what happens.

This is the part people most often get wrong when comparing multi-model apps: an allowance that is per-device is a very different product from one that is per-account, and a phone-first workflow gets no discount from either.

The one thing that is not shared is the local cache each device keeps for speed. It holds nothing the account does not already have, it is separate from the server record, and on a phone it is cleared by clearing app storage or by uninstalling the app. Conversations themselves live on the account, so a thread started on the phone is already there when you open the web app.

#### Where you subscribe decides where you cancel

This is the one operational difference between web and mobile that matters, and it catches people out.

If you subscribe on whizi.io, you cancel in account settings on the web. If you subscribe inside the iOS or Android app, Apple or Google owns that billing relationship and you must cancel through their subscription settings before the renewal deadline the store shows you. Whizi cannot cancel a store subscription on your behalf.

Deleting the app does not cancel a store subscription, and neither does deleting your Whizi account. Both of those remove your access while the store keeps billing. The full detail is in [billing and cancellation](https://whizi.io/docs/billing-and-cancellation).

#### What mobile is better and worse at

Worth being honest about the shape of the thing rather than claiming parity.

**Better on mobile.** Voice, obviously, since talking to a model while walking or cooking is the case voice was built for. Photographing something and asking about it, which beats finding and uploading a file. And catching a thought when you are away from a desk, where the transcript being in your history later is the whole value.

**Better on the web.** Side-by-side comparison, which wants screen width to be readable at all. Anything involving a long document or a spreadsheet. And any output you intend to copy into something else, which is simply easier with a keyboard and a real clipboard.

Neither list is a feature gate. The apps read the same catalogue and the same tier gate as the web app, so what differs is the shape of the device rather than what the plan opens.

The two are the same account rather than two products, so the sensible pattern is to use whichever is nearer and let the history reconcile itself.

Anything that is not device specific, from what a credit buys to what the trial includes, is answered on [the Whizi FAQ](https://whizi.io/docs/whizi-faq).

Key points:

- What is shared between the web app and the phone, and what is not
- Why where you subscribe determines where you can cancel
- The same model catalogue and the same tier gate on every device

Checklist:

- Web, iOS and Android are included on every plan
- One login covers all three
- Credits and quotas are shared across devices, not per device
- The model catalogue and tier gate are identical on every device
- Cancel where you subscribed: web in settings, app through Apple or Google
- Deleting the app does not cancel a store subscription
- Use the web for side-by-side and long documents

FAQ:

Q: Is the mobile app included in my subscription?
A: Yes, on every plan including Starter. Web, iOS and Android access all come with one subscription and one login, and there is no separate mobile purchase. Your conversations, credits and quotas follow the account rather than the device.

Q: Are my credits separate on mobile?
A: No, there is one allowance per account and it is shared across web and mobile. A message sent from the phone spends from the same monthly credits as a message sent from the desktop, and the same applies to image generations, voice minutes and video generations.

Q: Can I get Claude on my phone through Whizi?
A: Yes, on the Pro plan or above, because the mobile apps read the same model catalogue and the same tier gate as the web app. Pro reaches Claude Sonnet 5 and Powerhouse reaches Claude Opus 5, and both are available on iPhone and Android exactly as they are in a browser.

Q: How do I cancel if I subscribed in the app?
A: Through Apple or Google subscription settings, not through Whizi, because the store handles that billing. You need to cancel before the renewal deadline the store shows. Deleting the app or deleting your Whizi account will not stop the charge.

Q: Do my conversations sync between phone and computer?
A: Yes, conversations live on the account rather than on the device, so a thread started on the phone is there when you open the web app. Each device also keeps a local cache for speed, which is separate from the server record and can be cleared on mobile by clearing app storage or uninstalling.

### Write and debug code with multiple models in Whizi

URL: https://whizi.io/docs/write-and-debug-code-with-multiple-models/
Updated: August 2026

Quick answer: The coding specialists in Whizi sit on Powerhouse. Pro reaches no coder specialist at all, so its code path is the Claude Sonnet class, Kimi K3, and the curated general rows. Starter reaches four picker entries in total and none is a coding model. A free account can open any text model within 7 lifetime messages.

The coder-named rows in the Whizi catalogue, the plan each one sits behind, what each costs per message, and how a code chat actually works once the file is in it.

#### The short answer

The coding specialists are one of the categories held on Powerhouse. Pro reaches no coder specialist at all, so its code path is the Claude Sonnet class, [Kimi K3](https://whizi.io/docs/using-kimi-in-whizi), and the curated general rows. Starter reaches four picker entries in total and none of them is a coding model. A free account is not model-gated, so it can open any text model in the catalogue, coding specialists included, inside a lifetime allowance of 7 messages that never resets.

| Plan | What it reaches for code |
| --- | --- |
| Free | Any text model in the catalogue, within a lifetime allowance of 7 messages |
| Starter, $15.99/month | Four picker entries in total: Auto, the house model, one GPT row and one Gemini Flash row. No coding specialist |
| Pro, $29.99/month | The Claude Sonnet class, Kimi K3, and the curated general rows across GPT, Gemini, DeepSeek, [Llama](https://whizi.io/docs/using-llama-in-whizi), [Mistral](https://whizi.io/docs/using-mistral-in-whizi), Qwen and [GLM](https://whizi.io/docs/using-glm-in-whizi). Still no coding specialist |
| Powerhouse, $49.99/month | Everything above, plus the coding specialists |

The rest of this page is the coder-named rows with their costs, how to get code into a chat and what refuses it, and which existing page owns each half of the compare and switch loop.

#### Coder-named rows in the catalogue, and what each one costs

Anything not in the Starter set, the agent set, or the explicit Pro list resolves to Powerhouse, and the coding specialists are one of the categories deliberately held back there so the top tier has something to sell. They sit alongside the frontier reasoning tier, the 400B class open weights and the routing models.

Identifiers below are reference data, quoted as the catalogue carries them.

| Model | Credits per message | Lowest plan |
| --- | --- | --- |
| `openai gpt-5.3-codex` | 12 | Powerhouse |
| `openai gpt-5.2-codex` | 10 | Powerhouse |
| `moonshotai kimi-k2.7-code` | 10 | Powerhouse |
| `openai gpt-5.1-codex` | 8 | Powerhouse |
| `openai gpt-5.1-codex-max` | 8 | Powerhouse |
| `qwen qwen3-coder` | 3 | Powerhouse |
| `x-ai grok-build-0.1` | 3 | Powerhouse |
| Qwen3 Coder Next | 1 | Powerhouse |
| Codestral 2508 | 1 | Powerhouse |

The Cost Index prices several of these rows, quoting what one standard answer of 1,000 input plus 500 output tokens costs at provider list rates, read from OpenRouter on 2026-08-20.

| Model | Context window | Cost per standard answer | Credits |
| --- | --- | --- | --- |
| Qwen3 Coder Next | 262K | $0.00052 | 1 |
| Codestral 2508 | 256K | $0.00075 | 1 |
| Grok Build 0.1 | 256K | $0.002 | 3 |
| Kimi K2.7 Code | 262K | $0.00246 | 10 |

Of the priced rows above, two sit at 1 credit and one at 10, all behind the same plan.

A credit cost is a fixed integer per message. Nothing about the length of your code, the length of the thread, or the length of the answer enters the calculation, so a one line question and a 400 line file cost the same on the same model. The full ladder is in [how credits work](https://whizi.io/docs/credits).

#### What Pro uses instead

Pro is curated rather than complete: one or two current flagships per family, the fast tier from each family, and the high-volume workhorses. Sonnet is the Pro ceiling on Anthropic and the Opus class is Powerhouse, so for code on Pro the strongest hand-picked rows are the Claude Sonnet class at 10 credits and Kimi K3 at 10 credits.

Auto agrees with that. Auto sorts each message onto one of six fixed rungs, and its code rung is pinned to the Claude Sonnet class at 10 credits, with Pro as the lowest plan that reaches it. On Pro, the row Auto picks for a code message and the best code row you can pick by hand are the same row. The whole ladder is in [how Auto routing works](https://whizi.io/docs/how-auto-routing-works).

Difficulty is also sticky inside a thread: the model on the previous turn acts as a floor, so a follow-up to a code answer cannot drop back to the cheapest row on its own.

#### Getting code into the chat

Two routes accept code, and they accept different things.

**Per message.** The composer file picker takes source files directly, and the complete extension list is in [supported file types](https://whizi.io/docs/supported-file-types). The website composer stages at most 6 files per message, each capped at 10 MB.

**Pinned to a project.** This list is narrower and does not include source code: pinned files accept PDF, plain text, Markdown, CSV, JSON, .docx and .xlsx, and images are excluded on purpose because a pinned file rides along as prompt text on every turn. Picking a .ts or .py file returns `Choose a PDF, Word, Excel, text, Markdown, CSV, or JSON file.` as HTTP 415 `unsupported_project_file_type`. Saving the file as .txt or .md first is the way through. The other upload refusals are in [file upload failed](https://whizi.io/docs/file-upload-failed).

Pasting instead of attaching runs into the message cap. A single chat message is capped at 100,000 characters, and over it the send is refused with `This message is {count} characters, over the 100,000 character limit. Attach a shorter file, or ask about one section at a time.` as HTTP 400 `message_too_long`. That count includes text extracted from your files. See [a message will not send](https://whizi.io/docs/message-not-sending).

Whatever fits, most catalogue models then get a flat per turn budget of 40,000 input tokens, and a model whose context window sits under 93,000 tokens gets a smaller proportional budget instead, so a large file and a long thread compete for the same room. [When the conversation is too long](https://whizi.io/docs/context-too-long) has the budget in full.

#### The loop, and where each half is documented

Two techniques do the actual work, and each has its own page.

- **Sequential.** [Switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).
- **Parallel.** [Comparing models side by side](https://whizi.io/docs/compare-models-side-by-side).

One attachment rule is worth carrying into either one: only the most recent user message carrying attachments has its attachments forwarded to the provider, so a file attached five turns ago is not re-sent on the current turn.

Three constraints shape how that loop feels on code specifically.

**There is no regenerate control.** The way to re-run a turn is to edit your message and resend it, which creates a version of that turn, navigable with arrows that read like "2/2". A version is not just the edited message: it is that message and every reply after it, so switching versions swaps the whole tail of the conversation, and editing a turn discards the version history of every turn after it. Correcting a prompt three turns back throws away the branches you explored downstream of it.

**Only one generation runs per account at a time.** A per-user lock means two devices, or two tabs, can never generate concurrently for one account, and a second attempt is refused as a generation in progress conflict. A reply does survive the client, though: generation runs in a per-conversation coordinator, so closing the tab mid-stream no longer kills the answer, and a reopened tab can reattach to a stream already in flight.

**Picking a model above your plan is a refusal, not a downgrade.** It returns `Upgrade your plan to use this model.` as HTTP 403 `tier_upgrade_required`. Auto never produces that error, because it only ever picks rows your plan can already open. [Model unavailable](https://whizi.io/docs/model-unavailable) has the rest of the gate messages.

#### What Whizi does not have for code

There is no public or developer API, no API key issuance and no personal access tokens, so Whizi cannot be wired into an editor, a terminal or a CI job. Every route authenticates as an end user session rather than as a third-party integration. [Does Whizi have an API](https://whizi.io/docs/whizi-api) covers what to do instead.

There is no account-level custom instructions setting either. Per-project instructions, up to 32,000 characters and applied to every conversation in that project, are the only user-authored standing instruction in the product, and they are where a stack description or a house convention belongs.

There is no prompt template or saved prompt feature: no template library, no saved prompts and no snippet insertion anywhere in the chat UI. The closest real equivalent is [a project's instructions](https://whizi.io/docs/use-projects), which apply to every chat in that project automatically.

Key points:

- The coder-named rows in the catalogue, with credit cost and the plan that reaches each one
- Why a Pro subscriber has no coder specialist, and what Pro uses instead
- The three constraints that shape a code chat: no regenerate, one generation at a time, and a capped input budget

Checklist:

- The coding specialists are a Powerhouse-only category; Starter and Pro do not reach them
- Pro has no coder specialist; its code rows are the Claude Sonnet class and Kimi K3 at 10 credits
- Starter reaches four picker entries in total and none is a coding model
- A free account can try any text model, coding rows included, within 7 lifetime messages
- Two coder-named rows cost 1 credit and one costs 10, all behind Powerhouse
- Attach source files per message; project pinned files do not accept code extensions
- One chat message is capped at 100,000 characters, and most models get a 40,000 token input budget
- There is no regenerate control: edit the message and resend, which versions the whole tail

FAQ:

Q: Which model is best for coding in Whizi?
A: What Whizi publishes for each row is the plan gate, the credit rung and the context window, not a ranking. The nearest thing to a product opinion is Auto, whose code rung is pinned to Claude Sonnet at 10 credits and is reachable from Pro. Beyond that, the practical method is to run the same task against two rows and keep the one that wins on your own code.

Q: Can I use the Codex models on Pro?
A: No. The Pro OpenAI set is the general GPT rows, and every Codex identifier falls outside both the Starter set and the Pro set, which means Powerhouse. The Codex rows sit between 8 and 12 credits per message, so some are cheaper per turn than the Claude Sonnet class at 10 and the dearest is 12.

Q: Can I pin a source file to a project so every chat sees it?
A: Not with its original extension, and the caps matter more than the extension does. A pinned file is rebuilt into the prompt on every single turn in that project, inside the same input budget as the conversation, with each pinned file contributing at most 32,000 characters of extracted text and the whole project block capped at 120,000 characters across at most 10 files.

Q: Is my code used to train AI models?
A: No. Whizi does not use your prompts, files, conversations, voice transcripts, or generated content to train Whizi-owned AI models. Whizi does not sell that content as training data. Uploaded attachments are configured to expire after up to 30 days and may be deleted sooner by you or during account deletion. Processing by the upstream model providers is governed by the terms and configuration of each provider, and Whizi does not claim that every provider follows the same retention setting. The detail is in [what Whizi does with your data](https://whizi.io/docs/data-and-privacy).

## Model reference: price, context and credit cost

### Aion-3.0: price, context window and credit cost

URL: https://whizi.io/models/aion-3-0/
Updated: August 2026

Quick answer: Aion-3.0 costs $0.0060 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $3 per million input tokens and $6 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 10 credits per message on the Powerhouse plan and above.

What Aion-3.0 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Aion-3.0 is an Aion Labs model priced at $3 per million input tokens and $6 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0060. Inside Whizi it costs 10 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 15 of the 89 priced models cost more per answer, and 73 cost less.

#### What Aion-3.0 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $3 |
| Output, per million tokens | $6 |
| One standard answer | $0.0060 |
| One thousand answers | $6.00 |
| Context window | 131K tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 74 |

Prices are the published Aion Labs rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Aion-3.0 is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Aion-3.0 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Aion-3.0 costs 113 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 3.5 times more.

Its nearest neighbours from Aion Labs:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Aion-3.0-Mini | $0.0014 | 131K | 2 |
| Aion-3.0 (this page) | $0.0060 | 131K | 10 |

If Aion-3.0 is more model than a given task needs, Aion-3.0-Mini is the cheaper Aion Labs option at $0.0014 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Aion-3.0 accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0060 for one standard answer, ranking 74 of 89 priced models
- 131K token context window, below the catalogue median
- 10 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $3 and output $6 per million tokens
- $0.0060 for one standard answer, $6.00 for a thousand
- 131K token context window
- 10 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 74 of 89 priced models by cost per answer

FAQ:

Q: How much does Aion-3.0 cost?
A: $3 per million input tokens and $6 per million output tokens at list rates, which works out to $0.0060 for one standard answer of 1,000 tokens in and 500 out, or $6.00 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Aion-3.0?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 10 credits per message, a Powerhouse allowance of 8,000 credits covers 800 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Aion-3.0?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Aion-3.0 expensive compared to other models?
A: It ranks 74 of 89 priced models by cost per answer, so 15 cost more and 73 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Aion-3.0-Mini: price, context window and credit cost

URL: https://whizi.io/models/aion-3-0-mini/
Updated: August 2026

Quick answer: Aion-3.0-Mini costs $0.0014 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.7 per million input tokens and $1.4 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What Aion-3.0-Mini costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Aion-3.0-Mini is an Aion Labs model priced at $0.7 per million input tokens and $1.4 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0014. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 49 of the 89 priced models cost more per answer, and 39 cost less.

#### What Aion-3.0-Mini costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.7 |
| Output, per million tokens | $1.4 |
| One standard answer | $0.0014 |
| One thousand answers | $1.40 |
| Context window | 131K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 40 |

Prices are the published Aion Labs rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Aion-3.0-Mini is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Aion-3.0-Mini is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Aion-3.0-Mini costs 26 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.2 times less.

Its nearest neighbours from Aion Labs:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Aion-3.0-Mini (this page) | $0.0014 | 131K | 2 |
| Aion-3.0 | $0.0060 | 131K | 10 |

It is the cheapest Aion Labs model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Aion-3.0-Mini accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0014 for one standard answer, ranking 40 of 89 priced models
- 131K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.7 and output $1.4 per million tokens
- $0.0014 for one standard answer, $1.40 for a thousand
- 131K token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 40 of 89 priced models by cost per answer

FAQ:

Q: How much does Aion-3.0-Mini cost?
A: $0.7 per million input tokens and $1.4 per million output tokens at list rates, which works out to $0.0014 for one standard answer of 1,000 tokens in and 500 out, or $1.40 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Aion-3.0-Mini?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Aion-3.0-Mini?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Aion-3.0-Mini expensive compared to other models?
A: It ranks 40 of 89 priced models by cost per answer, so 49 cost more and 39 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Fable 5: price, context window and credit cost

URL: https://whizi.io/models/claude-fable-5/
Updated: August 2026

Quick answer: Claude Fable 5 costs $0.0350 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $10 per million input tokens and $50 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 50 credits per message on the Powerhouse plan and above.

What Claude Fable 5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Fable 5 is an Anthropic model priced at $10 per million input tokens and $50 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0350. Inside Whizi it costs 50 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 0 of the 89 priced models cost more per answer, and 88 cost less.

#### What Claude Fable 5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $10 |
| Output, per million tokens | $50 |
| One standard answer | $0.0350 |
| One thousand answers | $35.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 50 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 89 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Fable 5 is charged at 50 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 160 |

Claude Fable 5 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Fable 5 costs 660 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 20 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Sonnet 5 | $0.0070 | 1M | 10 |
| Claude Sonnet 4.6 | $0.0105 | 1M | 10 |
| Claude Opus 5 | $0.0175 | 1M | 20 |
| Claude Fable 5 (this page) | $0.0350 | 1M | 50 |

If Claude Fable 5 is more model than a given task needs, Claude Opus 5 is the cheaper Anthropic option at $0.0175 per answer and 20 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Claude Fable 5 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0350 for one standard answer, ranking 89 of 89 priced models
- 1M token context window, well above the catalogue median
- 50 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $10 and output $50 per million tokens
- $0.0350 for one standard answer, $35.00 for a thousand
- 1M token context window
- 50 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 89 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Fable 5 cost?
A: $10 per million input tokens and $50 per million output tokens at list rates, which works out to $0.0350 for one standard answer of 1,000 tokens in and 500 out, or $35.00 for a thousand of them. Inside Whizi it is 50 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Fable 5?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 50 credits per message, a Powerhouse allowance of 8,000 credits covers 160 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Fable 5?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Fable 5 expensive compared to other models?
A: It ranks 89 of 89 priced models by cost per answer, so 0 cost more and 88 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Haiku 4.5: price, context window and credit cost

URL: https://whizi.io/models/claude-haiku-4-5/
Updated: August 2026

Quick answer: Claude Haiku 4.5 costs $0.0035 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1 per million input tokens and $5 per million output tokens. The context window is 200K tokens, below the catalogue median. Inside Whizi it costs 4 credits per message on the Pro plan and above.

What Claude Haiku 4.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Haiku 4.5 is an Anthropic model priced at $1 per million input tokens and $5 per million output tokens, with a 200K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0035. Inside Whizi it costs 4 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 24 of the 89 priced models cost more per answer, and 64 cost less.

#### What Claude Haiku 4.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1 |
| Output, per million tokens | $5 |
| One standard answer | $0.0035 |
| One thousand answers | $3.50 |
| Context window | 200K tokens |
| Credits per message in Whizi | 4 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 65 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Haiku 4.5 is charged at 4 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 500 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,000 |

Claude Haiku 4.5 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Haiku 4.5 costs 66 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.0 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Haiku 4.5 (this page) | $0.0035 | 200K | 4 |
| Claude Sonnet 5 | $0.0070 | 1M | 10 |
| Claude Sonnet 4.6 | $0.0105 | 1M | 10 |
| Claude Opus 5 | $0.0175 | 1M | 20 |

It is the cheapest Anthropic model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Claude Haiku 4.5 accepts 200K tokens of context, which is roughly 300 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0035 for one standard answer, ranking 65 of 89 priced models
- 200K token context window, below the catalogue median
- 4 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $1 and output $5 per million tokens
- $0.0035 for one standard answer, $3.50 for a thousand
- 200K token context window
- 4 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 65 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Haiku 4.5 cost?
A: $1 per million input tokens and $5 per million output tokens at list rates, which works out to $0.0035 for one standard answer of 1,000 tokens in and 500 out, or $3.50 for a thousand of them. Inside Whizi it is 4 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Haiku 4.5?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 4 credits per message, a Pro allowance of 2,000 credits covers 500 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Haiku 4.5?
A: 200K tokens, which is roughly 300 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Haiku 4.5 expensive compared to other models?
A: It ranks 65 of 89 priced models by cost per answer, so 24 cost more and 64 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Opus 4.8: price, context window and credit cost

URL: https://whizi.io/models/claude-opus-4-8/
Updated: August 2026

Quick answer: Claude Opus 4.8 costs $0.0175 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $5 per million input tokens and $25 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 20 credits per message on the Powerhouse plan and above.

What Claude Opus 4.8 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Opus 4.8 is an Anthropic model priced at $5 per million input tokens and $25 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0175. Inside Whizi it costs 20 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 4 of the 89 priced models cost more per answer, and 84 cost less.

#### What Claude Opus 4.8 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $5 |
| Output, per million tokens | $25 |
| One standard answer | $0.0175 |
| One thousand answers | $17.50 |
| Context window | 1M tokens |
| Credits per message in Whizi | 20 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 85 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Opus 4.8 is charged at 20 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 400 |

Claude Opus 4.8 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Opus 4.8 costs 330 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 10 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Haiku 4.5 | $0.0035 | 200K | 4 |
| Claude Sonnet 5 | $0.0070 | 1M | 10 |
| Claude Sonnet 4.6 | $0.0105 | 1M | 10 |
| Claude Opus 4.8 (this page) | $0.0175 | 1M | 20 |

If Claude Opus 4.8 is more model than a given task needs, Claude Sonnet 4.6 is the cheaper Anthropic option at $0.0105 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Claude Opus 4.8 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0175 for one standard answer, ranking 85 of 89 priced models
- 1M token context window, well above the catalogue median
- 20 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $5 and output $25 per million tokens
- $0.0175 for one standard answer, $17.50 for a thousand
- 1M token context window
- 20 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 85 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Opus 4.8 cost?
A: $5 per million input tokens and $25 per million output tokens at list rates, which works out to $0.0175 for one standard answer of 1,000 tokens in and 500 out, or $17.50 for a thousand of them. Inside Whizi it is 20 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Opus 4.8?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 20 credits per message, a Powerhouse allowance of 8,000 credits covers 400 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Opus 4.8?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Opus 4.8 expensive compared to other models?
A: It ranks 85 of 89 priced models by cost per answer, so 4 cost more and 84 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Opus 5: price, context window and credit cost

URL: https://whizi.io/models/claude-opus-5/
Updated: August 2026

Quick answer: Claude Opus 5 costs $0.0175 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $5 per million input tokens and $25 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 20 credits per message on the Powerhouse plan and above.

What Claude Opus 5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Opus 5 is an Anthropic model priced at $5 per million input tokens and $25 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0175. Inside Whizi it costs 20 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 5 of the 89 priced models cost more per answer, and 83 cost less.

#### What Claude Opus 5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $5 |
| Output, per million tokens | $25 |
| One standard answer | $0.0175 |
| One thousand answers | $17.50 |
| Context window | 1M tokens |
| Credits per message in Whizi | 20 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 84 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Opus 5 is charged at 20 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 400 |

Claude Opus 5 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Opus 5 costs 330 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 10 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Haiku 4.5 | $0.0035 | 200K | 4 |
| Claude Sonnet 5 | $0.0070 | 1M | 10 |
| Claude Sonnet 4.6 | $0.0105 | 1M | 10 |
| Claude Opus 5 (this page) | $0.0175 | 1M | 20 |

If Claude Opus 5 is more model than a given task needs, Claude Sonnet 4.6 is the cheaper Anthropic option at $0.0105 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Claude Opus 5 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0175 for one standard answer, ranking 84 of 89 priced models
- 1M token context window, well above the catalogue median
- 20 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $5 and output $25 per million tokens
- $0.0175 for one standard answer, $17.50 for a thousand
- 1M token context window
- 20 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 84 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Opus 5 cost?
A: $5 per million input tokens and $25 per million output tokens at list rates, which works out to $0.0175 for one standard answer of 1,000 tokens in and 500 out, or $17.50 for a thousand of them. Inside Whizi it is 20 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Opus 5?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 20 credits per message, a Powerhouse allowance of 8,000 credits covers 400 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Opus 5?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Opus 5 expensive compared to other models?
A: It ranks 84 of 89 priced models by cost per answer, so 5 cost more and 83 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Sonnet 4.5: price, context window and credit cost

URL: https://whizi.io/models/claude-sonnet-4-5/
Updated: August 2026

Quick answer: Claude Sonnet 4.5 costs $0.0105 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $3 per million input tokens and $15 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 10 credits per message on the Pro plan and above.

What Claude Sonnet 4.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Sonnet 4.5 is an Anthropic model priced at $3 per million input tokens and $15 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0105. Inside Whizi it costs 10 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 7 of the 89 priced models cost more per answer, and 81 cost less.

#### What Claude Sonnet 4.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $3 |
| Output, per million tokens | $15 |
| One standard answer | $0.0105 |
| One thousand answers | $10.50 |
| Context window | 1M tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 82 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Sonnet 4.5 is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 200 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Claude Sonnet 4.5 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Sonnet 4.5 costs 198 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 6.1 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Haiku 4.5 | $0.0035 | 200K | 4 |
| Claude Sonnet 5 | $0.0070 | 1M | 10 |
| Claude Sonnet 4.5 (this page) | $0.0105 | 1M | 10 |
| Claude Opus 5 | $0.0175 | 1M | 20 |

If Claude Sonnet 4.5 is more model than a given task needs, Claude Sonnet 5 is the cheaper Anthropic option at $0.0070 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Claude Sonnet 4.5 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0105 for one standard answer, ranking 82 of 89 priced models
- 1M token context window, well above the catalogue median
- 10 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $3 and output $15 per million tokens
- $0.0105 for one standard answer, $10.50 for a thousand
- 1M token context window
- 10 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 82 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Sonnet 4.5 cost?
A: $3 per million input tokens and $15 per million output tokens at list rates, which works out to $0.0105 for one standard answer of 1,000 tokens in and 500 out, or $10.50 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Sonnet 4.5?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 10 credits per message, a Pro allowance of 2,000 credits covers 200 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Sonnet 4.5?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Sonnet 4.5 expensive compared to other models?
A: It ranks 82 of 89 priced models by cost per answer, so 7 cost more and 81 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Sonnet 4.6: price, context window and credit cost

URL: https://whizi.io/models/claude-sonnet-4-6/
Updated: August 2026

Quick answer: Claude Sonnet 4.6 costs $0.0105 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $3 per million input tokens and $15 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 10 credits per message on the Pro plan and above.

What Claude Sonnet 4.6 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Sonnet 4.6 is an Anthropic model priced at $3 per million input tokens and $15 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0105. Inside Whizi it costs 10 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 8 of the 89 priced models cost more per answer, and 80 cost less.

#### What Claude Sonnet 4.6 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $3 |
| Output, per million tokens | $15 |
| One standard answer | $0.0105 |
| One thousand answers | $10.50 |
| Context window | 1M tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 81 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Sonnet 4.6 is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 200 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Claude Sonnet 4.6 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Sonnet 4.6 costs 198 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 6.1 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Haiku 4.5 | $0.0035 | 200K | 4 |
| Claude Sonnet 5 | $0.0070 | 1M | 10 |
| Claude Sonnet 4.6 (this page) | $0.0105 | 1M | 10 |
| Claude Opus 5 | $0.0175 | 1M | 20 |

If Claude Sonnet 4.6 is more model than a given task needs, Claude Sonnet 5 is the cheaper Anthropic option at $0.0070 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Claude Sonnet 4.6 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0105 for one standard answer, ranking 81 of 89 priced models
- 1M token context window, well above the catalogue median
- 10 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $3 and output $15 per million tokens
- $0.0105 for one standard answer, $10.50 for a thousand
- 1M token context window
- 10 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 81 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Sonnet 4.6 cost?
A: $3 per million input tokens and $15 per million output tokens at list rates, which works out to $0.0105 for one standard answer of 1,000 tokens in and 500 out, or $10.50 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Sonnet 4.6?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 10 credits per message, a Pro allowance of 2,000 credits covers 200 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Sonnet 4.6?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Sonnet 4.6 expensive compared to other models?
A: It ranks 81 of 89 priced models by cost per answer, so 8 cost more and 80 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Claude Sonnet 5: price, context window and credit cost

URL: https://whizi.io/models/claude-sonnet-5/
Updated: August 2026

Quick answer: Claude Sonnet 5 costs $0.0070 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $10 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 10 credits per message on the Pro plan and above.

What Claude Sonnet 5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Claude Sonnet 5 is an Anthropic model priced at $2 per million input tokens and $10 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0070. Inside Whizi it costs 10 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 14 of the 89 priced models cost more per answer, and 74 cost less.

#### What Claude Sonnet 5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $10 |
| One standard answer | $0.0070 |
| One thousand answers | $7.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 75 |

Prices are the published Anthropic rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Claude Sonnet 5 is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 200 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Claude Sonnet 5 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Claude Sonnet 5 costs 132 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 4.1 times more.

Its nearest neighbours from Anthropic:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Claude Haiku 4.5 | $0.0035 | 200K | 4 |
| Claude Sonnet 5 (this page) | $0.0070 | 1M | 10 |
| Claude Sonnet 4.6 | $0.0105 | 1M | 10 |
| Claude Opus 5 | $0.0175 | 1M | 20 |

If Claude Sonnet 5 is more model than a given task needs, Claude Haiku 4.5 is the cheaper Anthropic option at $0.0035 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Claude Sonnet 5 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0070 for one standard answer, ranking 75 of 89 priced models
- 1M token context window, well above the catalogue median
- 10 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $2 and output $10 per million tokens
- $0.0070 for one standard answer, $7.00 for a thousand
- 1M token context window
- 10 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 75 of 89 priced models by cost per answer

FAQ:

Q: How much does Claude Sonnet 5 cost?
A: $2 per million input tokens and $10 per million output tokens at list rates, which works out to $0.0070 for one standard answer of 1,000 tokens in and 500 out, or $7.00 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Claude Sonnet 5?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 10 credits per message, a Pro allowance of 2,000 credits covers 200 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Claude Sonnet 5?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Claude Sonnet 5 expensive compared to other models?
A: It ranks 75 of 89 priced models by cost per answer, so 14 cost more and 74 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Codestral 2508: price, context window and credit cost

URL: https://whizi.io/models/codestral-2508/
Updated: August 2026

Quick answer: Codestral 2508 costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.3 per million input tokens and $0.9 per million output tokens. The context window is 256K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Codestral 2508 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Codestral 2508 is a Mistral model priced at $0.3 per million input tokens and $0.9 per million output tokens, with a 256K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 68 of the 89 priced models cost more per answer, and 20 cost less.

#### What Codestral 2508 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.3 |
| Output, per million tokens | $0.9 |
| One standard answer | $0.0008 |
| One thousand answers | $0.7500 |
| Context window | 256K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 21 |

Prices are the published Mistral rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Codestral 2508 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Codestral 2508 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Codestral 2508 costs 14 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.3 times less.

Its nearest neighbours from Mistral:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Codestral 2508 (this page) | $0.0008 | 256K | 1 |
| Mistral Large 3 2512 | $0.0013 | 262K | 2 |
| Mistral Medium 3.1 | $0.0014 | 131K | 2 |
| Mistral Medium 3.5 | $0.0053 | 262K | 6 |

It is the cheapest Mistral model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Codestral 2508 accepts 256K tokens of context, which is roughly 380 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 21 of 89 priced models
- 256K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.3 and output $0.9 per million tokens
- $0.0008 for one standard answer, $0.7500 for a thousand
- 256K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 21 of 89 priced models by cost per answer

FAQ:

Q: How much does Codestral 2508 cost?
A: $0.3 per million input tokens and $0.9 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.7500 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Codestral 2508?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Codestral 2508?
A: 256K tokens, which is roughly 380 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Codestral 2508 expensive compared to other models?
A: It ranks 21 of 89 priced models by cost per answer, so 68 cost more and 20 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Command A: price, context window and credit cost

URL: https://whizi.io/models/command-a/
Updated: August 2026

Quick answer: Command A costs $0.0075 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2.5 per million input tokens and $10 per million output tokens. The context window is 256K tokens, below the catalogue median. Inside Whizi it costs 10 credits per message on the Powerhouse plan and above.

What Command A costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Command A is a Cohere model priced at $2.5 per million input tokens and $10 per million output tokens, with a 256K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0075. Inside Whizi it costs 10 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 13 of the 89 priced models cost more per answer, and 75 cost less.

#### What Command A costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2.5 |
| Output, per million tokens | $10 |
| One standard answer | $0.0075 |
| One thousand answers | $7.50 |
| Context window | 256K tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 76 |

Prices are the published Cohere rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Command A is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Command A is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Command A costs 142 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 4.4 times more.

It is the only Cohere model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Cohere model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Command A accepts 256K tokens of context, which is roughly 380 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0075 for one standard answer, ranking 76 of 89 priced models
- 256K token context window, below the catalogue median
- 10 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $2.5 and output $10 per million tokens
- $0.0075 for one standard answer, $7.50 for a thousand
- 256K token context window
- 10 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 76 of 89 priced models by cost per answer

FAQ:

Q: How much does Command A cost?
A: $2.5 per million input tokens and $10 per million output tokens at list rates, which works out to $0.0075 for one standard answer of 1,000 tokens in and 500 out, or $7.50 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Command A?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 10 credits per message, a Powerhouse allowance of 8,000 credits covers 800 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Command A?
A: 256K tokens, which is roughly 380 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Command A expensive compared to other models?
A: It ranks 76 of 89 priced models by cost per answer, so 13 cost more and 75 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### DeepSeek V3.1: price, context window and credit cost

URL: https://whizi.io/models/deepseek-chat-v3-1/
Updated: August 2026

Quick answer: DeepSeek V3.1 costs $0.0007 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.25 per million input tokens and $0.95 per million output tokens. The context window is 164K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Pro plan and above.

What DeepSeek V3.1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

DeepSeek V3.1 is a DeepSeek model priced at $0.25 per million input tokens and $0.95 per million output tokens, with a 164K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0007. Inside Whizi it costs 1 credit per message and needs the Pro plan or above.

That makes it one of the cheaper models in the catalogue: 69 of the 89 priced models cost more per answer, and 19 cost less.

#### What DeepSeek V3.1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.25 |
| Output, per million tokens | $0.95 |
| One standard answer | $0.0007 |
| One thousand answers | $0.7250 |
| Context window | 164K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 20 |

Prices are the published DeepSeek rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and DeepSeek V3.1 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

DeepSeek V3.1 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, DeepSeek V3.1 costs 14 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.4 times less.

Its nearest neighbours from DeepSeek:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 | $0.0003 | 1M | 1 |
| DeepSeek V3.2 | $0.0005 | 164K | 1 |
| DeepSeek V3.1 (this page) | $0.0007 | 164K | 1 |
| R1 | $0.0019 | 64K | 2 |

If DeepSeek V3.1 is more model than a given task needs, DeepSeek V3.2 is the cheaper DeepSeek option at $0.0005 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

DeepSeek V3.1 accepts 164K tokens of context, which is roughly 250 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0007 for one standard answer, ranking 20 of 89 priced models
- 164K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.25 and output $0.95 per million tokens
- $0.0007 for one standard answer, $0.7250 for a thousand
- 164K token context window
- 1 credit per message inside Whizi
- Requires the Pro plan or above
- Ranks 20 of 89 priced models by cost per answer

FAQ:

Q: How much does DeepSeek V3.1 cost?
A: $0.25 per million input tokens and $0.95 per million output tokens at list rates, which works out to $0.0007 for one standard answer of 1,000 tokens in and 500 out, or $0.7250 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes DeepSeek V3.1?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 1 credits per message, a Pro allowance of 2,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of DeepSeek V3.1?
A: 164K tokens, which is roughly 250 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is DeepSeek V3.1 expensive compared to other models?
A: It ranks 20 of 89 priced models by cost per answer, so 69 cost more and 19 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### R1: price, context window and credit cost

URL: https://whizi.io/models/deepseek-r1/
Updated: August 2026

Quick answer: R1 costs $0.0019 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.7 per million input tokens and $2.5 per million output tokens. The context window is 64K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What R1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

R1 is a DeepSeek model priced at $0.7 per million input tokens and $2.5 per million output tokens, with a 64K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0019. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 41 of the 89 priced models cost more per answer, and 47 cost less.

#### What R1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.7 |
| Output, per million tokens | $2.5 |
| One standard answer | $0.0019 |
| One thousand answers | $1.95 |
| Context window | 64K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 48 |

Prices are the published DeepSeek rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and R1 is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

R1 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, R1 costs 37 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from DeepSeek:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| DeepSeek V3.2 | $0.0005 | 164K | 1 |
| DeepSeek V3.1 | $0.0007 | 164K | 1 |
| R1 (this page) | $0.0019 | 64K | 2 |
| DeepSeek V4 Pro 0813 | $0.0030 | 1M | 2 |

If R1 is more model than a given task needs, DeepSeek V3.1 is the cheaper DeepSeek option at $0.0007 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

R1 accepts 64K tokens of context, which is roughly 96 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0019 for one standard answer, ranking 48 of 89 priced models
- 64K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.7 and output $2.5 per million tokens
- $0.0019 for one standard answer, $1.95 for a thousand
- 64K token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 48 of 89 priced models by cost per answer

FAQ:

Q: How much does R1 cost?
A: $0.7 per million input tokens and $2.5 per million output tokens at list rates, which works out to $0.0019 for one standard answer of 1,000 tokens in and 500 out, or $1.95 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes R1?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of R1?
A: 64K tokens, which is roughly 96 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is R1 expensive compared to other models?
A: It ranks 48 of 89 priced models by cost per answer, so 41 cost more and 47 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### DeepSeek V3.2: price, context window and credit cost

URL: https://whizi.io/models/deepseek-v3-2/
Updated: August 2026

Quick answer: DeepSeek V3.2 costs $0.0005 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.269 per million input tokens and $0.4 per million output tokens. The context window is 164K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Pro plan and above.

What DeepSeek V3.2 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

DeepSeek V3.2 is a DeepSeek model priced at $0.269 per million input tokens and $0.4 per million output tokens, with a 164K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0005. Inside Whizi it costs 1 credit per message and needs the Pro plan or above.

That makes it one of the cheaper models in the catalogue: 73 of the 89 priced models cost more per answer, and 15 cost less.

#### What DeepSeek V3.2 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.269 |
| Output, per million tokens | $0.4 |
| One standard answer | $0.0005 |
| One thousand answers | $0.4690 |
| Context window | 164K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 16 |

Prices are the published DeepSeek rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and DeepSeek V3.2 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

DeepSeek V3.2 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, DeepSeek V3.2 costs 8.8 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 3.7 times less.

Its nearest neighbours from DeepSeek:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 | $0.0003 | 1M | 1 |
| DeepSeek V3.2 (this page) | $0.0005 | 164K | 1 |
| DeepSeek V3.1 | $0.0007 | 164K | 1 |
| R1 | $0.0019 | 64K | 2 |

If DeepSeek V3.2 is more model than a given task needs, DeepSeek V4 Flash 0731 is the cheaper DeepSeek option at $0.0003 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

DeepSeek V3.2 accepts 164K tokens of context, which is roughly 250 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0005 for one standard answer, ranking 16 of 89 priced models
- 164K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.269 and output $0.4 per million tokens
- $0.0005 for one standard answer, $0.4690 for a thousand
- 164K token context window
- 1 credit per message inside Whizi
- Requires the Pro plan or above
- Ranks 16 of 89 priced models by cost per answer

FAQ:

Q: How much does DeepSeek V3.2 cost?
A: $0.269 per million input tokens and $0.4 per million output tokens at list rates, which works out to $0.0005 for one standard answer of 1,000 tokens in and 500 out, or $0.4690 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes DeepSeek V3.2?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 1 credits per message, a Pro allowance of 2,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of DeepSeek V3.2?
A: 164K tokens, which is roughly 250 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is DeepSeek V3.2 expensive compared to other models?
A: It ranks 16 of 89 priced models by cost per answer, so 73 cost more and 15 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### DeepSeek V4 Flash 0731: price, context window and credit cost

URL: https://whizi.io/models/deepseek-v4-flash-0731/
Updated: August 2026

Quick answer: DeepSeek V4 Flash 0731 costs $0.0003 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.14 per million input tokens and $0.28 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What DeepSeek V4 Flash 0731 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

DeepSeek V4 Flash 0731 is a DeepSeek model priced at $0.14 per million input tokens and $0.28 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0003. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 78 of the 89 priced models cost more per answer, and 10 cost less.

#### What DeepSeek V4 Flash 0731 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.14 |
| Output, per million tokens | $0.28 |
| One standard answer | $0.0003 |
| One thousand answers | $0.2800 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 11 |

Prices are the published DeepSeek rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and DeepSeek V4 Flash 0731 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

DeepSeek V4 Flash 0731 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, DeepSeek V4 Flash 0731 costs 5.3 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 6.1 times less.

Its nearest neighbours from DeepSeek:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| DeepSeek V4 Flash 0731 (this page) | $0.0003 | 1M | 1 |
| DeepSeek V3.2 | $0.0005 | 164K | 1 |
| DeepSeek V3.1 | $0.0007 | 164K | 1 |
| R1 | $0.0019 | 64K | 2 |

It is the cheapest DeepSeek model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

DeepSeek V4 Flash 0731 accepts 1M tokens of context, which is roughly 2000 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0003 for one standard answer, ranking 11 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.14 and output $0.28 per million tokens
- $0.0003 for one standard answer, $0.2800 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 11 of 89 priced models by cost per answer

FAQ:

Q: How much does DeepSeek V4 Flash 0731 cost?
A: $0.14 per million input tokens and $0.28 per million output tokens at list rates, which works out to $0.0003 for one standard answer of 1,000 tokens in and 500 out, or $0.2800 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes DeepSeek V4 Flash 0731?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of DeepSeek V4 Flash 0731?
A: 1M tokens, which is roughly 2000 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is DeepSeek V4 Flash 0731 expensive compared to other models?
A: It ranks 11 of 89 priced models by cost per answer, so 78 cost more and 10 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### DeepSeek V4 Pro 0813: price, context window and credit cost

URL: https://whizi.io/models/deepseek-v4-pro-0813/
Updated: August 2026

Quick answer: DeepSeek V4 Pro 0813 costs $0.0030 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.188 per million input tokens and $3.564 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What DeepSeek V4 Pro 0813 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

DeepSeek V4 Pro 0813 is a DeepSeek model priced at $1.188 per million input tokens and $3.564 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0030. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 29 of the 89 priced models cost more per answer, and 59 cost less.

#### What DeepSeek V4 Pro 0813 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.188 |
| Output, per million tokens | $3.564 |
| One standard answer | $0.0030 |
| One thousand answers | $2.97 |
| Context window | 1M tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 60 |

Prices are the published DeepSeek rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and DeepSeek V4 Pro 0813 is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

DeepSeek V4 Pro 0813 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, DeepSeek V4 Pro 0813 costs 56 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.7 times more.

Its nearest neighbours from DeepSeek:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| DeepSeek V3.2 | $0.0005 | 164K | 1 |
| DeepSeek V3.1 | $0.0007 | 164K | 1 |
| R1 | $0.0019 | 64K | 2 |
| DeepSeek V4 Pro 0813 (this page) | $0.0030 | 1M | 2 |

If DeepSeek V4 Pro 0813 is more model than a given task needs, R1 is the cheaper DeepSeek option at $0.0019 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

DeepSeek V4 Pro 0813 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0030 for one standard answer, ranking 60 of 89 priced models
- 1M token context window, well above the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.188 and output $3.564 per million tokens
- $0.0030 for one standard answer, $2.97 for a thousand
- 1M token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 60 of 89 priced models by cost per answer

FAQ:

Q: How much does DeepSeek V4 Pro 0813 cost?
A: $1.188 per million input tokens and $3.564 per million output tokens at list rates, which works out to $0.0030 for one standard answer of 1,000 tokens in and 500 out, or $2.97 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes DeepSeek V4 Pro 0813?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of DeepSeek V4 Pro 0813?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is DeepSeek V4 Pro 0813 expensive compared to other models?
A: It ranks 60 of 89 priced models by cost per answer, so 29 cost more and 59 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Fugu Ultra: price, context window and credit cost

URL: https://whizi.io/models/fugu-ultra/
Updated: August 2026

Quick answer: Fugu Ultra costs $0.0200 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $5 per million input tokens and $30 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 25 credits per message on the Powerhouse plan and above.

What Fugu Ultra costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Fugu Ultra is a Sakana model priced at $5 per million input tokens and $30 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0200. Inside Whizi it costs 25 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 1 of the 89 priced models cost more per answer, and 87 cost less.

#### What Fugu Ultra costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $5 |
| Output, per million tokens | $30 |
| One standard answer | $0.0200 |
| One thousand answers | $20.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 25 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 88 |

Prices are the published Sakana rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Fugu Ultra is charged at 25 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 320 |

Fugu Ultra is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Fugu Ultra costs 377 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 12 times more.

Its nearest neighbours from Sakana:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Sakana Namazu | $0.0029 | 262K | 4 |
| Fugu Ultra (this page) | $0.0200 | 1M | 25 |

If Fugu Ultra is more model than a given task needs, Sakana Namazu is the cheaper Sakana option at $0.0029 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Fugu Ultra accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0200 for one standard answer, ranking 88 of 89 priced models
- 1M token context window, well above the catalogue median
- 25 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $5 and output $30 per million tokens
- $0.0200 for one standard answer, $20.00 for a thousand
- 1M token context window
- 25 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 88 of 89 priced models by cost per answer

FAQ:

Q: How much does Fugu Ultra cost?
A: $5 per million input tokens and $30 per million output tokens at list rates, which works out to $0.0200 for one standard answer of 1,000 tokens in and 500 out, or $20.00 for a thousand of them. Inside Whizi it is 25 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Fugu Ultra?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 25 credits per message, a Powerhouse allowance of 8,000 credits covers 320 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Fugu Ultra?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Fugu Ultra expensive compared to other models?
A: It ranks 88 of 89 priced models by cost per answer, so 1 cost more and 87 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 2.5 Flash: price, context window and credit cost

URL: https://whizi.io/models/gemini-2-5-flash/
Updated: August 2026

Quick answer: Gemini 2.5 Flash costs $0.0015 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.3 per million input tokens and $2.5 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 2 credits per message on the Pro plan and above.

What Gemini 2.5 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 2.5 Flash is a Google model priced at $0.3 per million input tokens and $2.5 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0015. Inside Whizi it costs 2 credits per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 47 of the 89 priced models cost more per answer, and 41 cost less.

#### What Gemini 2.5 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.3 |
| Output, per million tokens | $2.5 |
| One standard answer | $0.0015 |
| One thousand answers | $1.55 |
| Context window | 1M tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 42 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 2.5 Flash is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 1,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Gemini 2.5 Flash is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 2.5 Flash costs 29 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 3.1 Flash Lite | $0.0010 | 1M | 1 |
| Gemini 3.7 Flash | $0.0013 | 1M | 2 |
| Gemini 2.5 Flash (this page) | $0.0015 | 1M | 2 |
| Gemini 3.6 Flash | $0.0026 | 1M | 3 |

If Gemini 2.5 Flash is more model than a given task needs, Gemini 3.7 Flash is the cheaper Google option at $0.0013 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 2.5 Flash accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0015 for one standard answer, ranking 42 of 89 priced models
- 1M token context window, well above the catalogue median
- 2 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.3 and output $2.5 per million tokens
- $0.0015 for one standard answer, $1.55 for a thousand
- 1M token context window
- 2 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 42 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 2.5 Flash cost?
A: $0.3 per million input tokens and $2.5 per million output tokens at list rates, which works out to $0.0015 for one standard answer of 1,000 tokens in and 500 out, or $1.55 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 2.5 Flash?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 2 credits per message, a Pro allowance of 2,000 credits covers 1,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 2.5 Flash?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 2.5 Flash expensive compared to other models?
A: It ranks 42 of 89 priced models by cost per answer, so 47 cost more and 41 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 2.5 Flash Lite: price, context window and credit cost

URL: https://whizi.io/models/gemini-2-5-flash-lite/
Updated: August 2026

Quick answer: Gemini 2.5 Flash Lite costs $0.0003 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.1 per million input tokens and $0.4 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Pro plan and above.

What Gemini 2.5 Flash Lite costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 2.5 Flash Lite is a Google model priced at $0.1 per million input tokens and $0.4 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0003. Inside Whizi it costs 1 credit per message and needs the Pro plan or above.

That makes it one of the cheaper models in the catalogue: 77 of the 89 priced models cost more per answer, and 11 cost less.

#### What Gemini 2.5 Flash Lite costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.1 |
| Output, per million tokens | $0.4 |
| One standard answer | $0.0003 |
| One thousand answers | $0.3000 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 12 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 2.5 Flash Lite is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Gemini 2.5 Flash Lite is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 2.5 Flash Lite costs 5.7 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 5.7 times less.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 2.5 Flash Lite (this page) | $0.0003 | 1M | 1 |
| Gemini 3.1 Flash Lite | $0.0010 | 1M | 1 |
| Gemini 3.7 Flash | $0.0013 | 1M | 2 |
| Gemini 2.5 Flash | $0.0015 | 1M | 2 |

It is the cheapest Google model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Gemini 2.5 Flash Lite accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0003 for one standard answer, ranking 12 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.1 and output $0.4 per million tokens
- $0.0003 for one standard answer, $0.3000 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Pro plan or above
- Ranks 12 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 2.5 Flash Lite cost?
A: $0.1 per million input tokens and $0.4 per million output tokens at list rates, which works out to $0.0003 for one standard answer of 1,000 tokens in and 500 out, or $0.3000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 2.5 Flash Lite?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 1 credits per message, a Pro allowance of 2,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 2.5 Flash Lite?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 2.5 Flash Lite expensive compared to other models?
A: It ranks 12 of 89 priced models by cost per answer, so 77 cost more and 11 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 3.1 Flash Lite: price, context window and credit cost

URL: https://whizi.io/models/gemini-3-1-flash-lite/
Updated: August 2026

Quick answer: Gemini 3.1 Flash Lite costs $0.0010 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.25 per million input tokens and $1.5 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Gemini 3.1 Flash Lite costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 3.1 Flash Lite is a Google model priced at $0.25 per million input tokens and $1.5 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0010. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 57 of the 89 priced models cost more per answer, and 31 cost less.

#### What Gemini 3.1 Flash Lite costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.25 |
| Output, per million tokens | $1.5 |
| One standard answer | $0.0010 |
| One thousand answers | $1.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 32 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 3.1 Flash Lite is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Gemini 3.1 Flash Lite is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 3.1 Flash Lite costs 19 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.7 times less.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 2.5 Flash Lite | $0.0003 | 1M | 1 |
| Gemini 3.1 Flash Lite (this page) | $0.0010 | 1M | 1 |
| Gemini 3.7 Flash | $0.0013 | 1M | 2 |
| Gemini 2.5 Flash | $0.0015 | 1M | 2 |

If Gemini 3.1 Flash Lite is more model than a given task needs, Gemini 2.5 Flash Lite is the cheaper Google option at $0.0003 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 3.1 Flash Lite accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0010 for one standard answer, ranking 32 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.25 and output $1.5 per million tokens
- $0.0010 for one standard answer, $1.00 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 32 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 3.1 Flash Lite cost?
A: $0.25 per million input tokens and $1.5 per million output tokens at list rates, which works out to $0.0010 for one standard answer of 1,000 tokens in and 500 out, or $1.00 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 3.1 Flash Lite?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 3.1 Flash Lite?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 3.1 Flash Lite expensive compared to other models?
A: It ranks 32 of 89 priced models by cost per answer, so 57 cost more and 31 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 3.1 Pro Preview: price, context window and credit cost

URL: https://whizi.io/models/gemini-3-1-pro-preview/
Updated: August 2026

Quick answer: Gemini 3.1 Pro Preview costs $0.0080 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $12 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 10 credits per message on the Pro plan and above.

What Gemini 3.1 Pro Preview costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 3.1 Pro Preview is a Google model priced at $2 per million input tokens and $12 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0080. Inside Whizi it costs 10 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 12 of the 89 priced models cost more per answer, and 76 cost less.

#### What Gemini 3.1 Pro Preview costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $12 |
| One standard answer | $0.0080 |
| One thousand answers | $8.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 77 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 3.1 Pro Preview is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 200 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Gemini 3.1 Pro Preview is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 3.1 Pro Preview costs 151 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 4.7 times more.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 2.5 Flash | $0.0015 | 1M | 2 |
| Gemini 3.6 Flash | $0.0026 | 1M | 3 |
| Gemini 3.5 Flash | $0.0060 | 1M | 8 |
| Gemini 3.1 Pro Preview (this page) | $0.0080 | 1M | 10 |

If Gemini 3.1 Pro Preview is more model than a given task needs, Gemini 3.5 Flash is the cheaper Google option at $0.0060 per answer and 8 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 3.1 Pro Preview accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0080 for one standard answer, ranking 77 of 89 priced models
- 1M token context window, well above the catalogue median
- 10 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $2 and output $12 per million tokens
- $0.0080 for one standard answer, $8.00 for a thousand
- 1M token context window
- 10 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 77 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 3.1 Pro Preview cost?
A: $2 per million input tokens and $12 per million output tokens at list rates, which works out to $0.0080 for one standard answer of 1,000 tokens in and 500 out, or $8.00 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 3.1 Pro Preview?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 10 credits per message, a Pro allowance of 2,000 credits covers 200 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 3.1 Pro Preview?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 3.1 Pro Preview expensive compared to other models?
A: It ranks 77 of 89 priced models by cost per answer, so 12 cost more and 76 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 3.5 Flash: price, context window and credit cost

URL: https://whizi.io/models/gemini-3-5-flash/
Updated: August 2026

Quick answer: Gemini 3.5 Flash costs $0.0060 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.5 per million input tokens and $9 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 8 credits per message on the Pro plan and above.

What Gemini 3.5 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 3.5 Flash is a Google model priced at $1.5 per million input tokens and $9 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0060. Inside Whizi it costs 8 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 16 of the 89 priced models cost more per answer, and 72 cost less.

#### What Gemini 3.5 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.5 |
| Output, per million tokens | $9 |
| One standard answer | $0.0060 |
| One thousand answers | $6.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 8 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 73 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 3.5 Flash is charged at 8 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 250 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,000 |

Gemini 3.5 Flash is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 3.5 Flash costs 113 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 3.5 times more.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 2.5 Flash | $0.0015 | 1M | 2 |
| Gemini 3.6 Flash | $0.0026 | 1M | 3 |
| Gemini 3.5 Flash (this page) | $0.0060 | 1M | 8 |
| Gemini 3.1 Pro Preview | $0.0080 | 1M | 10 |

If Gemini 3.5 Flash is more model than a given task needs, Gemini 3.6 Flash is the cheaper Google option at $0.0026 per answer and 3 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 3.5 Flash accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0060 for one standard answer, ranking 73 of 89 priced models
- 1M token context window, well above the catalogue median
- 8 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $1.5 and output $9 per million tokens
- $0.0060 for one standard answer, $6.00 for a thousand
- 1M token context window
- 8 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 73 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 3.5 Flash cost?
A: $1.5 per million input tokens and $9 per million output tokens at list rates, which works out to $0.0060 for one standard answer of 1,000 tokens in and 500 out, or $6.00 for a thousand of them. Inside Whizi it is 8 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 3.5 Flash?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 8 credits per message, a Pro allowance of 2,000 credits covers 250 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 3.5 Flash?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 3.5 Flash expensive compared to other models?
A: It ranks 73 of 89 priced models by cost per answer, so 16 cost more and 72 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 3.5 Flash Lite: price, context window and credit cost

URL: https://whizi.io/models/gemini-3-5-flash-lite/
Updated: August 2026

Quick answer: Gemini 3.5 Flash Lite costs $0.0015 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.3 per million input tokens and $2.5 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What Gemini 3.5 Flash Lite costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 3.5 Flash Lite is a Google model priced at $0.3 per million input tokens and $2.5 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0015. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 46 of the 89 priced models cost more per answer, and 42 cost less.

#### What Gemini 3.5 Flash Lite costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.3 |
| Output, per million tokens | $2.5 |
| One standard answer | $0.0015 |
| One thousand answers | $1.55 |
| Context window | 1M tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 43 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 3.5 Flash Lite is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Gemini 3.5 Flash Lite is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 3.5 Flash Lite costs 29 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 3.1 Flash Lite | $0.0010 | 1M | 1 |
| Gemini 3.7 Flash | $0.0013 | 1M | 2 |
| Gemini 3.5 Flash Lite (this page) | $0.0015 | 1M | 2 |
| Gemini 3.6 Flash | $0.0026 | 1M | 3 |

If Gemini 3.5 Flash Lite is more model than a given task needs, Gemini 3.7 Flash is the cheaper Google option at $0.0013 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 3.5 Flash Lite accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0015 for one standard answer, ranking 43 of 89 priced models
- 1M token context window, well above the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.3 and output $2.5 per million tokens
- $0.0015 for one standard answer, $1.55 for a thousand
- 1M token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 43 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 3.5 Flash Lite cost?
A: $0.3 per million input tokens and $2.5 per million output tokens at list rates, which works out to $0.0015 for one standard answer of 1,000 tokens in and 500 out, or $1.55 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 3.5 Flash Lite?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 3.5 Flash Lite?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 3.5 Flash Lite expensive compared to other models?
A: It ranks 43 of 89 priced models by cost per answer, so 46 cost more and 42 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 3.6 Flash: price, context window and credit cost

URL: https://whizi.io/models/gemini-3-6-flash/
Updated: August 2026

Quick answer: Gemini 3.6 Flash costs $0.0026 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.75 per million input tokens and $3.75 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 3 credits per message on the Pro plan and above.

What Gemini 3.6 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 3.6 Flash is a Google model priced at $0.75 per million input tokens and $3.75 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0026. Inside Whizi it costs 3 credits per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 31 of the 89 priced models cost more per answer, and 57 cost less.

#### What Gemini 3.6 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.75 |
| Output, per million tokens | $3.75 |
| One standard answer | $0.0026 |
| One thousand answers | $2.63 |
| Context window | 1M tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 58 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 3.6 Flash is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 666 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

Gemini 3.6 Flash is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 3.6 Flash costs 50 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.5 times more.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 3.1 Flash Lite | $0.0010 | 1M | 1 |
| Gemini 3.7 Flash | $0.0013 | 1M | 2 |
| Gemini 2.5 Flash | $0.0015 | 1M | 2 |
| Gemini 3.6 Flash (this page) | $0.0026 | 1M | 3 |

If Gemini 3.6 Flash is more model than a given task needs, Gemini 2.5 Flash is the cheaper Google option at $0.0015 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 3.6 Flash accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0026 for one standard answer, ranking 58 of 89 priced models
- 1M token context window, well above the catalogue median
- 3 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.75 and output $3.75 per million tokens
- $0.0026 for one standard answer, $2.63 for a thousand
- 1M token context window
- 3 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 58 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 3.6 Flash cost?
A: $0.75 per million input tokens and $3.75 per million output tokens at list rates, which works out to $0.0026 for one standard answer of 1,000 tokens in and 500 out, or $2.63 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 3.6 Flash?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 3 credits per message, a Pro allowance of 2,000 credits covers 666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 3.6 Flash?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 3.6 Flash expensive compared to other models?
A: It ranks 58 of 89 priced models by cost per answer, so 31 cost more and 57 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Gemini 3.7 Flash: price, context window and credit cost

URL: https://whizi.io/models/gemini-3-7-flash/
Updated: August 2026

Quick answer: Gemini 3.7 Flash costs $0.0013 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.375 per million input tokens and $1.875 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 2 credits per message on the Pro plan and above.

What Gemini 3.7 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Gemini 3.7 Flash is a Google model priced at $0.375 per million input tokens and $1.875 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0013. Inside Whizi it costs 2 credits per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 51 of the 89 priced models cost more per answer, and 37 cost less.

#### What Gemini 3.7 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.375 |
| Output, per million tokens | $1.875 |
| One standard answer | $0.0013 |
| One thousand answers | $1.31 |
| Context window | 1M tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 38 |

Prices are the published Google rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Gemini 3.7 Flash is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 1,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Gemini 3.7 Flash is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Gemini 3.7 Flash costs 25 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.3 times less.

Its nearest neighbours from Google:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Gemini 2.5 Flash Lite | $0.0003 | 1M | 1 |
| Gemini 3.1 Flash Lite | $0.0010 | 1M | 1 |
| Gemini 3.7 Flash (this page) | $0.0013 | 1M | 2 |
| Gemini 2.5 Flash | $0.0015 | 1M | 2 |

If Gemini 3.7 Flash is more model than a given task needs, Gemini 3.1 Flash Lite is the cheaper Google option at $0.0010 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Gemini 3.7 Flash accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0013 for one standard answer, ranking 38 of 89 priced models
- 1M token context window, well above the catalogue median
- 2 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.375 and output $1.875 per million tokens
- $0.0013 for one standard answer, $1.31 for a thousand
- 1M token context window
- 2 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 38 of 89 priced models by cost per answer

FAQ:

Q: How much does Gemini 3.7 Flash cost?
A: $0.375 per million input tokens and $1.875 per million output tokens at list rates, which works out to $0.0013 for one standard answer of 1,000 tokens in and 500 out, or $1.31 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Gemini 3.7 Flash?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 2 credits per message, a Pro allowance of 2,000 credits covers 1,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Gemini 3.7 Flash?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Gemini 3.7 Flash expensive compared to other models?
A: It ranks 38 of 89 priced models by cost per answer, so 51 cost more and 37 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GLM 4.6: price, context window and credit cost

URL: https://whizi.io/models/glm-4-6/
Updated: August 2026

Quick answer: GLM 4.6 costs $0.0015 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.5 per million input tokens and $2 per million output tokens. The context window is 205K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What GLM 4.6 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GLM 4.6 is a Z.ai model priced at $0.5 per million input tokens and $2 per million output tokens, with a 205K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0015. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 48 of the 89 priced models cost more per answer, and 40 cost less.

#### What GLM 4.6 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.5 |
| Output, per million tokens | $2 |
| One standard answer | $0.0015 |
| One thousand answers | $1.50 |
| Context window | 205K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 41 |

Prices are the published Z.ai rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GLM 4.6 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

GLM 4.6 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GLM 4.6 costs 28 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Z.ai:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GLM 4.7 | $0.0013 | 205K | 2 |
| GLM 4.6 (this page) | $0.0015 | 205K | 1 |
| GLM 5 | $0.0016 | 205K | 2 |
| GLM 5.2 | $0.0025 | 1M | 3 |

If GLM 4.6 is more model than a given task needs, GLM 4.7 is the cheaper Z.ai option at $0.0013 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GLM 4.6 accepts 205K tokens of context, which is roughly 310 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0015 for one standard answer, ranking 41 of 89 priced models
- 205K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.5 and output $2 per million tokens
- $0.0015 for one standard answer, $1.50 for a thousand
- 205K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 41 of 89 priced models by cost per answer

FAQ:

Q: How much does GLM 4.6 cost?
A: $0.5 per million input tokens and $2 per million output tokens at list rates, which works out to $0.0015 for one standard answer of 1,000 tokens in and 500 out, or $1.50 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GLM 4.6?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GLM 4.6?
A: 205K tokens, which is roughly 310 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GLM 4.6 expensive compared to other models?
A: It ranks 41 of 89 priced models by cost per answer, so 48 cost more and 40 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GLM 4.7: price, context window and credit cost

URL: https://whizi.io/models/glm-4-7/
Updated: August 2026

Quick answer: GLM 4.7 costs $0.0013 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.4 per million input tokens and $1.75 per million output tokens. The context window is 205K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What GLM 4.7 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GLM 4.7 is a Z.ai model priced at $0.4 per million input tokens and $1.75 per million output tokens, with a 205K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0013. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 52 of the 89 priced models cost more per answer, and 36 cost less.

#### What GLM 4.7 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.4 |
| Output, per million tokens | $1.75 |
| One standard answer | $0.0013 |
| One thousand answers | $1.27 |
| Context window | 205K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 37 |

Prices are the published Z.ai rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GLM 4.7 is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

GLM 4.7 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GLM 4.7 costs 24 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.3 times less.

Its nearest neighbours from Z.ai:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GLM 4.7 (this page) | $0.0013 | 205K | 2 |
| GLM 4.6 | $0.0015 | 205K | 1 |
| GLM 5 | $0.0016 | 205K | 2 |
| GLM 5.2 | $0.0025 | 1M | 3 |

It is the cheapest Z.ai model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

GLM 4.7 accepts 205K tokens of context, which is roughly 310 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0013 for one standard answer, ranking 37 of 89 priced models
- 205K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.4 and output $1.75 per million tokens
- $0.0013 for one standard answer, $1.27 for a thousand
- 205K token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 37 of 89 priced models by cost per answer

FAQ:

Q: How much does GLM 4.7 cost?
A: $0.4 per million input tokens and $1.75 per million output tokens at list rates, which works out to $0.0013 for one standard answer of 1,000 tokens in and 500 out, or $1.27 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GLM 4.7?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GLM 4.7?
A: 205K tokens, which is roughly 310 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GLM 4.7 expensive compared to other models?
A: It ranks 37 of 89 priced models by cost per answer, so 52 cost more and 36 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GLM 5: price, context window and credit cost

URL: https://whizi.io/models/glm-5/
Updated: August 2026

Quick answer: GLM 5 costs $0.0016 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.6 per million input tokens and $1.92 per million output tokens. The context window is 205K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Pro plan and above.

What GLM 5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GLM 5 is a Z.ai model priced at $0.6 per million input tokens and $1.92 per million output tokens, with a 205K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0016. Inside Whizi it costs 2 credits per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 45 of the 89 priced models cost more per answer, and 43 cost less.

#### What GLM 5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.6 |
| Output, per million tokens | $1.92 |
| One standard answer | $0.0016 |
| One thousand answers | $1.56 |
| Context window | 205K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 44 |

Prices are the published Z.ai rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GLM 5 is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 1,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

GLM 5 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GLM 5 costs 29 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Z.ai:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GLM 4.7 | $0.0013 | 205K | 2 |
| GLM 4.6 | $0.0015 | 205K | 1 |
| GLM 5 (this page) | $0.0016 | 205K | 2 |
| GLM 5.2 | $0.0025 | 1M | 3 |

If GLM 5 is more model than a given task needs, GLM 4.6 is the cheaper Z.ai option at $0.0015 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GLM 5 accepts 205K tokens of context, which is roughly 310 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0016 for one standard answer, ranking 44 of 89 priced models
- 205K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.6 and output $1.92 per million tokens
- $0.0016 for one standard answer, $1.56 for a thousand
- 205K token context window
- 2 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 44 of 89 priced models by cost per answer

FAQ:

Q: How much does GLM 5 cost?
A: $0.6 per million input tokens and $1.92 per million output tokens at list rates, which works out to $0.0016 for one standard answer of 1,000 tokens in and 500 out, or $1.56 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GLM 5?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 2 credits per message, a Pro allowance of 2,000 credits covers 1,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GLM 5?
A: 205K tokens, which is roughly 310 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GLM 5 expensive compared to other models?
A: It ranks 44 of 89 priced models by cost per answer, so 45 cost more and 43 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GLM 5.2: price, context window and credit cost

URL: https://whizi.io/models/glm-5-2/
Updated: August 2026

Quick answer: GLM 5.2 costs $0.0025 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.966 per million input tokens and $3.036 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What GLM 5.2 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GLM 5.2 is a Z.ai model priced at $0.966 per million input tokens and $3.036 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0025. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 33 of the 89 priced models cost more per answer, and 55 cost less.

#### What GLM 5.2 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.966 |
| Output, per million tokens | $3.036 |
| One standard answer | $0.0025 |
| One thousand answers | $2.48 |
| Context window | 1M tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 56 |

Prices are the published Z.ai rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GLM 5.2 is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

GLM 5.2 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GLM 5.2 costs 47 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.4 times more.

Its nearest neighbours from Z.ai:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GLM 4.6 | $0.0015 | 205K | 1 |
| GLM 5 | $0.0016 | 205K | 2 |
| GLM 5.2 (this page) | $0.0025 | 1M | 3 |
| GLM 5.3 | $0.0036 | 1M | 5 |

If GLM 5.2 is more model than a given task needs, GLM 5 is the cheaper Z.ai option at $0.0016 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GLM 5.2 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0025 for one standard answer, ranking 56 of 89 priced models
- 1M token context window, well above the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.966 and output $3.036 per million tokens
- $0.0025 for one standard answer, $2.48 for a thousand
- 1M token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 56 of 89 priced models by cost per answer

FAQ:

Q: How much does GLM 5.2 cost?
A: $0.966 per million input tokens and $3.036 per million output tokens at list rates, which works out to $0.0025 for one standard answer of 1,000 tokens in and 500 out, or $2.48 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GLM 5.2?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GLM 5.2?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GLM 5.2 expensive compared to other models?
A: It ranks 56 of 89 priced models by cost per answer, so 33 cost more and 55 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GLM 5.3: price, context window and credit cost

URL: https://whizi.io/models/glm-5-3/
Updated: August 2026

Quick answer: GLM 5.3 costs $0.0036 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.4 per million input tokens and $4.4 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 5 credits per message on the Powerhouse plan and above.

What GLM 5.3 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GLM 5.3 is a Z.ai model priced at $1.4 per million input tokens and $4.4 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0036. Inside Whizi it costs 5 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 23 of the 89 priced models cost more per answer, and 65 cost less.

#### What GLM 5.3 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.4 |
| Output, per million tokens | $4.4 |
| One standard answer | $0.0036 |
| One thousand answers | $3.60 |
| Context window | 1M tokens |
| Credits per message in Whizi | 5 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 66 |

Prices are the published Z.ai rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GLM 5.3 is charged at 5 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,600 |

GLM 5.3 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GLM 5.3 costs 68 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.1 times more.

Its nearest neighbours from Z.ai:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GLM 4.6 | $0.0015 | 205K | 1 |
| GLM 5 | $0.0016 | 205K | 2 |
| GLM 5.2 | $0.0025 | 1M | 3 |
| GLM 5.3 (this page) | $0.0036 | 1M | 5 |

If GLM 5.3 is more model than a given task needs, GLM 5.2 is the cheaper Z.ai option at $0.0025 per answer and 3 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GLM 5.3 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0036 for one standard answer, ranking 66 of 89 priced models
- 1M token context window, well above the catalogue median
- 5 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.4 and output $4.4 per million tokens
- $0.0036 for one standard answer, $3.60 for a thousand
- 1M token context window
- 5 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 66 of 89 priced models by cost per answer

FAQ:

Q: How much does GLM 5.3 cost?
A: $1.4 per million input tokens and $4.4 per million output tokens at list rates, which works out to $0.0036 for one standard answer of 1,000 tokens in and 500 out, or $3.60 for a thousand of them. Inside Whizi it is 5 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GLM 5.3?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 5 credits per message, a Powerhouse allowance of 8,000 credits covers 1,600 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GLM 5.3?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GLM 5.3 expensive compared to other models?
A: It ranks 66 of 89 priced models by cost per answer, so 23 cost more and 65 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.4: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-4/
Updated: August 2026

Quick answer: GPT-5.4 costs $0.0100 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2.5 per million input tokens and $15 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 15 credits per message on the Powerhouse plan and above.

What GPT-5.4 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.4 is an OpenAI model priced at $2.5 per million input tokens and $15 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0100. Inside Whizi it costs 15 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 10 of the 89 priced models cost more per answer, and 78 cost less.

#### What GPT-5.4 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2.5 |
| Output, per million tokens | $15 |
| One standard answer | $0.0100 |
| One thousand answers | $10.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 15 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 79 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.4 is charged at 15 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 533 |

GPT-5.4 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.4 costs 189 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 5.8 times more.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.4 Nano | $0.0008 | 400K | 2 |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |
| GPT-5.4 (this page) | $0.0100 | 1M | 15 |

If GPT-5.4 is more model than a given task needs, GPT-5.6 Terra is the cheaper OpenAI option at $0.0080 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT-5.4 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0100 for one standard answer, ranking 79 of 89 priced models
- 1M token context window, well above the catalogue median
- 15 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $2.5 and output $15 per million tokens
- $0.0100 for one standard answer, $10.00 for a thousand
- 1M token context window
- 15 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 79 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.4 cost?
A: $2.5 per million input tokens and $15 per million output tokens at list rates, which works out to $0.0100 for one standard answer of 1,000 tokens in and 500 out, or $10.00 for a thousand of them. Inside Whizi it is 15 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.4?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 15 credits per message, a Powerhouse allowance of 8,000 credits covers 533 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT-5.4?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.4 expensive compared to other models?
A: It ranks 79 of 89 priced models by cost per answer, so 10 cost more and 78 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.4 Mini: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-4-mini/
Updated: August 2026

Quick answer: GPT-5.4 Mini costs $0.0030 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.75 per million input tokens and $4.5 per million output tokens. The context window is 400K tokens, around the catalogue median. Inside Whizi it costs 4 credits per message on the Powerhouse plan and above.

What GPT-5.4 Mini costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.4 Mini is an OpenAI model priced at $0.75 per million input tokens and $4.5 per million output tokens, with a 400K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0030. Inside Whizi it costs 4 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 27 of the 89 priced models cost more per answer, and 61 cost less.

#### What GPT-5.4 Mini costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.75 |
| Output, per million tokens | $4.5 |
| One standard answer | $0.0030 |
| One thousand answers | $3.00 |
| Context window | 400K tokens |
| Credits per message in Whizi | 4 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 62 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.4 Mini is charged at 4 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,000 |

GPT-5.4 Mini is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.4 Mini costs 57 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.7 times more.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.6 Luna | $0.0008 | 1M | 1 |
| GPT-5.4 Nano | $0.0008 | 400K | 2 |
| GPT-5.4 Mini (this page) | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |

If GPT-5.4 Mini is more model than a given task needs, GPT-5.4 Nano is the cheaper OpenAI option at $0.0008 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT-5.4 Mini accepts 400K tokens of context, which is roughly 600 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0030 for one standard answer, ranking 62 of 89 priced models
- 400K token context window, around the catalogue median
- 4 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.75 and output $4.5 per million tokens
- $0.0030 for one standard answer, $3.00 for a thousand
- 400K token context window
- 4 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 62 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.4 Mini cost?
A: $0.75 per million input tokens and $4.5 per million output tokens at list rates, which works out to $0.0030 for one standard answer of 1,000 tokens in and 500 out, or $3.00 for a thousand of them. Inside Whizi it is 4 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.4 Mini?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 4 credits per message, a Powerhouse allowance of 8,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT-5.4 Mini?
A: 400K tokens, which is roughly 600 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.4 Mini expensive compared to other models?
A: It ranks 62 of 89 priced models by cost per answer, so 27 cost more and 61 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.4 Nano: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-4-nano/
Updated: August 2026

Quick answer: GPT-5.4 Nano costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.2 per million input tokens and $1.25 per million output tokens. The context window is 400K tokens, around the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What GPT-5.4 Nano costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.4 Nano is an OpenAI model priced at $0.2 per million input tokens and $1.25 per million output tokens, with a 400K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 62 of the 89 priced models cost more per answer, and 26 cost less.

#### What GPT-5.4 Nano costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.2 |
| Output, per million tokens | $1.25 |
| One standard answer | $0.0008 |
| One thousand answers | $0.8250 |
| Context window | 400K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 27 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.4 Nano is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

GPT-5.4 Nano is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.4 Nano costs 16 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.1 times less.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.6 Luna | $0.0008 | 1M | 1 |
| GPT-5.4 Nano (this page) | $0.0008 | 400K | 2 |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |

If GPT-5.4 Nano is more model than a given task needs, GPT-5.6 Luna is the cheaper OpenAI option at $0.0008 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT-5.4 Nano accepts 400K tokens of context, which is roughly 600 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 27 of 89 priced models
- 400K token context window, around the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.2 and output $1.25 per million tokens
- $0.0008 for one standard answer, $0.8250 for a thousand
- 400K token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 27 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.4 Nano cost?
A: $0.2 per million input tokens and $1.25 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.8250 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.4 Nano?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT-5.4 Nano?
A: 400K tokens, which is roughly 600 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.4 Nano expensive compared to other models?
A: It ranks 27 of 89 priced models by cost per answer, so 62 cost more and 26 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.5: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-5/
Updated: August 2026

Quick answer: GPT-5.5 costs $0.0200 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $5 per million input tokens and $30 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 20 credits per message on the Pro plan and above.

What GPT-5.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.5 is an OpenAI model priced at $5 per million input tokens and $30 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0200. Inside Whizi it costs 20 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 3 of the 89 priced models cost more per answer, and 85 cost less.

#### What GPT-5.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $5 |
| Output, per million tokens | $30 |
| One standard answer | $0.0200 |
| One thousand answers | $20.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 20 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 86 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.5 is charged at 20 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 100 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 400 |

GPT-5.5 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.5 costs 377 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 12 times more.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |
| GPT-5.4 | $0.0100 | 1M | 15 |
| GPT-5.5 (this page) | $0.0200 | 1M | 20 |

If GPT-5.5 is more model than a given task needs, GPT-5.4 is the cheaper OpenAI option at $0.0100 per answer and 15 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT-5.5 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0200 for one standard answer, ranking 86 of 89 priced models
- 1M token context window, well above the catalogue median
- 20 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $5 and output $30 per million tokens
- $0.0200 for one standard answer, $20.00 for a thousand
- 1M token context window
- 20 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 86 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.5 cost?
A: $5 per million input tokens and $30 per million output tokens at list rates, which works out to $0.0200 for one standard answer of 1,000 tokens in and 500 out, or $20.00 for a thousand of them. Inside Whizi it is 20 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.5?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 20 credits per message, a Pro allowance of 2,000 credits covers 100 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT-5.5?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.5 expensive compared to other models?
A: It ranks 86 of 89 priced models by cost per answer, so 3 cost more and 85 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.6 Luna: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-6-luna/
Updated: August 2026

Quick answer: GPT-5.6 Luna costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.2 per million input tokens and $1.2 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on every plan, including Starter.

What GPT-5.6 Luna costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.6 Luna is an OpenAI model priced at $0.2 per million input tokens and $1.2 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 1 credit per message and needs the Starter plan or above.

That makes it one of the cheaper models in the catalogue: 63 of the 89 priced models cost more per answer, and 25 cost less.

#### What GPT-5.6 Luna costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.2 |
| Output, per million tokens | $1.2 |
| One standard answer | $0.0008 |
| One thousand answers | $0.8000 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Starter |
| Rank by cost, of 89 priced models | 26 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.6 Luna is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Starter | $15.99, or $10.99 billed yearly | 400 | 400 |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

GPT-5.6 Luna is included on every plan, including Starter. The Starter models are the ones that cost 1 credit per message, which is what makes a 400 credit allowance go a long way. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.6 Luna costs 15 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.1 times less.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.6 Luna (this page) | $0.0008 | 1M | 1 |
| GPT-5.4 Nano | $0.0008 | 400K | 2 |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |

It is the cheapest OpenAI model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

GPT-5.6 Luna accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 26 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Starter plan and above

Checklist:

- Input $0.2 and output $1.2 per million tokens
- $0.0008 for one standard answer, $0.8000 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Starter plan or above
- Ranks 26 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.6 Luna cost?
A: $0.2 per million input tokens and $1.2 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.8000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.6 Luna?
A: Every plan, including Starter at $15.99 per month or $10.99 billed yearly. At 1 credit per message, a Starter allowance of 400 credits covers 400 messages on it.

Q: What is the context window of GPT-5.6 Luna?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.6 Luna expensive compared to other models?
A: It ranks 26 of 89 priced models by cost per answer, so 63 cost more and 25 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.6 Sol: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-6-sol/
Updated: August 2026

Quick answer: GPT-5.6 Sol costs $0.0100 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2.5 per million input tokens and $15 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 20 credits per message on the Powerhouse plan and above.

What GPT-5.6 Sol costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.6 Sol is an OpenAI model priced at $2.5 per million input tokens and $15 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0100. Inside Whizi it costs 20 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 9 of the 89 priced models cost more per answer, and 79 cost less.

#### What GPT-5.6 Sol costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2.5 |
| Output, per million tokens | $15 |
| One standard answer | $0.0100 |
| One thousand answers | $10.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 20 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 80 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.6 Sol is charged at 20 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 400 |

GPT-5.6 Sol is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.6 Sol costs 189 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 5.8 times more.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.4 Nano | $0.0008 | 400K | 2 |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |
| GPT-5.6 Sol (this page) | $0.0100 | 1M | 20 |

If GPT-5.6 Sol is more model than a given task needs, GPT-5.6 Terra is the cheaper OpenAI option at $0.0080 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT-5.6 Sol accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0100 for one standard answer, ranking 80 of 89 priced models
- 1M token context window, well above the catalogue median
- 20 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $2.5 and output $15 per million tokens
- $0.0100 for one standard answer, $10.00 for a thousand
- 1M token context window
- 20 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 80 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.6 Sol cost?
A: $2.5 per million input tokens and $15 per million output tokens at list rates, which works out to $0.0100 for one standard answer of 1,000 tokens in and 500 out, or $10.00 for a thousand of them. Inside Whizi it is 20 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.6 Sol?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 20 credits per message, a Powerhouse allowance of 8,000 credits covers 400 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT-5.6 Sol?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.6 Sol expensive compared to other models?
A: It ranks 80 of 89 priced models by cost per answer, so 9 cost more and 79 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT-5.6 Terra: price, context window and credit cost

URL: https://whizi.io/models/gpt-5-6-terra/
Updated: August 2026

Quick answer: GPT-5.6 Terra costs $0.0080 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $12 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 4 credits per message on the Pro plan and above.

What GPT-5.6 Terra costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT-5.6 Terra is an OpenAI model priced at $2 per million input tokens and $12 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0080. Inside Whizi it costs 4 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 11 of the 89 priced models cost more per answer, and 77 cost less.

#### What GPT-5.6 Terra costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $12 |
| One standard answer | $0.0080 |
| One thousand answers | $8.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 4 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 78 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT-5.6 Terra is charged at 4 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 500 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,000 |

GPT-5.6 Terra is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT-5.6 Terra costs 151 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 4.7 times more.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.4 Nano | $0.0008 | 400K | 2 |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra (this page) | $0.0080 | 1M | 4 |
| GPT-5.4 | $0.0100 | 1M | 15 |

If GPT-5.6 Terra is more model than a given task needs, GPT-5.4 Mini is the cheaper OpenAI option at $0.0030 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT-5.6 Terra accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0080 for one standard answer, ranking 78 of 89 priced models
- 1M token context window, well above the catalogue median
- 4 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $2 and output $12 per million tokens
- $0.0080 for one standard answer, $8.00 for a thousand
- 1M token context window
- 4 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 78 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT-5.6 Terra cost?
A: $2 per million input tokens and $12 per million output tokens at list rates, which works out to $0.0080 for one standard answer of 1,000 tokens in and 500 out, or $8.00 for a thousand of them. Inside Whizi it is 4 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT-5.6 Terra?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 4 credits per message, a Pro allowance of 2,000 credits covers 500 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT-5.6 Terra?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT-5.6 Terra expensive compared to other models?
A: It ranks 78 of 89 priced models by cost per answer, so 11 cost more and 77 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### GPT Chat Latest: price, context window and credit cost

URL: https://whizi.io/models/gpt-chat-latest/
Updated: August 2026

Quick answer: GPT Chat Latest costs $0.0200 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $5 per million input tokens and $30 per million output tokens. The context window is 400K tokens, around the catalogue median. Inside Whizi it costs 25 credits per message on the Powerhouse plan and above.

What GPT Chat Latest costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

GPT Chat Latest is an OpenAI model priced at $5 per million input tokens and $30 per million output tokens, with a 400K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0200. Inside Whizi it costs 25 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 2 of the 89 priced models cost more per answer, and 86 cost less.

#### What GPT Chat Latest costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $5 |
| Output, per million tokens | $30 |
| One standard answer | $0.0200 |
| One thousand answers | $20.00 |
| Context window | 400K tokens |
| Credits per message in Whizi | 25 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 87 |

Prices are the published OpenAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and GPT Chat Latest is charged at 25 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 320 |

GPT Chat Latest is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, GPT Chat Latest costs 377 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 12 times more.

Its nearest neighbours from OpenAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| GPT-5.4 Mini | $0.0030 | 400K | 4 |
| GPT-5.6 Terra | $0.0080 | 1M | 4 |
| GPT-5.4 | $0.0100 | 1M | 15 |
| GPT Chat Latest (this page) | $0.0200 | 400K | 25 |

If GPT Chat Latest is more model than a given task needs, GPT-5.4 is the cheaper OpenAI option at $0.0100 per answer and 15 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

GPT Chat Latest accepts 400K tokens of context, which is roughly 600 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0200 for one standard answer, ranking 87 of 89 priced models
- 400K token context window, around the catalogue median
- 25 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $5 and output $30 per million tokens
- $0.0200 for one standard answer, $20.00 for a thousand
- 400K token context window
- 25 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 87 of 89 priced models by cost per answer

FAQ:

Q: How much does GPT Chat Latest cost?
A: $5 per million input tokens and $30 per million output tokens at list rates, which works out to $0.0200 for one standard answer of 1,000 tokens in and 500 out, or $20.00 for a thousand of them. Inside Whizi it is 25 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes GPT Chat Latest?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 25 credits per message, a Powerhouse allowance of 8,000 credits covers 320 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of GPT Chat Latest?
A: 400K tokens, which is roughly 600 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is GPT Chat Latest expensive compared to other models?
A: It ranks 87 of 89 priced models by cost per answer, so 2 cost more and 86 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Granite 4.1 8B: price, context window and credit cost

URL: https://whizi.io/models/granite-4-1-8b/
Updated: August 2026

Quick answer: Granite 4.1 8B costs $0.0001 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.05 per million input tokens and $0.1 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Granite 4.1 8B costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Granite 4.1 8B is an Ibm Granite model priced at $0.05 per million input tokens and $0.1 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0001. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 84 of the 89 priced models cost more per answer, and 4 cost less.

#### What Granite 4.1 8B costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.05 |
| Output, per million tokens | $0.1 |
| One standard answer | $0.0001 |
| One thousand answers | $0.1000 |
| Context window | 131K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 5 |

Prices are the published Ibm Granite rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Granite 4.1 8B is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Granite 4.1 8B is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Granite 4.1 8B costs 1.9 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 17.2 times less.

It is the only Ibm Granite model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Ibm Granite model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Granite 4.1 8B accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0001 for one standard answer, ranking 5 of 89 priced models
- 131K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.05 and output $0.1 per million tokens
- $0.0001 for one standard answer, $0.1000 for a thousand
- 131K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 5 of 89 priced models by cost per answer

FAQ:

Q: How much does Granite 4.1 8B cost?
A: $0.05 per million input tokens and $0.1 per million output tokens at list rates, which works out to $0.0001 for one standard answer of 1,000 tokens in and 500 out, or $0.1000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Granite 4.1 8B?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Granite 4.1 8B?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Granite 4.1 8B expensive compared to other models?
A: It ranks 5 of 89 priced models by cost per answer, so 84 cost more and 4 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Grok 4.3: price, context window and credit cost

URL: https://whizi.io/models/grok-4-3/
Updated: August 2026

Quick answer: Grok 4.3 costs $0.0025 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.25 per million input tokens and $2.5 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 4 credits per message on the Powerhouse plan and above.

What Grok 4.3 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Grok 4.3 is a xAI model priced at $1.25 per million input tokens and $2.5 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0025. Inside Whizi it costs 4 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 32 of the 89 priced models cost more per answer, and 56 cost less.

#### What Grok 4.3 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.25 |
| Output, per million tokens | $2.5 |
| One standard answer | $0.0025 |
| One thousand answers | $2.50 |
| Context window | 1M tokens |
| Credits per message in Whizi | 4 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 57 |

Prices are the published xAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Grok 4.3 is charged at 4 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,000 |

Grok 4.3 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Grok 4.3 costs 47 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.5 times more.

Its nearest neighbours from xAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Grok Build 0.1 | $0.0020 | 256K | 3 |
| Grok 4.3 (this page) | $0.0025 | 1M | 4 |
| Grok 4.5 | $0.0050 | 500K | 6 |

If Grok 4.3 is more model than a given task needs, Grok Build 0.1 is the cheaper xAI option at $0.0020 per answer and 3 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Grok 4.3 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0025 for one standard answer, ranking 57 of 89 priced models
- 1M token context window, well above the catalogue median
- 4 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.25 and output $2.5 per million tokens
- $0.0025 for one standard answer, $2.50 for a thousand
- 1M token context window
- 4 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 57 of 89 priced models by cost per answer

FAQ:

Q: How much does Grok 4.3 cost?
A: $1.25 per million input tokens and $2.5 per million output tokens at list rates, which works out to $0.0025 for one standard answer of 1,000 tokens in and 500 out, or $2.50 for a thousand of them. Inside Whizi it is 4 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Grok 4.3?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 4 credits per message, a Powerhouse allowance of 8,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Grok 4.3?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Grok 4.3 expensive compared to other models?
A: It ranks 57 of 89 priced models by cost per answer, so 32 cost more and 56 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Grok 4.5: price, context window and credit cost

URL: https://whizi.io/models/grok-4-5/
Updated: August 2026

Quick answer: Grok 4.5 costs $0.0050 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $6 per million output tokens. The context window is 500K tokens, around the catalogue median. Inside Whizi it costs 6 credits per message on the Powerhouse plan and above.

What Grok 4.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Grok 4.5 is a xAI model priced at $2 per million input tokens and $6 per million output tokens, with a 500K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0050. Inside Whizi it costs 6 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 21 of the 89 priced models cost more per answer, and 67 cost less.

#### What Grok 4.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $6 |
| One standard answer | $0.0050 |
| One thousand answers | $5.00 |
| Context window | 500K tokens |
| Credits per message in Whizi | 6 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 68 |

Prices are the published xAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Grok 4.5 is charged at 6 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,333 |

Grok 4.5 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Grok 4.5 costs 94 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.9 times more.

Its nearest neighbours from xAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Grok Build 0.1 | $0.0020 | 256K | 3 |
| Grok 4.3 | $0.0025 | 1M | 4 |
| Grok 4.5 (this page) | $0.0050 | 500K | 6 |

If Grok 4.5 is more model than a given task needs, Grok 4.3 is the cheaper xAI option at $0.0025 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Grok 4.5 accepts 500K tokens of context, which is roughly 750 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0050 for one standard answer, ranking 68 of 89 priced models
- 500K token context window, around the catalogue median
- 6 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $2 and output $6 per million tokens
- $0.0050 for one standard answer, $5.00 for a thousand
- 500K token context window
- 6 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 68 of 89 priced models by cost per answer

FAQ:

Q: How much does Grok 4.5 cost?
A: $2 per million input tokens and $6 per million output tokens at list rates, which works out to $0.0050 for one standard answer of 1,000 tokens in and 500 out, or $5.00 for a thousand of them. Inside Whizi it is 6 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Grok 4.5?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 6 credits per message, a Powerhouse allowance of 8,000 credits covers 1,333 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Grok 4.5?
A: 500K tokens, which is roughly 750 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Grok 4.5 expensive compared to other models?
A: It ranks 68 of 89 priced models by cost per answer, so 21 cost more and 67 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Grok 4.6: price, context window and credit cost

URL: https://whizi.io/models/grok-4-6/
Updated: August 2026

Quick answer: Grok 4.6 costs $0.0050 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $6 per million output tokens. The context window is 500K tokens, around the catalogue median. Inside Whizi it costs 6 credits per message on the Pro plan and above.

What Grok 4.6 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Grok 4.6 is a xAI model priced at $2 per million input tokens and $6 per million output tokens, with a 500K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0050. Inside Whizi it costs 6 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 18 of the 89 priced models cost more per answer, and 70 cost less.

#### What Grok 4.6 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $6 |
| One standard answer | $0.0050 |
| One thousand answers | $5.00 |
| Context window | 500K tokens |
| Credits per message in Whizi | 6 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 71 |

Prices are the published xAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Grok 4.6 is charged at 6 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 333 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,333 |

Grok 4.6 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Grok 4.6 costs 94 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.9 times more.

Its nearest neighbours from xAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Grok Build 0.1 | $0.0020 | 256K | 3 |
| Grok 4.3 | $0.0025 | 1M | 4 |
| Grok 4.6 (this page) | $0.0050 | 500K | 6 |

If Grok 4.6 is more model than a given task needs, Grok 4.3 is the cheaper xAI option at $0.0025 per answer and 4 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Grok 4.6 accepts 500K tokens of context, which is roughly 750 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0050 for one standard answer, ranking 71 of 89 priced models
- 500K token context window, around the catalogue median
- 6 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $2 and output $6 per million tokens
- $0.0050 for one standard answer, $5.00 for a thousand
- 500K token context window
- 6 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 71 of 89 priced models by cost per answer

FAQ:

Q: How much does Grok 4.6 cost?
A: $2 per million input tokens and $6 per million output tokens at list rates, which works out to $0.0050 for one standard answer of 1,000 tokens in and 500 out, or $5.00 for a thousand of them. Inside Whizi it is 6 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Grok 4.6?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 6 credits per message, a Pro allowance of 2,000 credits covers 333 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Grok 4.6?
A: 500K tokens, which is roughly 750 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Grok 4.6 expensive compared to other models?
A: It ranks 71 of 89 priced models by cost per answer, so 18 cost more and 70 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Grok Build 0.1: price, context window and credit cost

URL: https://whizi.io/models/grok-build-0-1/
Updated: August 2026

Quick answer: Grok Build 0.1 costs $0.0020 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1 per million input tokens and $2 per million output tokens. The context window is 256K tokens, below the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What Grok Build 0.1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Grok Build 0.1 is a xAI model priced at $1 per million input tokens and $2 per million output tokens, with a 256K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0020. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 39 of the 89 priced models cost more per answer, and 49 cost less.

#### What Grok Build 0.1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1 |
| Output, per million tokens | $2 |
| One standard answer | $0.0020 |
| One thousand answers | $2.00 |
| Context window | 256K tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 50 |

Prices are the published xAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Grok Build 0.1 is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

Grok Build 0.1 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Grok Build 0.1 costs 38 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.2 times more.

Its nearest neighbours from xAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Grok Build 0.1 (this page) | $0.0020 | 256K | 3 |
| Grok 4.3 | $0.0025 | 1M | 4 |
| Grok 4.5 | $0.0050 | 500K | 6 |

It is the cheapest xAI model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Grok Build 0.1 accepts 256K tokens of context, which is roughly 380 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0020 for one standard answer, ranking 50 of 89 priced models
- 256K token context window, below the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1 and output $2 per million tokens
- $0.0020 for one standard answer, $2.00 for a thousand
- 256K token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 50 of 89 priced models by cost per answer

FAQ:

Q: How much does Grok Build 0.1 cost?
A: $1 per million input tokens and $2 per million output tokens at list rates, which works out to $0.0020 for one standard answer of 1,000 tokens in and 500 out, or $2.00 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Grok Build 0.1?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Grok Build 0.1?
A: 256K tokens, which is roughly 380 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Grok Build 0.1 expensive compared to other models?
A: It ranks 50 of 89 priced models by cost per answer, so 39 cost more and 49 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Hy3: price, context window and credit cost

URL: https://whizi.io/models/hy3/
Updated: August 2026

Quick answer: Hy3 costs $0.0004 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.132 per million input tokens and $0.528 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Hy3 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Hy3 is a Tencent model priced at $0.132 per million input tokens and $0.528 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0004. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 75 of the 89 priced models cost more per answer, and 13 cost less.

#### What Hy3 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.132 |
| Output, per million tokens | $0.528 |
| One standard answer | $0.0004 |
| One thousand answers | $0.3960 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 14 |

Prices are the published Tencent rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Hy3 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Hy3 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Hy3 costs 7.5 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 4.3 times less.

It is the only Tencent model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Tencent model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Hy3 accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0004 for one standard answer, ranking 14 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.132 and output $0.528 per million tokens
- $0.0004 for one standard answer, $0.3960 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 14 of 89 priced models by cost per answer

FAQ:

Q: How much does Hy3 cost?
A: $0.132 per million input tokens and $0.528 per million output tokens at list rates, which works out to $0.0004 for one standard answer of 1,000 tokens in and 500 out, or $0.3960 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Hy3?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Hy3?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Hy3 expensive compared to other models?
A: It ranks 14 of 89 priced models by cost per answer, so 75 cost more and 13 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Inkling: price, context window and credit cost

URL: https://whizi.io/models/inkling/
Updated: August 2026

Quick answer: Inkling costs $0.0030 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.95 per million input tokens and $4.05 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 4 credits per message on the Powerhouse plan and above.

What Inkling costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Inkling is a Thinkingmachines model priced at $0.95 per million input tokens and $4.05 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0030. Inside Whizi it costs 4 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 28 of the 89 priced models cost more per answer, and 60 cost less.

#### What Inkling costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.95 |
| Output, per million tokens | $4.05 |
| One standard answer | $0.0030 |
| One thousand answers | $2.98 |
| Context window | 1M tokens |
| Credits per message in Whizi | 4 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 61 |

Prices are the published Thinkingmachines rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Inkling is charged at 4 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,000 |

Inkling is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Inkling costs 56 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.7 times more.

Its nearest neighbours from Thinkingmachines:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Inkling Small | $0.0010 | 524K | 1 |
| Inkling (this page) | $0.0030 | 1M | 4 |

If Inkling is more model than a given task needs, Inkling Small is the cheaper Thinkingmachines option at $0.0010 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Inkling accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0030 for one standard answer, ranking 61 of 89 priced models
- 1M token context window, well above the catalogue median
- 4 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.95 and output $4.05 per million tokens
- $0.0030 for one standard answer, $2.98 for a thousand
- 1M token context window
- 4 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 61 of 89 priced models by cost per answer

FAQ:

Q: How much does Inkling cost?
A: $0.95 per million input tokens and $4.05 per million output tokens at list rates, which works out to $0.0030 for one standard answer of 1,000 tokens in and 500 out, or $2.98 for a thousand of them. Inside Whizi it is 4 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Inkling?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 4 credits per message, a Powerhouse allowance of 8,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Inkling?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Inkling expensive compared to other models?
A: It ranks 61 of 89 priced models by cost per answer, so 28 cost more and 60 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Inkling Small: price, context window and credit cost

URL: https://whizi.io/models/inkling-small/
Updated: August 2026

Quick answer: Inkling Small costs $0.0010 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.45 per million input tokens and $1.2 per million output tokens. The context window is 524K tokens, around the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Inkling Small costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Inkling Small is a Thinkingmachines model priced at $0.45 per million input tokens and $1.2 per million output tokens, with a 524K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0010. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 56 of the 89 priced models cost more per answer, and 32 cost less.

#### What Inkling Small costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.45 |
| Output, per million tokens | $1.2 |
| One standard answer | $0.0010 |
| One thousand answers | $1.05 |
| Context window | 524K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 33 |

Prices are the published Thinkingmachines rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Inkling Small is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Inkling Small is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Inkling Small costs 20 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.6 times less.

Its nearest neighbours from Thinkingmachines:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Inkling Small (this page) | $0.0010 | 524K | 1 |
| Inkling | $0.0030 | 1M | 4 |

It is the cheapest Thinkingmachines model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Inkling Small accepts 524K tokens of context, which is roughly 790 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0010 for one standard answer, ranking 33 of 89 priced models
- 524K token context window, around the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.45 and output $1.2 per million tokens
- $0.0010 for one standard answer, $1.05 for a thousand
- 524K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 33 of 89 priced models by cost per answer

FAQ:

Q: How much does Inkling Small cost?
A: $0.45 per million input tokens and $1.2 per million output tokens at list rates, which works out to $0.0010 for one standard answer of 1,000 tokens in and 500 out, or $1.05 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Inkling Small?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Inkling Small?
A: 524K tokens, which is roughly 790 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Inkling Small expensive compared to other models?
A: It ranks 33 of 89 priced models by cost per answer, so 56 cost more and 32 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### KAT-Coder-Air V2.5: price, context window and credit cost

URL: https://whizi.io/models/kat-coder-air-v2-5/
Updated: August 2026

Quick answer: KAT-Coder-Air V2.5 costs $0.0004 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.15 per million input tokens and $0.6 per million output tokens. The context window is 256K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What KAT-Coder-Air V2.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

KAT-Coder-Air V2.5 is a Kwaipilot model priced at $0.15 per million input tokens and $0.6 per million output tokens, with a 256K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0004. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 74 of the 89 priced models cost more per answer, and 14 cost less.

#### What KAT-Coder-Air V2.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.15 |
| Output, per million tokens | $0.6 |
| One standard answer | $0.0004 |
| One thousand answers | $0.4500 |
| Context window | 256K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 15 |

Prices are the published Kwaipilot rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and KAT-Coder-Air V2.5 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

KAT-Coder-Air V2.5 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, KAT-Coder-Air V2.5 costs 8.5 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 3.8 times less.

Its nearest neighbours from Kwaipilot:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| KAT-Coder-Air V2.5 (this page) | $0.0004 | 256K | 1 |
| KAT-Coder-Pro V2.5 | $0.0022 | 256K | 3 |

It is the cheapest Kwaipilot model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

KAT-Coder-Air V2.5 accepts 256K tokens of context, which is roughly 380 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0004 for one standard answer, ranking 15 of 89 priced models
- 256K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.15 and output $0.6 per million tokens
- $0.0004 for one standard answer, $0.4500 for a thousand
- 256K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 15 of 89 priced models by cost per answer

FAQ:

Q: How much does KAT-Coder-Air V2.5 cost?
A: $0.15 per million input tokens and $0.6 per million output tokens at list rates, which works out to $0.0004 for one standard answer of 1,000 tokens in and 500 out, or $0.4500 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes KAT-Coder-Air V2.5?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of KAT-Coder-Air V2.5?
A: 256K tokens, which is roughly 380 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is KAT-Coder-Air V2.5 expensive compared to other models?
A: It ranks 15 of 89 priced models by cost per answer, so 74 cost more and 14 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### KAT-Coder-Pro V2.5: price, context window and credit cost

URL: https://whizi.io/models/kat-coder-pro-v2-5/
Updated: August 2026

Quick answer: KAT-Coder-Pro V2.5 costs $0.0022 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.74 per million input tokens and $2.96 per million output tokens. The context window is 256K tokens, below the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What KAT-Coder-Pro V2.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

KAT-Coder-Pro V2.5 is a Kwaipilot model priced at $0.74 per million input tokens and $2.96 per million output tokens, with a 256K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0022. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 37 of the 89 priced models cost more per answer, and 51 cost less.

#### What KAT-Coder-Pro V2.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.74 |
| Output, per million tokens | $2.96 |
| One standard answer | $0.0022 |
| One thousand answers | $2.22 |
| Context window | 256K tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 52 |

Prices are the published Kwaipilot rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and KAT-Coder-Pro V2.5 is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

KAT-Coder-Pro V2.5 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, KAT-Coder-Pro V2.5 costs 42 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.3 times more.

Its nearest neighbours from Kwaipilot:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| KAT-Coder-Air V2.5 | $0.0004 | 256K | 1 |
| KAT-Coder-Pro V2.5 (this page) | $0.0022 | 256K | 3 |

If KAT-Coder-Pro V2.5 is more model than a given task needs, KAT-Coder-Air V2.5 is the cheaper Kwaipilot option at $0.0004 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

KAT-Coder-Pro V2.5 accepts 256K tokens of context, which is roughly 380 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0022 for one standard answer, ranking 52 of 89 priced models
- 256K token context window, below the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.74 and output $2.96 per million tokens
- $0.0022 for one standard answer, $2.22 for a thousand
- 256K token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 52 of 89 priced models by cost per answer

FAQ:

Q: How much does KAT-Coder-Pro V2.5 cost?
A: $0.74 per million input tokens and $2.96 per million output tokens at list rates, which works out to $0.0022 for one standard answer of 1,000 tokens in and 500 out, or $2.22 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes KAT-Coder-Pro V2.5?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of KAT-Coder-Pro V2.5?
A: 256K tokens, which is roughly 380 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is KAT-Coder-Pro V2.5 expensive compared to other models?
A: It ranks 52 of 89 priced models by cost per answer, so 37 cost more and 51 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Kimi K2 0711: price, context window and credit cost

URL: https://whizi.io/models/kimi-k2/
Updated: August 2026

Quick answer: Kimi K2 0711 costs $0.0017 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.57 per million input tokens and $2.3 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 10 credits per message on the Powerhouse plan and above.

What Kimi K2 0711 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Kimi K2 0711 is a Moonshot model priced at $0.57 per million input tokens and $2.3 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0017. Inside Whizi it costs 10 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 44 of the 89 priced models cost more per answer, and 44 cost less.

#### What Kimi K2 0711 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.57 |
| Output, per million tokens | $2.3 |
| One standard answer | $0.0017 |
| One thousand answers | $1.72 |
| Context window | 131K tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 45 |

Prices are the published Moonshot rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Kimi K2 0711 is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Kimi K2 0711 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Kimi K2 0711 costs 32 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Moonshot:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Kimi K2 0711 (this page) | $0.0017 | 131K | 10 |
| Kimi K2 Thinking | $0.0019 | 262K | 10 |
| Kimi K2.7 Code | $0.0025 | 262K | 10 |
| Kimi K3 | $0.0105 | 1M | 10 |

It is the cheapest Moonshot model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Kimi K2 0711 accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0017 for one standard answer, ranking 45 of 89 priced models
- 131K token context window, below the catalogue median
- 10 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.57 and output $2.3 per million tokens
- $0.0017 for one standard answer, $1.72 for a thousand
- 131K token context window
- 10 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 45 of 89 priced models by cost per answer

FAQ:

Q: How much does Kimi K2 0711 cost?
A: $0.57 per million input tokens and $2.3 per million output tokens at list rates, which works out to $0.0017 for one standard answer of 1,000 tokens in and 500 out, or $1.72 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Kimi K2 0711?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 10 credits per message, a Powerhouse allowance of 8,000 credits covers 800 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Kimi K2 0711?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Kimi K2 0711 expensive compared to other models?
A: It ranks 45 of 89 priced models by cost per answer, so 44 cost more and 44 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Kimi K2.7 Code: price, context window and credit cost

URL: https://whizi.io/models/kimi-k2-7-code/
Updated: August 2026

Quick answer: Kimi K2.7 Code costs $0.0025 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.71 per million input tokens and $3.5 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 10 credits per message on the Powerhouse plan and above.

What Kimi K2.7 Code costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Kimi K2.7 Code is a Moonshot model priced at $0.71 per million input tokens and $3.5 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0025. Inside Whizi it costs 10 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 34 of the 89 priced models cost more per answer, and 54 cost less.

#### What Kimi K2.7 Code costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.71 |
| Output, per million tokens | $3.5 |
| One standard answer | $0.0025 |
| One thousand answers | $2.46 |
| Context window | 262K tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 55 |

Prices are the published Moonshot rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Kimi K2.7 Code is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Kimi K2.7 Code is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Kimi K2.7 Code costs 46 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.4 times more.

Its nearest neighbours from Moonshot:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Kimi K2 0711 | $0.0017 | 131K | 10 |
| Kimi K2 Thinking | $0.0019 | 262K | 10 |
| Kimi K2.7 Code (this page) | $0.0025 | 262K | 10 |
| Kimi K3 | $0.0105 | 1M | 10 |

If Kimi K2.7 Code is more model than a given task needs, Kimi K2 Thinking is the cheaper Moonshot option at $0.0019 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Kimi K2.7 Code accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0025 for one standard answer, ranking 55 of 89 priced models
- 262K token context window, below the catalogue median
- 10 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.71 and output $3.5 per million tokens
- $0.0025 for one standard answer, $2.46 for a thousand
- 262K token context window
- 10 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 55 of 89 priced models by cost per answer

FAQ:

Q: How much does Kimi K2.7 Code cost?
A: $0.71 per million input tokens and $3.5 per million output tokens at list rates, which works out to $0.0025 for one standard answer of 1,000 tokens in and 500 out, or $2.46 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Kimi K2.7 Code?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 10 credits per message, a Powerhouse allowance of 8,000 credits covers 800 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Kimi K2.7 Code?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Kimi K2.7 Code expensive compared to other models?
A: It ranks 55 of 89 priced models by cost per answer, so 34 cost more and 54 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Kimi K2 Thinking: price, context window and credit cost

URL: https://whizi.io/models/kimi-k2-thinking/
Updated: August 2026

Quick answer: Kimi K2 Thinking costs $0.0019 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.6 per million input tokens and $2.5 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 10 credits per message on the Powerhouse plan and above.

What Kimi K2 Thinking costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Kimi K2 Thinking is a Moonshot model priced at $0.6 per million input tokens and $2.5 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0019. Inside Whizi it costs 10 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 42 of the 89 priced models cost more per answer, and 46 cost less.

#### What Kimi K2 Thinking costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.6 |
| Output, per million tokens | $2.5 |
| One standard answer | $0.0019 |
| One thousand answers | $1.85 |
| Context window | 262K tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 47 |

Prices are the published Moonshot rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Kimi K2 Thinking is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Kimi K2 Thinking is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Kimi K2 Thinking costs 35 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Moonshot:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Kimi K2 0711 | $0.0017 | 131K | 10 |
| Kimi K2 Thinking (this page) | $0.0019 | 262K | 10 |
| Kimi K2.7 Code | $0.0025 | 262K | 10 |
| Kimi K3 | $0.0105 | 1M | 10 |

If Kimi K2 Thinking is more model than a given task needs, Kimi K2 0711 is the cheaper Moonshot option at $0.0017 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Kimi K2 Thinking accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0019 for one standard answer, ranking 47 of 89 priced models
- 262K token context window, below the catalogue median
- 10 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.6 and output $2.5 per million tokens
- $0.0019 for one standard answer, $1.85 for a thousand
- 262K token context window
- 10 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 47 of 89 priced models by cost per answer

FAQ:

Q: How much does Kimi K2 Thinking cost?
A: $0.6 per million input tokens and $2.5 per million output tokens at list rates, which works out to $0.0019 for one standard answer of 1,000 tokens in and 500 out, or $1.85 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Kimi K2 Thinking?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 10 credits per message, a Powerhouse allowance of 8,000 credits covers 800 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Kimi K2 Thinking?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Kimi K2 Thinking expensive compared to other models?
A: It ranks 47 of 89 priced models by cost per answer, so 42 cost more and 46 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Kimi K3: price, context window and credit cost

URL: https://whizi.io/models/kimi-k3/
Updated: August 2026

Quick answer: Kimi K3 costs $0.0105 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $3 per million input tokens and $15 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 10 credits per message on the Pro plan and above.

What Kimi K3 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Kimi K3 is a Moonshot model priced at $3 per million input tokens and $15 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0105. Inside Whizi it costs 10 credits per message and needs the Pro plan or above.

That makes it one of the more expensive models in the catalogue: 6 of the 89 priced models cost more per answer, and 82 cost less.

#### What Kimi K3 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $3 |
| Output, per million tokens | $15 |
| One standard answer | $0.0105 |
| One thousand answers | $10.50 |
| Context window | 1M tokens |
| Credits per message in Whizi | 10 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 83 |

Prices are the published Moonshot rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Kimi K3 is charged at 10 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 200 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 800 |

Kimi K3 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Kimi K3 costs 198 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 6.1 times more.

Its nearest neighbours from Moonshot:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Kimi K2 0711 | $0.0017 | 131K | 10 |
| Kimi K2 Thinking | $0.0019 | 262K | 10 |
| Kimi K2.7 Code | $0.0025 | 262K | 10 |
| Kimi K3 (this page) | $0.0105 | 1M | 10 |

If Kimi K3 is more model than a given task needs, Kimi K2.7 Code is the cheaper Moonshot option at $0.0025 per answer and 10 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Kimi K3 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0105 for one standard answer, ranking 83 of 89 priced models
- 1M token context window, well above the catalogue median
- 10 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $3 and output $15 per million tokens
- $0.0105 for one standard answer, $10.50 for a thousand
- 1M token context window
- 10 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 83 of 89 priced models by cost per answer

FAQ:

Q: How much does Kimi K3 cost?
A: $3 per million input tokens and $15 per million output tokens at list rates, which works out to $0.0105 for one standard answer of 1,000 tokens in and 500 out, or $10.50 for a thousand of them. Inside Whizi it is 10 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Kimi K3?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 10 credits per message, a Pro allowance of 2,000 credits covers 200 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Kimi K3?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Kimi K3 expensive compared to other models?
A: It ranks 83 of 89 priced models by cost per answer, so 6 cost more and 82 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Laguna S 2.1: price, context window and credit cost

URL: https://whizi.io/models/laguna-s-2-1/
Updated: August 2026

Quick answer: Laguna S 2.1 costs $0.0002 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.09 per million input tokens and $0.18 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Laguna S 2.1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Laguna S 2.1 is a Poolside model priced at $0.09 per million input tokens and $0.18 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0002. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 80 of the 89 priced models cost more per answer, and 8 cost less.

#### What Laguna S 2.1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.09 |
| Output, per million tokens | $0.18 |
| One standard answer | $0.0002 |
| One thousand answers | $0.1800 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 9 |

Prices are the published Poolside rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Laguna S 2.1 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Laguna S 2.1 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Laguna S 2.1 costs 3.4 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 9.6 times less.

Its nearest neighbours from Poolside:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Laguna XS 2.1 | $0.0001 | 262K | 1 |
| Laguna S 2.1 (this page) | $0.0002 | 1M | 1 |

If Laguna S 2.1 is more model than a given task needs, Laguna XS 2.1 is the cheaper Poolside option at $0.0001 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Laguna S 2.1 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0002 for one standard answer, ranking 9 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.09 and output $0.18 per million tokens
- $0.0002 for one standard answer, $0.1800 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 9 of 89 priced models by cost per answer

FAQ:

Q: How much does Laguna S 2.1 cost?
A: $0.09 per million input tokens and $0.18 per million output tokens at list rates, which works out to $0.0002 for one standard answer of 1,000 tokens in and 500 out, or $0.1800 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Laguna S 2.1?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Laguna S 2.1?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Laguna S 2.1 expensive compared to other models?
A: It ranks 9 of 89 priced models by cost per answer, so 80 cost more and 8 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Laguna XS 2.1: price, context window and credit cost

URL: https://whizi.io/models/laguna-xs-2-1/
Updated: August 2026

Quick answer: Laguna XS 2.1 costs $0.0001 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.06 per million input tokens and $0.12 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Laguna XS 2.1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Laguna XS 2.1 is a Poolside model priced at $0.06 per million input tokens and $0.12 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0001. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 83 of the 89 priced models cost more per answer, and 5 cost less.

#### What Laguna XS 2.1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.06 |
| Output, per million tokens | $0.12 |
| One standard answer | $0.0001 |
| One thousand answers | $0.1200 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 6 |

Prices are the published Poolside rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Laguna XS 2.1 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Laguna XS 2.1 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Laguna XS 2.1 costs 2.3 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 14.3 times less.

Its nearest neighbours from Poolside:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Laguna XS 2.1 (this page) | $0.0001 | 262K | 1 |
| Laguna S 2.1 | $0.0002 | 1M | 1 |

It is the cheapest Poolside model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Laguna XS 2.1 accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0001 for one standard answer, ranking 6 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.06 and output $0.12 per million tokens
- $0.0001 for one standard answer, $0.1200 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 6 of 89 priced models by cost per answer

FAQ:

Q: How much does Laguna XS 2.1 cost?
A: $0.06 per million input tokens and $0.12 per million output tokens at list rates, which works out to $0.0001 for one standard answer of 1,000 tokens in and 500 out, or $0.1200 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Laguna XS 2.1?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Laguna XS 2.1?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Laguna XS 2.1 expensive compared to other models?
A: It ranks 6 of 89 priced models by cost per answer, so 83 cost more and 5 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Ling-3.0-flash: price, context window and credit cost

URL: https://whizi.io/models/ling-3-0-flash/
Updated: August 2026

Quick answer: Ling-3.0-flash costs $0.000053 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.021 per million input tokens and $0.063 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Ling-3.0-flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Ling-3.0-flash is an inclusionAI model priced at $0.021 per million input tokens and $0.063 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.000053. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 88 of the 89 priced models cost more per answer, and 0 cost less.

#### What Ling-3.0-flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.021 |
| Output, per million tokens | $0.063 |
| One standard answer | $0.000053 |
| One thousand answers | $0.0525 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 1 |

Prices are the published inclusionAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Ling-3.0-flash is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Ling-3.0-flash is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Ling-3.0-flash costs about the same. Against the median, Kimi K2 0711 at $0.0017, it costs 32.5 times less.

Its nearest neighbours from inclusionAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Ling-3.0-flash (this page) | $0.000053 | 262K | 1 |
| Ring-2.6-1T | $0.0004 | 262K | 1 |

It is the cheapest inclusionAI model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Ling-3.0-flash accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.000053 for one standard answer, ranking 1 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.021 and output $0.063 per million tokens
- $0.000053 for one standard answer, $0.0525 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 1 of 89 priced models by cost per answer

FAQ:

Q: How much does Ling-3.0-flash cost?
A: $0.021 per million input tokens and $0.063 per million output tokens at list rates, which works out to $0.000053 for one standard answer of 1,000 tokens in and 500 out, or $0.0525 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Ling-3.0-flash?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Ling-3.0-flash?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Ling-3.0-flash expensive compared to other models?
A: It ranks 1 of 89 priced models by cost per answer, so 88 cost more and 0 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Llama 3.3 70B Instruct: price, context window and credit cost

URL: https://whizi.io/models/llama-3-3-70b-instruct/
Updated: August 2026

Quick answer: Llama 3.3 70B Instruct costs $0.0003 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.1 per million input tokens and $0.32 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Pro plan and above.

What Llama 3.3 70B Instruct costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Llama 3.3 70B Instruct is a Meta model priced at $0.1 per million input tokens and $0.32 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0003. Inside Whizi it costs 1 credit per message and needs the Pro plan or above.

That makes it one of the cheaper models in the catalogue: 79 of the 89 priced models cost more per answer, and 9 cost less.

#### What Llama 3.3 70B Instruct costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.1 |
| Output, per million tokens | $0.32 |
| One standard answer | $0.0003 |
| One thousand answers | $0.2600 |
| Context window | 131K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 10 |

Prices are the published Meta rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Llama 3.3 70B Instruct is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Llama 3.3 70B Instruct is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Llama 3.3 70B Instruct costs 4.9 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 6.6 times less.

Its nearest neighbours from Meta:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Llama 3.3 70B Instruct (this page) | $0.0003 | 131K | 1 |
| Llama 4 Maverick | $0.0006 | 1M | 1 |
| Muse Glimmer 30B | $0.0011 | 131K | 2 |
| Muse Spark 1.2 | $0.0034 | 1M | 5 |

It is the cheapest Meta model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Llama 3.3 70B Instruct accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0003 for one standard answer, ranking 10 of 89 priced models
- 131K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.1 and output $0.32 per million tokens
- $0.0003 for one standard answer, $0.2600 for a thousand
- 131K token context window
- 1 credit per message inside Whizi
- Requires the Pro plan or above
- Ranks 10 of 89 priced models by cost per answer

FAQ:

Q: How much does Llama 3.3 70B Instruct cost?
A: $0.1 per million input tokens and $0.32 per million output tokens at list rates, which works out to $0.0003 for one standard answer of 1,000 tokens in and 500 out, or $0.2600 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Llama 3.3 70B Instruct?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 1 credits per message, a Pro allowance of 2,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Llama 3.3 70B Instruct?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Llama 3.3 70B Instruct expensive compared to other models?
A: It ranks 10 of 89 priced models by cost per answer, so 79 cost more and 9 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Llama 4 Maverick: price, context window and credit cost

URL: https://whizi.io/models/llama-4-maverick/
Updated: August 2026

Quick answer: Llama 4 Maverick costs $0.0006 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.2 per million input tokens and $0.8 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Pro plan and above.

What Llama 4 Maverick costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Llama 4 Maverick is a Meta model priced at $0.2 per million input tokens and $0.8 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0006. Inside Whizi it costs 1 credit per message and needs the Pro plan or above.

That makes it one of the cheaper models in the catalogue: 71 of the 89 priced models cost more per answer, and 17 cost less.

#### What Llama 4 Maverick costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.2 |
| Output, per million tokens | $0.8 |
| One standard answer | $0.0006 |
| One thousand answers | $0.6000 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 18 |

Prices are the published Meta rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Llama 4 Maverick is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Llama 4 Maverick is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Llama 4 Maverick costs 11 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.9 times less.

Its nearest neighbours from Meta:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Llama 3.3 70B Instruct | $0.0003 | 131K | 1 |
| Llama 4 Maverick (this page) | $0.0006 | 1M | 1 |
| Muse Glimmer 30B | $0.0011 | 131K | 2 |
| Muse Spark 1.2 | $0.0034 | 1M | 5 |

If Llama 4 Maverick is more model than a given task needs, Llama 3.3 70B Instruct is the cheaper Meta option at $0.0003 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Llama 4 Maverick accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0006 for one standard answer, ranking 18 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.2 and output $0.8 per million tokens
- $0.0006 for one standard answer, $0.6000 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Pro plan or above
- Ranks 18 of 89 priced models by cost per answer

FAQ:

Q: How much does Llama 4 Maverick cost?
A: $0.2 per million input tokens and $0.8 per million output tokens at list rates, which works out to $0.0006 for one standard answer of 1,000 tokens in and 500 out, or $0.6000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Llama 4 Maverick?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 1 credits per message, a Pro allowance of 2,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Llama 4 Maverick?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Llama 4 Maverick expensive compared to other models?
A: It ranks 18 of 89 priced models by cost per answer, so 71 cost more and 17 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### LongCat 2.0: price, context window and credit cost

URL: https://whizi.io/models/longcat-2-0/
Updated: August 2026

Quick answer: LongCat 2.0 costs $0.0009 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.3 per million input tokens and $1.2 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What LongCat 2.0 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

LongCat 2.0 is a Meituan model priced at $0.3 per million input tokens and $1.2 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0009. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 60 of the 89 priced models cost more per answer, and 28 cost less.

#### What LongCat 2.0 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.3 |
| Output, per million tokens | $1.2 |
| One standard answer | $0.0009 |
| One thousand answers | $0.9000 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 29 |

Prices are the published Meituan rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and LongCat 2.0 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

LongCat 2.0 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, LongCat 2.0 costs 17 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.9 times less.

It is the only Meituan model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Meituan model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

LongCat 2.0 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0009 for one standard answer, ranking 29 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.3 and output $1.2 per million tokens
- $0.0009 for one standard answer, $0.9000 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 29 of 89 priced models by cost per answer

FAQ:

Q: How much does LongCat 2.0 cost?
A: $0.3 per million input tokens and $1.2 per million output tokens at list rates, which works out to $0.0009 for one standard answer of 1,000 tokens in and 500 out, or $0.9000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes LongCat 2.0?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of LongCat 2.0?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is LongCat 2.0 expensive compared to other models?
A: It ranks 29 of 89 priced models by cost per answer, so 60 cost more and 28 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### MiniMax M2: price, context window and credit cost

URL: https://whizi.io/models/minimax-m2/
Updated: August 2026

Quick answer: MiniMax M2 costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.255 per million input tokens and $1.02 per million output tokens. The context window is 205K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What MiniMax M2 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

MiniMax M2 is a MiniMax model priced at $0.255 per million input tokens and $1.02 per million output tokens, with a 205K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 65 of the 89 priced models cost more per answer, and 23 cost less.

#### What MiniMax M2 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.255 |
| Output, per million tokens | $1.02 |
| One standard answer | $0.0008 |
| One thousand answers | $0.7650 |
| Context window | 205K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 24 |

Prices are the published MiniMax rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and MiniMax M2 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

MiniMax M2 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, MiniMax M2 costs 14 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.2 times less.

Its nearest neighbours from MiniMax:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| MiniMax M2 (this page) | $0.0008 | 205K | 1 |
| MiniMax M3 | $0.0009 | 1M | 1 |

It is the cheapest MiniMax model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

MiniMax M2 accepts 205K tokens of context, which is roughly 310 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 24 of 89 priced models
- 205K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.255 and output $1.02 per million tokens
- $0.0008 for one standard answer, $0.7650 for a thousand
- 205K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 24 of 89 priced models by cost per answer

FAQ:

Q: How much does MiniMax M2 cost?
A: $0.255 per million input tokens and $1.02 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.7650 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes MiniMax M2?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of MiniMax M2?
A: 205K tokens, which is roughly 310 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is MiniMax M2 expensive compared to other models?
A: It ranks 24 of 89 priced models by cost per answer, so 65 cost more and 23 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### MiniMax M3: price, context window and credit cost

URL: https://whizi.io/models/minimax-m3/
Updated: August 2026

Quick answer: MiniMax M3 costs $0.0009 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.3 per million input tokens and $1.2 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What MiniMax M3 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

MiniMax M3 is a MiniMax model priced at $0.3 per million input tokens and $1.2 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0009. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 61 of the 89 priced models cost more per answer, and 27 cost less.

#### What MiniMax M3 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.3 |
| Output, per million tokens | $1.2 |
| One standard answer | $0.0009 |
| One thousand answers | $0.9000 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 28 |

Prices are the published MiniMax rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and MiniMax M3 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

MiniMax M3 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, MiniMax M3 costs 17 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.9 times less.

Its nearest neighbours from MiniMax:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| MiniMax M2 | $0.0008 | 205K | 1 |
| MiniMax M3 (this page) | $0.0009 | 1M | 1 |

If MiniMax M3 is more model than a given task needs, MiniMax M2 is the cheaper MiniMax option at $0.0008 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

MiniMax M3 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0009 for one standard answer, ranking 28 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.3 and output $1.2 per million tokens
- $0.0009 for one standard answer, $0.9000 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 28 of 89 priced models by cost per answer

FAQ:

Q: How much does MiniMax M3 cost?
A: $0.3 per million input tokens and $1.2 per million output tokens at list rates, which works out to $0.0009 for one standard answer of 1,000 tokens in and 500 out, or $0.9000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes MiniMax M3?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of MiniMax M3?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is MiniMax M3 expensive compared to other models?
A: It ranks 28 of 89 priced models by cost per answer, so 61 cost more and 27 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Mistral Large 3 2512: price, context window and credit cost

URL: https://whizi.io/models/mistral-large-2512/
Updated: August 2026

Quick answer: Mistral Large 3 2512 costs $0.0013 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.5 per million input tokens and $1.5 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Pro plan and above.

What Mistral Large 3 2512 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Mistral Large 3 2512 is a Mistral model priced at $0.5 per million input tokens and $1.5 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0013. Inside Whizi it costs 2 credits per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 53 of the 89 priced models cost more per answer, and 35 cost less.

#### What Mistral Large 3 2512 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.5 |
| Output, per million tokens | $1.5 |
| One standard answer | $0.0013 |
| One thousand answers | $1.25 |
| Context window | 262K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 36 |

Prices are the published Mistral rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Mistral Large 3 2512 is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 1,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Mistral Large 3 2512 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Mistral Large 3 2512 costs 24 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.4 times less.

Its nearest neighbours from Mistral:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Codestral 2508 | $0.0008 | 256K | 1 |
| Mistral Large 3 2512 (this page) | $0.0013 | 262K | 2 |
| Mistral Medium 3.1 | $0.0014 | 131K | 2 |
| Mistral Medium 3.5 | $0.0053 | 262K | 6 |

If Mistral Large 3 2512 is more model than a given task needs, Codestral 2508 is the cheaper Mistral option at $0.0008 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Mistral Large 3 2512 accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0013 for one standard answer, ranking 36 of 89 priced models
- 262K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.5 and output $1.5 per million tokens
- $0.0013 for one standard answer, $1.25 for a thousand
- 262K token context window
- 2 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 36 of 89 priced models by cost per answer

FAQ:

Q: How much does Mistral Large 3 2512 cost?
A: $0.5 per million input tokens and $1.5 per million output tokens at list rates, which works out to $0.0013 for one standard answer of 1,000 tokens in and 500 out, or $1.25 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Mistral Large 3 2512?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 2 credits per message, a Pro allowance of 2,000 credits covers 1,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Mistral Large 3 2512?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Mistral Large 3 2512 expensive compared to other models?
A: It ranks 36 of 89 priced models by cost per answer, so 53 cost more and 35 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Mistral Medium 3.1: price, context window and credit cost

URL: https://whizi.io/models/mistral-medium-3-1/
Updated: August 2026

Quick answer: Mistral Medium 3.1 costs $0.0014 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.4 per million input tokens and $2 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Pro plan and above.

What Mistral Medium 3.1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Mistral Medium 3.1 is a Mistral model priced at $0.4 per million input tokens and $2 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0014. Inside Whizi it costs 2 credits per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 50 of the 89 priced models cost more per answer, and 38 cost less.

#### What Mistral Medium 3.1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.4 |
| Output, per million tokens | $2 |
| One standard answer | $0.0014 |
| One thousand answers | $1.40 |
| Context window | 131K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 39 |

Prices are the published Mistral rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Mistral Medium 3.1 is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 1,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Mistral Medium 3.1 is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Mistral Medium 3.1 costs 26 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.2 times less.

Its nearest neighbours from Mistral:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Codestral 2508 | $0.0008 | 256K | 1 |
| Mistral Large 3 2512 | $0.0013 | 262K | 2 |
| Mistral Medium 3.1 (this page) | $0.0014 | 131K | 2 |
| Mistral Medium 3.5 | $0.0053 | 262K | 6 |

If Mistral Medium 3.1 is more model than a given task needs, Mistral Large 3 2512 is the cheaper Mistral option at $0.0013 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Mistral Medium 3.1 accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0014 for one standard answer, ranking 39 of 89 priced models
- 131K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.4 and output $2 per million tokens
- $0.0014 for one standard answer, $1.40 for a thousand
- 131K token context window
- 2 credits per message inside Whizi
- Requires the Pro plan or above
- Ranks 39 of 89 priced models by cost per answer

FAQ:

Q: How much does Mistral Medium 3.1 cost?
A: $0.4 per million input tokens and $2 per million output tokens at list rates, which works out to $0.0014 for one standard answer of 1,000 tokens in and 500 out, or $1.40 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Mistral Medium 3.1?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 2 credits per message, a Pro allowance of 2,000 credits covers 1,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Mistral Medium 3.1?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Mistral Medium 3.1 expensive compared to other models?
A: It ranks 39 of 89 priced models by cost per answer, so 50 cost more and 38 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Mistral Medium 3.5: price, context window and credit cost

URL: https://whizi.io/models/mistral-medium-3-5/
Updated: August 2026

Quick answer: Mistral Medium 3.5 costs $0.0053 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.5 per million input tokens and $7.5 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 6 credits per message on the Powerhouse plan and above.

What Mistral Medium 3.5 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Mistral Medium 3.5 is a Mistral model priced at $1.5 per million input tokens and $7.5 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0053. Inside Whizi it costs 6 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 17 of the 89 priced models cost more per answer, and 71 cost less.

#### What Mistral Medium 3.5 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.5 |
| Output, per million tokens | $7.5 |
| One standard answer | $0.0053 |
| One thousand answers | $5.25 |
| Context window | 262K tokens |
| Credits per message in Whizi | 6 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 72 |

Prices are the published Mistral rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Mistral Medium 3.5 is charged at 6 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,333 |

Mistral Medium 3.5 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Mistral Medium 3.5 costs 99 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 3.1 times more.

Its nearest neighbours from Mistral:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Codestral 2508 | $0.0008 | 256K | 1 |
| Mistral Large 3 2512 | $0.0013 | 262K | 2 |
| Mistral Medium 3.1 | $0.0014 | 131K | 2 |
| Mistral Medium 3.5 (this page) | $0.0053 | 262K | 6 |

If Mistral Medium 3.5 is more model than a given task needs, Mistral Medium 3.1 is the cheaper Mistral option at $0.0014 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Mistral Medium 3.5 accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0053 for one standard answer, ranking 72 of 89 priced models
- 262K token context window, below the catalogue median
- 6 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.5 and output $7.5 per million tokens
- $0.0053 for one standard answer, $5.25 for a thousand
- 262K token context window
- 6 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 72 of 89 priced models by cost per answer

FAQ:

Q: How much does Mistral Medium 3.5 cost?
A: $1.5 per million input tokens and $7.5 per million output tokens at list rates, which works out to $0.0053 for one standard answer of 1,000 tokens in and 500 out, or $5.25 for a thousand of them. Inside Whizi it is 6 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Mistral Medium 3.5?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 6 credits per message, a Powerhouse allowance of 8,000 credits covers 1,333 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Mistral Medium 3.5?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Mistral Medium 3.5 expensive compared to other models?
A: It ranks 72 of 89 priced models by cost per answer, so 17 cost more and 71 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Muse Glimmer 30B: price, context window and credit cost

URL: https://whizi.io/models/muse-glimmer-30b/
Updated: August 2026

Quick answer: Muse Glimmer 30B costs $0.0011 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.35 per million input tokens and $1.5 per million output tokens. The context window is 131K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What Muse Glimmer 30B costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Muse Glimmer 30B is a Meta model priced at $0.35 per million input tokens and $1.5 per million output tokens, with a 131K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0011. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 55 of the 89 priced models cost more per answer, and 33 cost less.

#### What Muse Glimmer 30B costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.35 |
| Output, per million tokens | $1.5 |
| One standard answer | $0.0011 |
| One thousand answers | $1.10 |
| Context window | 131K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 34 |

Prices are the published Meta rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Muse Glimmer 30B is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Muse Glimmer 30B is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Muse Glimmer 30B costs 21 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.6 times less.

Its nearest neighbours from Meta:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Llama 3.3 70B Instruct | $0.0003 | 131K | 1 |
| Llama 4 Maverick | $0.0006 | 1M | 1 |
| Muse Glimmer 30B (this page) | $0.0011 | 131K | 2 |
| Muse Spark 1.2 | $0.0034 | 1M | 5 |

If Muse Glimmer 30B is more model than a given task needs, Llama 4 Maverick is the cheaper Meta option at $0.0006 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Muse Glimmer 30B accepts 131K tokens of context, which is roughly 200 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0011 for one standard answer, ranking 34 of 89 priced models
- 131K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.35 and output $1.5 per million tokens
- $0.0011 for one standard answer, $1.10 for a thousand
- 131K token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 34 of 89 priced models by cost per answer

FAQ:

Q: How much does Muse Glimmer 30B cost?
A: $0.35 per million input tokens and $1.5 per million output tokens at list rates, which works out to $0.0011 for one standard answer of 1,000 tokens in and 500 out, or $1.10 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Muse Glimmer 30B?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Muse Glimmer 30B?
A: 131K tokens, which is roughly 200 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Muse Glimmer 30B expensive compared to other models?
A: It ranks 34 of 89 priced models by cost per answer, so 55 cost more and 33 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Muse Spark 1.1: price, context window and credit cost

URL: https://whizi.io/models/muse-spark-1-1/
Updated: August 2026

Quick answer: Muse Spark 1.1 costs $0.0034 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.25 per million input tokens and $4.25 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 5 credits per message on the Powerhouse plan and above.

What Muse Spark 1.1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Muse Spark 1.1 is a Meta model priced at $1.25 per million input tokens and $4.25 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0034. Inside Whizi it costs 5 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 25 of the 89 priced models cost more per answer, and 63 cost less.

#### What Muse Spark 1.1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.25 |
| Output, per million tokens | $4.25 |
| One standard answer | $0.0034 |
| One thousand answers | $3.38 |
| Context window | 1M tokens |
| Credits per message in Whizi | 5 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 64 |

Prices are the published Meta rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Muse Spark 1.1 is charged at 5 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,600 |

Muse Spark 1.1 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Muse Spark 1.1 costs 64 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.0 times more.

Its nearest neighbours from Meta:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Llama 3.3 70B Instruct | $0.0003 | 131K | 1 |
| Llama 4 Maverick | $0.0006 | 1M | 1 |
| Muse Glimmer 30B | $0.0011 | 131K | 2 |
| Muse Spark 1.1 (this page) | $0.0034 | 1M | 5 |

If Muse Spark 1.1 is more model than a given task needs, Muse Glimmer 30B is the cheaper Meta option at $0.0011 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Muse Spark 1.1 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0034 for one standard answer, ranking 64 of 89 priced models
- 1M token context window, well above the catalogue median
- 5 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.25 and output $4.25 per million tokens
- $0.0034 for one standard answer, $3.38 for a thousand
- 1M token context window
- 5 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 64 of 89 priced models by cost per answer

FAQ:

Q: How much does Muse Spark 1.1 cost?
A: $1.25 per million input tokens and $4.25 per million output tokens at list rates, which works out to $0.0034 for one standard answer of 1,000 tokens in and 500 out, or $3.38 for a thousand of them. Inside Whizi it is 5 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Muse Spark 1.1?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 5 credits per message, a Powerhouse allowance of 8,000 credits covers 1,600 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Muse Spark 1.1?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Muse Spark 1.1 expensive compared to other models?
A: It ranks 64 of 89 priced models by cost per answer, so 25 cost more and 63 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Muse Spark 1.2: price, context window and credit cost

URL: https://whizi.io/models/muse-spark-1-2/
Updated: August 2026

Quick answer: Muse Spark 1.2 costs $0.0034 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.25 per million input tokens and $4.25 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 5 credits per message on the Powerhouse plan and above.

What Muse Spark 1.2 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Muse Spark 1.2 is a Meta model priced at $1.25 per million input tokens and $4.25 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0034. Inside Whizi it costs 5 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 26 of the 89 priced models cost more per answer, and 62 cost less.

#### What Muse Spark 1.2 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.25 |
| Output, per million tokens | $4.25 |
| One standard answer | $0.0034 |
| One thousand answers | $3.38 |
| Context window | 1M tokens |
| Credits per message in Whizi | 5 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 63 |

Prices are the published Meta rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Muse Spark 1.2 is charged at 5 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,600 |

Muse Spark 1.2 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Muse Spark 1.2 costs 64 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.0 times more.

Its nearest neighbours from Meta:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Llama 3.3 70B Instruct | $0.0003 | 131K | 1 |
| Llama 4 Maverick | $0.0006 | 1M | 1 |
| Muse Glimmer 30B | $0.0011 | 131K | 2 |
| Muse Spark 1.2 (this page) | $0.0034 | 1M | 5 |

If Muse Spark 1.2 is more model than a given task needs, Muse Glimmer 30B is the cheaper Meta option at $0.0011 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Muse Spark 1.2 accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0034 for one standard answer, ranking 63 of 89 priced models
- 1M token context window, well above the catalogue median
- 5 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.25 and output $4.25 per million tokens
- $0.0034 for one standard answer, $3.38 for a thousand
- 1M token context window
- 5 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 63 of 89 priced models by cost per answer

FAQ:

Q: How much does Muse Spark 1.2 cost?
A: $1.25 per million input tokens and $4.25 per million output tokens at list rates, which works out to $0.0034 for one standard answer of 1,000 tokens in and 500 out, or $3.38 for a thousand of them. Inside Whizi it is 5 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Muse Spark 1.2?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 5 credits per message, a Powerhouse allowance of 8,000 credits covers 1,600 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Muse Spark 1.2?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Muse Spark 1.2 expensive compared to other models?
A: It ranks 63 of 89 priced models by cost per answer, so 26 cost more and 62 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Nemotron 3.5 Lightning: price, context window and credit cost

URL: https://whizi.io/models/nemotron-3-5-lightning/
Updated: August 2026

Quick answer: Nemotron 3.5 Lightning costs $0.0002 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.08 per million input tokens and $0.2 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Nemotron 3.5 Lightning costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Nemotron 3.5 Lightning is a NVIDIA model priced at $0.08 per million input tokens and $0.2 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0002. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 81 of the 89 priced models cost more per answer, and 7 cost less.

#### What Nemotron 3.5 Lightning costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.08 |
| Output, per million tokens | $0.2 |
| One standard answer | $0.0002 |
| One thousand answers | $0.1800 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 8 |

Prices are the published NVIDIA rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Nemotron 3.5 Lightning is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Nemotron 3.5 Lightning is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Nemotron 3.5 Lightning costs 3.4 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 9.6 times less.

Its nearest neighbours from NVIDIA:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Nemotron 3.5 Lightning (this page) | $0.0002 | 262K | 1 |
| Nemotron 3 Ultra | $0.0024 | 512K | 3 |

It is the cheapest NVIDIA model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Nemotron 3.5 Lightning accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0002 for one standard answer, ranking 8 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.08 and output $0.2 per million tokens
- $0.0002 for one standard answer, $0.1800 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 8 of 89 priced models by cost per answer

FAQ:

Q: How much does Nemotron 3.5 Lightning cost?
A: $0.08 per million input tokens and $0.2 per million output tokens at list rates, which works out to $0.0002 for one standard answer of 1,000 tokens in and 500 out, or $0.1800 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Nemotron 3.5 Lightning?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Nemotron 3.5 Lightning?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Nemotron 3.5 Lightning expensive compared to other models?
A: It ranks 8 of 89 priced models by cost per answer, so 81 cost more and 7 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Nemotron 3 Ultra: price, context window and credit cost

URL: https://whizi.io/models/nemotron-3-ultra-550b-a55b/
Updated: August 2026

Quick answer: Nemotron 3 Ultra costs $0.0024 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.6 per million input tokens and $3.6 per million output tokens. The context window is 512K tokens, around the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What Nemotron 3 Ultra costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Nemotron 3 Ultra is a NVIDIA model priced at $0.6 per million input tokens and $3.6 per million output tokens, with a 512K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0024. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 35 of the 89 priced models cost more per answer, and 53 cost less.

#### What Nemotron 3 Ultra costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.6 |
| Output, per million tokens | $3.6 |
| One standard answer | $0.0024 |
| One thousand answers | $2.40 |
| Context window | 512K tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 54 |

Prices are the published NVIDIA rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Nemotron 3 Ultra is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

Nemotron 3 Ultra is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Nemotron 3 Ultra costs 45 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.4 times more.

Its nearest neighbours from NVIDIA:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Nemotron 3.5 Lightning | $0.0002 | 262K | 1 |
| Nemotron 3 Ultra (this page) | $0.0024 | 512K | 3 |

If Nemotron 3 Ultra is more model than a given task needs, Nemotron 3.5 Lightning is the cheaper NVIDIA option at $0.0002 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Nemotron 3 Ultra accepts 512K tokens of context, which is roughly 770 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0024 for one standard answer, ranking 54 of 89 priced models
- 512K token context window, around the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.6 and output $3.6 per million tokens
- $0.0024 for one standard answer, $2.40 for a thousand
- 512K token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 54 of 89 priced models by cost per answer

FAQ:

Q: How much does Nemotron 3 Ultra cost?
A: $0.6 per million input tokens and $3.6 per million output tokens at list rates, which works out to $0.0024 for one standard answer of 1,000 tokens in and 500 out, or $2.40 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Nemotron 3 Ultra?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Nemotron 3 Ultra?
A: 512K tokens, which is roughly 770 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Nemotron 3 Ultra expensive compared to other models?
A: It ranks 54 of 89 priced models by cost per answer, so 35 cost more and 53 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Nex-N2-Mini: price, context window and credit cost

URL: https://whizi.io/models/nex-n2-mini/
Updated: August 2026

Quick answer: Nex-N2-Mini costs $0.000075 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.025 per million input tokens and $0.1 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Nex-N2-Mini costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Nex-N2-Mini is a Nex Agi model priced at $0.025 per million input tokens and $0.1 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.000075. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 87 of the 89 priced models cost more per answer, and 1 cost less.

#### What Nex-N2-Mini costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.025 |
| Output, per million tokens | $0.1 |
| One standard answer | $0.000075 |
| One thousand answers | $0.0750 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 2 |

Prices are the published Nex Agi rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Nex-N2-Mini is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Nex-N2-Mini is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Nex-N2-Mini costs 1.4 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 22.9 times less.

Its nearest neighbours from Nex Agi:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Nex-N2-Mini (this page) | $0.000075 | 262K | 1 |
| Nex-N2-Pro | $0.0008 | 262K | 1 |

It is the cheapest Nex Agi model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Nex-N2-Mini accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.000075 for one standard answer, ranking 2 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.025 and output $0.1 per million tokens
- $0.000075 for one standard answer, $0.0750 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 2 of 89 priced models by cost per answer

FAQ:

Q: How much does Nex-N2-Mini cost?
A: $0.025 per million input tokens and $0.1 per million output tokens at list rates, which works out to $0.000075 for one standard answer of 1,000 tokens in and 500 out, or $0.0750 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Nex-N2-Mini?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Nex-N2-Mini?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Nex-N2-Mini expensive compared to other models?
A: It ranks 2 of 89 priced models by cost per answer, so 87 cost more and 1 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Nex-N2-Pro: price, context window and credit cost

URL: https://whizi.io/models/nex-n2-pro/
Updated: August 2026

Quick answer: Nex-N2-Pro costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.25 per million input tokens and $1 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Nex-N2-Pro costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Nex-N2-Pro is a Nex Agi model priced at $0.25 per million input tokens and $1 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 66 of the 89 priced models cost more per answer, and 22 cost less.

#### What Nex-N2-Pro costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.25 |
| Output, per million tokens | $1 |
| One standard answer | $0.0008 |
| One thousand answers | $0.7500 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 23 |

Prices are the published Nex Agi rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Nex-N2-Pro is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Nex-N2-Pro is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Nex-N2-Pro costs 14 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.3 times less.

Its nearest neighbours from Nex Agi:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Nex-N2-Mini | $0.000075 | 262K | 1 |
| Nex-N2-Pro (this page) | $0.0008 | 262K | 1 |

If Nex-N2-Pro is more model than a given task needs, Nex-N2-Mini is the cheaper Nex Agi option at $0.000075 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Nex-N2-Pro accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 23 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.25 and output $1 per million tokens
- $0.0008 for one standard answer, $0.7500 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 23 of 89 priced models by cost per answer

FAQ:

Q: How much does Nex-N2-Pro cost?
A: $0.25 per million input tokens and $1 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.7500 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Nex-N2-Pro?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Nex-N2-Pro?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Nex-N2-Pro expensive compared to other models?
A: It ranks 23 of 89 priced models by cost per answer, so 66 cost more and 22 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Nova Pro 1.0: price, context window and credit cost

URL: https://whizi.io/models/nova-pro-v1/
Updated: August 2026

Quick answer: Nova Pro 1.0 costs $0.0024 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.8 per million input tokens and $3.2 per million output tokens. The context window is 300K tokens, below the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What Nova Pro 1.0 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Nova Pro 1.0 is an Amazon model priced at $0.8 per million input tokens and $3.2 per million output tokens, with a 300K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0024. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 36 of the 89 priced models cost more per answer, and 52 cost less.

#### What Nova Pro 1.0 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.8 |
| Output, per million tokens | $3.2 |
| One standard answer | $0.0024 |
| One thousand answers | $2.40 |
| Context window | 300K tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 53 |

Prices are the published Amazon rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Nova Pro 1.0 is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

Nova Pro 1.0 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Nova Pro 1.0 costs 45 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.4 times more.

It is the only Amazon model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Amazon model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Nova Pro 1.0 accepts 300K tokens of context, which is roughly 450 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0024 for one standard answer, ranking 53 of 89 priced models
- 300K token context window, below the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.8 and output $3.2 per million tokens
- $0.0024 for one standard answer, $2.40 for a thousand
- 300K token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 53 of 89 priced models by cost per answer

FAQ:

Q: How much does Nova Pro 1.0 cost?
A: $0.8 per million input tokens and $3.2 per million output tokens at list rates, which works out to $0.0024 for one standard answer of 1,000 tokens in and 500 out, or $2.40 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Nova Pro 1.0?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Nova Pro 1.0?
A: 300K tokens, which is roughly 450 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Nova Pro 1.0 expensive compared to other models?
A: It ranks 53 of 89 priced models by cost per answer, so 36 cost more and 52 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Perceptron Mk1: price, context window and credit cost

URL: https://whizi.io/models/perceptron-mk1/
Updated: August 2026

Quick answer: Perceptron Mk1 costs $0.0009 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.15 per million input tokens and $1.5 per million output tokens. The context window is 33K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Perceptron Mk1 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Perceptron Mk1 is a Perceptron model priced at $0.15 per million input tokens and $1.5 per million output tokens, with a 33K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0009. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 59 of the 89 priced models cost more per answer, and 29 cost less.

#### What Perceptron Mk1 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.15 |
| Output, per million tokens | $1.5 |
| One standard answer | $0.0009 |
| One thousand answers | $0.9000 |
| Context window | 33K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 30 |

Prices are the published Perceptron rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Perceptron Mk1 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Perceptron Mk1 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Perceptron Mk1 costs 17 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.9 times less.

It is the only Perceptron model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Perceptron model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Perceptron Mk1 accepts 33K tokens of context, which is roughly 50 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0009 for one standard answer, ranking 30 of 89 priced models
- 33K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.15 and output $1.5 per million tokens
- $0.0009 for one standard answer, $0.9000 for a thousand
- 33K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 30 of 89 priced models by cost per answer

FAQ:

Q: How much does Perceptron Mk1 cost?
A: $0.15 per million input tokens and $1.5 per million output tokens at list rates, which works out to $0.0009 for one standard answer of 1,000 tokens in and 500 out, or $0.9000 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Perceptron Mk1?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Perceptron Mk1?
A: 33K tokens, which is roughly 50 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Perceptron Mk1 expensive compared to other models?
A: It ranks 30 of 89 priced models by cost per answer, so 59 cost more and 29 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Phi 4: price, context window and credit cost

URL: https://whizi.io/models/phi-4/
Updated: August 2026

Quick answer: Phi 4 costs $0.0001 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.07 per million input tokens and $0.14 per million output tokens. The context window is 16K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Phi 4 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Phi 4 is a Microsoft model priced at $0.07 per million input tokens and $0.14 per million output tokens, with a 16K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0001. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 82 of the 89 priced models cost more per answer, and 6 cost less.

#### What Phi 4 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.07 |
| Output, per million tokens | $0.14 |
| One standard answer | $0.0001 |
| One thousand answers | $0.1400 |
| Context window | 16K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 7 |

Prices are the published Microsoft rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Phi 4 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Phi 4 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Phi 4 costs 2.6 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 12.3 times less.

It is the only Microsoft model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Microsoft model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Phi 4 accepts 16K tokens of context, which is roughly 24 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0001 for one standard answer, ranking 7 of 89 priced models
- 16K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.07 and output $0.14 per million tokens
- $0.0001 for one standard answer, $0.1400 for a thousand
- 16K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 7 of 89 priced models by cost per answer

FAQ:

Q: How much does Phi 4 cost?
A: $0.07 per million input tokens and $0.14 per million output tokens at list rates, which works out to $0.0001 for one standard answer of 1,000 tokens in and 500 out, or $0.1400 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Phi 4?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Phi 4?
A: 16K tokens, which is roughly 24 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Phi 4 expensive compared to other models?
A: It ranks 7 of 89 priced models by cost per answer, so 82 cost more and 6 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.5 Plus 2026-04-20: price, context window and credit cost

URL: https://whizi.io/models/qwen3-5-plus-20260420/
Updated: August 2026

Quick answer: Qwen3.5 Plus 2026-04-20 costs $0.0012 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.3 per million input tokens and $1.8 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Qwen3.5 Plus 2026-04-20 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.5 Plus 2026-04-20 is a Qwen model priced at $0.3 per million input tokens and $1.8 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0012. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 54 of the 89 priced models cost more per answer, and 34 cost less.

#### What Qwen3.5 Plus 2026-04-20 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.3 |
| Output, per million tokens | $1.8 |
| One standard answer | $0.0012 |
| One thousand answers | $1.20 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 35 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.5 Plus 2026-04-20 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Qwen3.5 Plus 2026-04-20 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.5 Plus 2026-04-20 costs 23 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.4 times less.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.6 35B A3B | $0.0006 | 262K | 1 |
| Qwen3.6 Flash | $0.0008 | 1M | 1 |
| Qwen3.7 Plus | $0.0010 | 1M | 1 |
| Qwen3.5 Plus 2026-04-20 (this page) | $0.0012 | 1M | 1 |

If Qwen3.5 Plus 2026-04-20 is more model than a given task needs, Qwen3.7 Plus is the cheaper Qwen option at $0.0010 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.5 Plus 2026-04-20 accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0012 for one standard answer, ranking 35 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.3 and output $1.8 per million tokens
- $0.0012 for one standard answer, $1.20 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 35 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.5 Plus 2026-04-20 cost?
A: $0.3 per million input tokens and $1.8 per million output tokens at list rates, which works out to $0.0012 for one standard answer of 1,000 tokens in and 500 out, or $1.20 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.5 Plus 2026-04-20?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.5 Plus 2026-04-20?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.5 Plus 2026-04-20 expensive compared to other models?
A: It ranks 35 of 89 priced models by cost per answer, so 54 cost more and 34 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.6 35B A3B: price, context window and credit cost

URL: https://whizi.io/models/qwen3-6-35b-a3b/
Updated: August 2026

Quick answer: Qwen3.6 35B A3B costs $0.0006 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.14 per million input tokens and $1 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Qwen3.6 35B A3B costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.6 35B A3B is a Qwen model priced at $0.14 per million input tokens and $1 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0006. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 70 of the 89 priced models cost more per answer, and 18 cost less.

#### What Qwen3.6 35B A3B costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.14 |
| Output, per million tokens | $1 |
| One standard answer | $0.0006 |
| One thousand answers | $0.6400 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 19 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.6 35B A3B is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Qwen3.6 35B A3B is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.6 35B A3B costs 12 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.7 times less.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3 Coder Next | $0.0005 | 262K | 1 |
| Qwen3.6 35B A3B (this page) | $0.0006 | 262K | 1 |
| Qwen3.6 Flash | $0.0008 | 1M | 1 |
| Qwen3.7 Plus | $0.0010 | 1M | 1 |

If Qwen3.6 35B A3B is more model than a given task needs, Qwen3 Coder Next is the cheaper Qwen option at $0.0005 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.6 35B A3B accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0006 for one standard answer, ranking 19 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.14 and output $1 per million tokens
- $0.0006 for one standard answer, $0.6400 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 19 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.6 35B A3B cost?
A: $0.14 per million input tokens and $1 per million output tokens at list rates, which works out to $0.0006 for one standard answer of 1,000 tokens in and 500 out, or $0.6400 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.6 35B A3B?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.6 35B A3B?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.6 35B A3B expensive compared to other models?
A: It ranks 19 of 89 priced models by cost per answer, so 70 cost more and 18 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.6 Flash: price, context window and credit cost

URL: https://whizi.io/models/qwen3-6-flash/
Updated: August 2026

Quick answer: Qwen3.6 Flash costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.1875 per million input tokens and $1.125 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Qwen3.6 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.6 Flash is a Qwen model priced at $0.1875 per million input tokens and $1.125 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 67 of the 89 priced models cost more per answer, and 21 cost less.

#### What Qwen3.6 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.1875 |
| Output, per million tokens | $1.125 |
| One standard answer | $0.0008 |
| One thousand answers | $0.7500 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 22 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.6 Flash is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Qwen3.6 Flash is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.6 Flash costs 14 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.3 times less.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3 Coder Next | $0.0005 | 262K | 1 |
| Qwen3.6 35B A3B | $0.0006 | 262K | 1 |
| Qwen3.6 Flash (this page) | $0.0008 | 1M | 1 |
| Qwen3.7 Plus | $0.0010 | 1M | 1 |

If Qwen3.6 Flash is more model than a given task needs, Qwen3.6 35B A3B is the cheaper Qwen option at $0.0006 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.6 Flash accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 22 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.1875 and output $1.125 per million tokens
- $0.0008 for one standard answer, $0.7500 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 22 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.6 Flash cost?
A: $0.1875 per million input tokens and $1.125 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.7500 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.6 Flash?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.6 Flash?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.6 Flash expensive compared to other models?
A: It ranks 22 of 89 priced models by cost per answer, so 67 cost more and 21 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.7 Flash: price, context window and credit cost

URL: https://whizi.io/models/qwen3-7-flash/
Updated: August 2026

Quick answer: Qwen3.7 Flash costs $0.000095 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.03 per million input tokens and $0.13 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Qwen3.7 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.7 Flash is a Qwen model priced at $0.03 per million input tokens and $0.13 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.000095. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 85 of the 89 priced models cost more per answer, and 3 cost less.

#### What Qwen3.7 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.03 |
| Output, per million tokens | $0.13 |
| One standard answer | $0.000095 |
| One thousand answers | $0.0950 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 4 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.7 Flash is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Qwen3.7 Flash is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.7 Flash costs 1.8 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 18.1 times less.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.7 Flash (this page) | $0.000095 | 1M | 1 |
| Qwen3 Coder Next | $0.0005 | 262K | 1 |
| Qwen3.6 35B A3B | $0.0006 | 262K | 1 |
| Qwen3.6 Flash | $0.0008 | 1M | 1 |

It is the cheapest Qwen model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Qwen3.7 Flash accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.000095 for one standard answer, ranking 4 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.03 and output $0.13 per million tokens
- $0.000095 for one standard answer, $0.0950 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 4 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.7 Flash cost?
A: $0.03 per million input tokens and $0.13 per million output tokens at list rates, which works out to $0.000095 for one standard answer of 1,000 tokens in and 500 out, or $0.0950 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.7 Flash?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.7 Flash?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.7 Flash expensive compared to other models?
A: It ranks 4 of 89 priced models by cost per answer, so 85 cost more and 3 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.7 Max: price, context window and credit cost

URL: https://whizi.io/models/qwen3-7-max/
Updated: August 2026

Quick answer: Qwen3.7 Max costs $0.0037 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $1.475 per million input tokens and $4.425 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 5 credits per message on the Powerhouse plan and above.

What Qwen3.7 Max costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.7 Max is a Qwen model priced at $1.475 per million input tokens and $4.425 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0037. Inside Whizi it costs 5 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 22 of the 89 priced models cost more per answer, and 66 cost less.

#### What Qwen3.7 Max costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $1.475 |
| Output, per million tokens | $4.425 |
| One standard answer | $0.0037 |
| One thousand answers | $3.69 |
| Context window | 1M tokens |
| Credits per message in Whizi | 5 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 67 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.7 Max is charged at 5 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,600 |

Qwen3.7 Max is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.7 Max costs 70 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.1 times more.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.5 Plus 2026-04-20 | $0.0012 | 1M | 1 |
| Qwen3.8 27B | $0.0021 | 1M | 3 |
| Qwen3.7 Max (this page) | $0.0037 | 1M | 5 |
| Qwen3.8 Max | $0.0050 | 1M | 6 |

If Qwen3.7 Max is more model than a given task needs, Qwen3.8 27B is the cheaper Qwen option at $0.0021 per answer and 3 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.7 Max accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0037 for one standard answer, ranking 67 of 89 priced models
- 1M token context window, well above the catalogue median
- 5 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $1.475 and output $4.425 per million tokens
- $0.0037 for one standard answer, $3.69 for a thousand
- 1M token context window
- 5 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 67 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.7 Max cost?
A: $1.475 per million input tokens and $4.425 per million output tokens at list rates, which works out to $0.0037 for one standard answer of 1,000 tokens in and 500 out, or $3.69 for a thousand of them. Inside Whizi it is 5 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.7 Max?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 5 credits per message, a Powerhouse allowance of 8,000 credits covers 1,600 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.7 Max?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.7 Max expensive compared to other models?
A: It ranks 67 of 89 priced models by cost per answer, so 22 cost more and 66 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.7 Plus: price, context window and credit cost

URL: https://whizi.io/models/qwen3-7-plus/
Updated: August 2026

Quick answer: Qwen3.7 Plus costs $0.0010 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.32 per million input tokens and $1.28 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 1 credit per message on the Pro plan and above.

What Qwen3.7 Plus costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.7 Plus is a Qwen model priced at $0.32 per million input tokens and $1.28 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0010. Inside Whizi it costs 1 credit per message and needs the Pro plan or above.

That makes it mid-priced within the catalogue: 58 of the 89 priced models cost more per answer, and 30 cost less.

#### What Qwen3.7 Plus costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.32 |
| Output, per million tokens | $1.28 |
| One standard answer | $0.0010 |
| One thousand answers | $0.9600 |
| Context window | 1M tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Pro |
| Rank by cost, of 89 priced models | 31 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.7 Plus is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Pro | $29.99, or $19.99 billed yearly | 2,000 | 2,000 |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Qwen3.7 Plus is included from the Pro plan up. Pro is a curated tier rather than everything cheap: one or two current flagships per model family, each fast tier, and the high volume workhorses. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.7 Plus costs 18 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.8 times less.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.6 35B A3B | $0.0006 | 262K | 1 |
| Qwen3.6 Flash | $0.0008 | 1M | 1 |
| Qwen3.7 Plus (this page) | $0.0010 | 1M | 1 |
| Qwen3.5 Plus 2026-04-20 | $0.0012 | 1M | 1 |

If Qwen3.7 Plus is more model than a given task needs, Qwen3.6 Flash is the cheaper Qwen option at $0.0008 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.7 Plus accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0010 for one standard answer, ranking 31 of 89 priced models
- 1M token context window, well above the catalogue median
- 1 credit per message in Whizi, on the Pro plan and above

Checklist:

- Input $0.32 and output $1.28 per million tokens
- $0.0010 for one standard answer, $0.9600 for a thousand
- 1M token context window
- 1 credit per message inside Whizi
- Requires the Pro plan or above
- Ranks 31 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.7 Plus cost?
A: $0.32 per million input tokens and $1.28 per million output tokens at list rates, which works out to $0.0010 for one standard answer of 1,000 tokens in and 500 out, or $0.9600 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.7 Plus?
A: Pro and above, which is $29.99 per month or $19.99 billed yearly. At 1 credits per message, a Pro allowance of 2,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.7 Plus?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.7 Plus expensive compared to other models?
A: It ranks 31 of 89 priced models by cost per answer, so 58 cost more and 30 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.8 2.4T A95B: price, context window and credit cost

URL: https://whizi.io/models/qwen3-8-2-4t-a95b/
Updated: August 2026

Quick answer: Qwen3.8 2.4T A95B costs $0.0050 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $6 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 8 credits per message on the Powerhouse plan and above.

What Qwen3.8 2.4T A95B costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.8 2.4T A95B is a Qwen model priced at $2 per million input tokens and $6 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0050. Inside Whizi it costs 8 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 19 of the 89 priced models cost more per answer, and 69 cost less.

#### What Qwen3.8 2.4T A95B costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $6 |
| One standard answer | $0.0050 |
| One thousand answers | $5.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 8 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 70 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.8 2.4T A95B is charged at 8 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,000 |

Qwen3.8 2.4T A95B is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.8 2.4T A95B costs 94 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.9 times more.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.5 Plus 2026-04-20 | $0.0012 | 1M | 1 |
| Qwen3.8 27B | $0.0021 | 1M | 3 |
| Qwen3.7 Max | $0.0037 | 1M | 5 |
| Qwen3.8 2.4T A95B (this page) | $0.0050 | 1M | 8 |

If Qwen3.8 2.4T A95B is more model than a given task needs, Qwen3.7 Max is the cheaper Qwen option at $0.0037 per answer and 5 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.8 2.4T A95B accepts 1M tokens of context, which is roughly 1600 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0050 for one standard answer, ranking 70 of 89 priced models
- 1M token context window, well above the catalogue median
- 8 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $2 and output $6 per million tokens
- $0.0050 for one standard answer, $5.00 for a thousand
- 1M token context window
- 8 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 70 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.8 2.4T A95B cost?
A: $2 per million input tokens and $6 per million output tokens at list rates, which works out to $0.0050 for one standard answer of 1,000 tokens in and 500 out, or $5.00 for a thousand of them. Inside Whizi it is 8 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.8 2.4T A95B?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 8 credits per message, a Powerhouse allowance of 8,000 credits covers 1,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.8 2.4T A95B?
A: 1M tokens, which is roughly 1600 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.8 2.4T A95B expensive compared to other models?
A: It ranks 70 of 89 priced models by cost per answer, so 19 cost more and 69 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.8 27B: price, context window and credit cost

URL: https://whizi.io/models/qwen3-8-27b/
Updated: August 2026

Quick answer: Qwen3.8 27B costs $0.0021 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.45 per million input tokens and $3.2 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What Qwen3.8 27B costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.8 27B is a Qwen model priced at $0.45 per million input tokens and $3.2 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0021. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 38 of the 89 priced models cost more per answer, and 50 cost less.

#### What Qwen3.8 27B costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.45 |
| Output, per million tokens | $3.2 |
| One standard answer | $0.0021 |
| One thousand answers | $2.05 |
| Context window | 1M tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 51 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.8 27B is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

Qwen3.8 27B is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.8 27B costs 39 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.2 times more.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.6 Flash | $0.0008 | 1M | 1 |
| Qwen3.7 Plus | $0.0010 | 1M | 1 |
| Qwen3.5 Plus 2026-04-20 | $0.0012 | 1M | 1 |
| Qwen3.8 27B (this page) | $0.0021 | 1M | 3 |

If Qwen3.8 27B is more model than a given task needs, Qwen3.5 Plus 2026-04-20 is the cheaper Qwen option at $0.0012 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.8 27B accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0021 for one standard answer, ranking 51 of 89 priced models
- 1M token context window, well above the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.45 and output $3.2 per million tokens
- $0.0021 for one standard answer, $2.05 for a thousand
- 1M token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 51 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.8 27B cost?
A: $0.45 per million input tokens and $3.2 per million output tokens at list rates, which works out to $0.0021 for one standard answer of 1,000 tokens in and 500 out, or $2.05 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.8 27B?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.8 27B?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.8 27B expensive compared to other models?
A: It ranks 51 of 89 priced models by cost per answer, so 38 cost more and 50 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3.8 Max: price, context window and credit cost

URL: https://whizi.io/models/qwen3-8-max/
Updated: August 2026

Quick answer: Qwen3.8 Max costs $0.0050 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $2 per million input tokens and $6 per million output tokens. The context window is 1M tokens, well above the catalogue median. Inside Whizi it costs 6 credits per message on the Powerhouse plan and above.

What Qwen3.8 Max costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3.8 Max is a Qwen model priced at $2 per million input tokens and $6 per million output tokens, with a 1M token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0050. Inside Whizi it costs 6 credits per message and needs the Powerhouse plan or above.

That makes it one of the more expensive models in the catalogue: 20 of the 89 priced models cost more per answer, and 68 cost less.

#### What Qwen3.8 Max costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $2 |
| Output, per million tokens | $6 |
| One standard answer | $0.0050 |
| One thousand answers | $5.00 |
| Context window | 1M tokens |
| Credits per message in Whizi | 6 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 69 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3.8 Max is charged at 6 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 1,333 |

Qwen3.8 Max is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3.8 Max costs 94 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.9 times more.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.5 Plus 2026-04-20 | $0.0012 | 1M | 1 |
| Qwen3.8 27B | $0.0021 | 1M | 3 |
| Qwen3.7 Max | $0.0037 | 1M | 5 |
| Qwen3.8 Max (this page) | $0.0050 | 1M | 6 |

If Qwen3.8 Max is more model than a given task needs, Qwen3.7 Max is the cheaper Qwen option at $0.0037 per answer and 5 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3.8 Max accepts 1M tokens of context, which is roughly 1500 pages of text. That is well above the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0050 for one standard answer, ranking 69 of 89 priced models
- 1M token context window, well above the catalogue median
- 6 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $2 and output $6 per million tokens
- $0.0050 for one standard answer, $5.00 for a thousand
- 1M token context window
- 6 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 69 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3.8 Max cost?
A: $2 per million input tokens and $6 per million output tokens at list rates, which works out to $0.0050 for one standard answer of 1,000 tokens in and 500 out, or $5.00 for a thousand of them. Inside Whizi it is 6 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3.8 Max?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 6 credits per message, a Powerhouse allowance of 8,000 credits covers 1,333 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3.8 Max?
A: 1M tokens, which is roughly 1500 pages of text and sits well above the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3.8 Max expensive compared to other models?
A: It ranks 69 of 89 priced models by cost per answer, so 20 cost more and 68 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Qwen3 Coder Next: price, context window and credit cost

URL: https://whizi.io/models/qwen3-coder-next/
Updated: August 2026

Quick answer: Qwen3 Coder Next costs $0.0005 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.12 per million input tokens and $0.8 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Qwen3 Coder Next costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Qwen3 Coder Next is a Qwen model priced at $0.12 per million input tokens and $0.8 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0005. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 72 of the 89 priced models cost more per answer, and 16 cost less.

#### What Qwen3 Coder Next costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.12 |
| Output, per million tokens | $0.8 |
| One standard answer | $0.0005 |
| One thousand answers | $0.5200 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 17 |

Prices are the published Qwen rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Qwen3 Coder Next is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Qwen3 Coder Next is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Qwen3 Coder Next costs 9.8 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 3.3 times less.

Its nearest neighbours from Qwen:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Qwen3.7 Flash | $0.000095 | 1M | 1 |
| Qwen3 Coder Next (this page) | $0.0005 | 262K | 1 |
| Qwen3.6 35B A3B | $0.0006 | 262K | 1 |
| Qwen3.6 Flash | $0.0008 | 1M | 1 |

If Qwen3 Coder Next is more model than a given task needs, Qwen3.7 Flash is the cheaper Qwen option at $0.000095 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Qwen3 Coder Next accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0005 for one standard answer, ranking 17 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.12 and output $0.8 per million tokens
- $0.0005 for one standard answer, $0.5200 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 17 of 89 priced models by cost per answer

FAQ:

Q: How much does Qwen3 Coder Next cost?
A: $0.12 per million input tokens and $0.8 per million output tokens at list rates, which works out to $0.0005 for one standard answer of 1,000 tokens in and 500 out, or $0.5200 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Qwen3 Coder Next?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Qwen3 Coder Next?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Qwen3 Coder Next expensive compared to other models?
A: It ranks 17 of 89 priced models by cost per answer, so 72 cost more and 16 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Ring-2.6-1T: price, context window and credit cost

URL: https://whizi.io/models/ring-2-6-1t/
Updated: August 2026

Quick answer: Ring-2.6-1T costs $0.0004 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.075 per million input tokens and $0.625 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Ring-2.6-1T costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Ring-2.6-1T is an inclusionAI model priced at $0.075 per million input tokens and $0.625 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0004. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 76 of the 89 priced models cost more per answer, and 12 cost less.

#### What Ring-2.6-1T costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.075 |
| Output, per million tokens | $0.625 |
| One standard answer | $0.0004 |
| One thousand answers | $0.3875 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 13 |

Prices are the published inclusionAI rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Ring-2.6-1T is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Ring-2.6-1T is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Ring-2.6-1T costs 7.3 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 4.4 times less.

Its nearest neighbours from inclusionAI:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Ling-3.0-flash | $0.000053 | 262K | 1 |
| Ring-2.6-1T (this page) | $0.0004 | 262K | 1 |

If Ring-2.6-1T is more model than a given task needs, Ling-3.0-flash is the cheaper inclusionAI option at $0.000053 per answer and 1 credit per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Ring-2.6-1T accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0004 for one standard answer, ranking 13 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.075 and output $0.625 per million tokens
- $0.0004 for one standard answer, $0.3875 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 13 of 89 priced models by cost per answer

FAQ:

Q: How much does Ring-2.6-1T cost?
A: $0.075 per million input tokens and $0.625 per million output tokens at list rates, which works out to $0.0004 for one standard answer of 1,000 tokens in and 500 out, or $0.3875 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Ring-2.6-1T?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Ring-2.6-1T?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Ring-2.6-1T expensive compared to other models?
A: It ranks 13 of 89 priced models by cost per answer, so 76 cost more and 12 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Sakana Namazu: price, context window and credit cost

URL: https://whizi.io/models/sakana-namazu/
Updated: August 2026

Quick answer: Sakana Namazu costs $0.0029 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.95 per million input tokens and $4 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 4 credits per message on the Powerhouse plan and above.

What Sakana Namazu costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Sakana Namazu is a Sakana model priced at $0.95 per million input tokens and $4 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0029. Inside Whizi it costs 4 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 30 of the 89 priced models cost more per answer, and 58 cost less.

#### What Sakana Namazu costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.95 |
| Output, per million tokens | $4 |
| One standard answer | $0.0029 |
| One thousand answers | $2.95 |
| Context window | 262K tokens |
| Credits per message in Whizi | 4 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 59 |

Prices are the published Sakana rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Sakana Namazu is charged at 4 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,000 |

Sakana Namazu is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Sakana Namazu costs 56 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.7 times more.

Its nearest neighbours from Sakana:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Sakana Namazu (this page) | $0.0029 | 262K | 4 |
| Fugu Ultra | $0.0200 | 1M | 25 |

It is the cheapest Sakana model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Sakana Namazu accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0029 for one standard answer, ranking 59 of 89 priced models
- 262K token context window, below the catalogue median
- 4 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.95 and output $4 per million tokens
- $0.0029 for one standard answer, $2.95 for a thousand
- 262K token context window
- 4 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 59 of 89 priced models by cost per answer

FAQ:

Q: How much does Sakana Namazu cost?
A: $0.95 per million input tokens and $4 per million output tokens at list rates, which works out to $0.0029 for one standard answer of 1,000 tokens in and 500 out, or $2.95 for a thousand of them. Inside Whizi it is 4 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Sakana Namazu?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 4 credits per message, a Powerhouse allowance of 8,000 credits covers 2,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Sakana Namazu?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Sakana Namazu expensive compared to other models?
A: It ranks 59 of 89 priced models by cost per answer, so 30 cost more and 58 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Seed-2.0-Code: price, context window and credit cost

URL: https://whizi.io/models/seed-2-0-code/
Updated: August 2026

Quick answer: Seed-2.0-Code costs $0.0020 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.5 per million input tokens and $3 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 3 credits per message on the Powerhouse plan and above.

What Seed-2.0-Code costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Seed-2.0-Code is a Bytedance Seed model priced at $0.5 per million input tokens and $3 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0020. Inside Whizi it costs 3 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 40 of the 89 priced models cost more per answer, and 48 cost less.

#### What Seed-2.0-Code costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.5 |
| Output, per million tokens | $3 |
| One standard answer | $0.0020 |
| One thousand answers | $2.00 |
| Context window | 262K tokens |
| Credits per message in Whizi | 3 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 49 |

Prices are the published Bytedance Seed rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Seed-2.0-Code is charged at 3 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 2,666 |

Seed-2.0-Code is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Seed-2.0-Code costs 38 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 1.2 times more.

Its nearest neighbours from Bytedance Seed:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Seed 2.1 Turbo | $0.0018 | 262K | 2 |
| Seed-2.0-Code (this page) | $0.0020 | 262K | 3 |

If Seed-2.0-Code is more model than a given task needs, Seed 2.1 Turbo is the cheaper Bytedance Seed option at $0.0018 per answer and 2 credits per message. Routing routine work down a tier is the single biggest thing you can do to make an allowance last.

#### The context window in practice

Seed-2.0-Code accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0020 for one standard answer, ranking 49 of 89 priced models
- 262K token context window, below the catalogue median
- 3 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.5 and output $3 per million tokens
- $0.0020 for one standard answer, $2.00 for a thousand
- 262K token context window
- 3 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 49 of 89 priced models by cost per answer

FAQ:

Q: How much does Seed-2.0-Code cost?
A: $0.5 per million input tokens and $3 per million output tokens at list rates, which works out to $0.0020 for one standard answer of 1,000 tokens in and 500 out, or $2.00 for a thousand of them. Inside Whizi it is 3 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Seed-2.0-Code?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 3 credits per message, a Powerhouse allowance of 8,000 credits covers 2,666 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Seed-2.0-Code?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Seed-2.0-Code expensive compared to other models?
A: It ranks 49 of 89 priced models by cost per answer, so 40 cost more and 48 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Seed 2.1 Turbo: price, context window and credit cost

URL: https://whizi.io/models/seed-2-1-turbo/
Updated: August 2026

Quick answer: Seed 2.1 Turbo costs $0.0018 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.5 per million input tokens and $2.5 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 2 credits per message on the Powerhouse plan and above.

What Seed 2.1 Turbo costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Seed 2.1 Turbo is a Bytedance Seed model priced at $0.5 per million input tokens and $2.5 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0018. Inside Whizi it costs 2 credits per message and needs the Powerhouse plan or above.

That makes it mid-priced within the catalogue: 43 of the 89 priced models cost more per answer, and 45 cost less.

#### What Seed 2.1 Turbo costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.5 |
| Output, per million tokens | $2.5 |
| One standard answer | $0.0018 |
| One thousand answers | $1.75 |
| Context window | 262K tokens |
| Credits per message in Whizi | 2 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 46 |

Prices are the published Bytedance Seed rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Seed 2.1 Turbo is charged at 2 credits per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 4,000 |

Seed 2.1 Turbo is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Seed 2.1 Turbo costs 33 times more. Against the median, Kimi K2 0711 at $0.0017, it costs about the same.

Its nearest neighbours from Bytedance Seed:

| Model | One answer | Context | Credits |
| --- | --- | --- | --- |
| Seed 2.1 Turbo (this page) | $0.0018 | 262K | 2 |
| Seed-2.0-Code | $0.0020 | 262K | 3 |

It is the cheapest Bytedance Seed model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Seed 2.1 Turbo accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0018 for one standard answer, ranking 46 of 89 priced models
- 262K token context window, below the catalogue median
- 2 credits per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.5 and output $2.5 per million tokens
- $0.0018 for one standard answer, $1.75 for a thousand
- 262K token context window
- 2 credits per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 46 of 89 priced models by cost per answer

FAQ:

Q: How much does Seed 2.1 Turbo cost?
A: $0.5 per million input tokens and $2.5 per million output tokens at list rates, which works out to $0.0018 for one standard answer of 1,000 tokens in and 500 out, or $1.75 for a thousand of them. Inside Whizi it is 2 credits per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Seed 2.1 Turbo?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 2 credits per message, a Powerhouse allowance of 8,000 credits covers 4,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Seed 2.1 Turbo?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Seed 2.1 Turbo expensive compared to other models?
A: It ranks 46 of 89 priced models by cost per answer, so 43 cost more and 45 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Solar Pro 4: price, context window and credit cost

URL: https://whizi.io/models/solar-pro4/
Updated: August 2026

Quick answer: Solar Pro 4 costs $0.000090 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.03 per million input tokens and $0.12 per million output tokens. The context window is 524K tokens, around the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Solar Pro 4 costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Solar Pro 4 is an Upstage model priced at $0.03 per million input tokens and $0.12 per million output tokens, with a 524K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.000090. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 86 of the 89 priced models cost more per answer, and 2 cost less.

#### What Solar Pro 4 costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.03 |
| Output, per million tokens | $0.12 |
| One standard answer | $0.000090 |
| One thousand answers | $0.0900 |
| Context window | 524K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 3 |

Prices are the published Upstage rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Solar Pro 4 is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Solar Pro 4 is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Solar Pro 4 costs 1.7 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 19.1 times less.

It is the only Upstage model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Upstage model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Solar Pro 4 accepts 524K tokens of context, which is roughly 790 pages of text. That is around the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.000090 for one standard answer, ranking 3 of 89 priced models
- 524K token context window, around the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.03 and output $0.12 per million tokens
- $0.000090 for one standard answer, $0.0900 for a thousand
- 524K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 3 of 89 priced models by cost per answer

FAQ:

Q: How much does Solar Pro 4 cost?
A: $0.03 per million input tokens and $0.12 per million output tokens at list rates, which works out to $0.000090 for one standard answer of 1,000 tokens in and 500 out, or $0.0900 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Solar Pro 4?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Solar Pro 4?
A: 524K tokens, which is roughly 790 pages of text and sits around the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Solar Pro 4 expensive compared to other models?
A: It ranks 3 of 89 priced models by cost per answer, so 86 cost more and 2 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

### Step 3.7 Flash: price, context window and credit cost

URL: https://whizi.io/models/step-3-7-flash/
Updated: August 2026

Quick answer: Step 3.7 Flash costs $0.0008 for one standard answer, meaning 1,000 tokens in and 500 out. List rates are $0.2 per million input tokens and $1.15 per million output tokens. The context window is 262K tokens, below the catalogue median. Inside Whizi it costs 1 credit per message on the Powerhouse plan and above.

What Step 3.7 Flash costs per token and per answer, how big its context window is, how many credits it spends in Whizi, and which plan unlocks it.

#### The short answer

Step 3.7 Flash is a Stepfun model priced at $0.2 per million input tokens and $1.15 per million output tokens, with a 262K token context window. One standard answer, meaning 1,000 tokens in and 500 tokens out, costs $0.0008. Inside Whizi it costs 1 credit per message and needs the Powerhouse plan or above.

That makes it one of the cheaper models in the catalogue: 64 of the 89 priced models cost more per answer, and 24 cost less.

#### What Step 3.7 Flash costs

List rates are what the provider charges per token. The cost of one standard answer is the more useful number, because it prices the same unit of work across every model regardless of how each one is billed.

| | |
| --- | --- |
| Input, per million tokens | $0.2 |
| Output, per million tokens | $1.15 |
| One standard answer | $0.0008 |
| One thousand answers | $0.7750 |
| Context window | 262K tokens |
| Credits per message in Whizi | 1 |
| Minimum Whizi plan | Powerhouse |
| Rank by cost, of 89 priced models | 25 |

Prices are the published Stepfun rates as carried by OpenRouter, last refreshed 2026-08-20. The full dataset covering all 100 models from 29 providers is the [AI model cost index](https://whizi.io/tools/model-cost-index), which is free to reuse with attribution.

#### What it costs inside Whizi

Whizi meters messages in credits rather than tokens, and Step 3.7 Flash is charged at 1 credit per message. The cost does not change with the length of your message or the length of the answer, so a one line question and a long document analysis cost the same.

| Plan | Price per month | Monthly credits | Messages on this model |
| --- | --- | --- | --- |
| Powerhouse | $49.99, or $34.99 billed yearly | 8,000 | 8,000 |

Step 3.7 Flash is part of the Powerhouse catalogue. Anything not explicitly listed on the Starter or Pro tiers resolves to Powerhouse, which is the deliberate default for the frontier and specialist models. See [plans and limits](https://whizi.io/docs/plans-and-limits).

#### How it compares on price

Against the cheapest model in the index, Ling-3.0-flash at $0.000053 per answer, Step 3.7 Flash costs 15 times more. Against the median, Kimi K2 0711 at $0.0017, it costs 2.2 times less.

It is the only Stepfun model in the priced index, so there is no same-provider comparison to draw.

It is the cheapest Stepfun model in the priced index, so within this provider there is nothing to route down to.

#### The context window in practice

Step 3.7 Flash accepts 262K tokens of context, which is roughly 390 pages of text. That is below the median for the priced catalogue.

Advertised context and usable context are not the same thing. Recall tends to degrade before the stated limit is reached, particularly for material in the middle of a long input, so the practical test is to upload a long document and ask about something buried halfway through rather than trusting the number.

The number matters most when you are working from your own files. If a document is too long for the model you picked, switching to a larger-context model in the same conversation carries the thread across, so you are not starting over. See [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation).

Key points:

- $0.0008 for one standard answer, ranking 25 of 89 priced models
- 262K token context window, below the catalogue median
- 1 credit per message in Whizi, on the Powerhouse plan and above

Checklist:

- Input $0.2 and output $1.15 per million tokens
- $0.0008 for one standard answer, $0.7750 for a thousand
- 262K token context window
- 1 credit per message inside Whizi
- Requires the Powerhouse plan or above
- Ranks 25 of 89 priced models by cost per answer

FAQ:

Q: How much does Step 3.7 Flash cost?
A: $0.2 per million input tokens and $1.15 per million output tokens at list rates, which works out to $0.0008 for one standard answer of 1,000 tokens in and 500 out, or $0.7750 for a thousand of them. Inside Whizi it is 1 credit per message on a monthly allowance, rather than metered per token.

Q: Which Whizi plan includes Step 3.7 Flash?
A: Powerhouse and above, which is $49.99 per month or $34.99 billed yearly. At 1 credits per message, a Powerhouse allowance of 8,000 credits covers 8,000 messages on this model if you spend the whole allowance here, and most people mix in a 1 credit model for routine work so it goes further.

Q: What is the context window of Step 3.7 Flash?
A: 262K tokens, which is roughly 390 pages of text and sits below the median for the priced catalogue. Treat that as the ceiling rather than the working figure: recall usually degrades before the stated limit, especially for content in the middle of a long input.

Q: Is Step 3.7 Flash expensive compared to other models?
A: It ranks 25 of 89 priced models by cost per answer, so 64 cost more and 24 cost less. The spread across the whole catalogue is about 2000 times from cheapest to priciest, which is far wider than most people assume, so "expensive" only means anything relative to the specific alternative you would otherwise use.

## Use cases by role and task

### How to analyze spreadsheets with AI in Whizi

URL: https://whizi.io/use-cases/analyze-spreadsheets-with-ai/
Updated: August 2026

Quick answer: To analyze a spreadsheet with AI, upload the CSV or Excel file and profile it before asking anything: row counts, missing values, duplicates, and distinct values per column. Then ask your real question and require the method alongside every number. Verify the arithmetic by spot checking one group and cross-checking with a second model.

Upload a CSV or Excel file to Whizi, ask GPT for structured analysis, and cross-check with Claude, without writing a single formula.

#### Profile the data before you ask it anything

The most common way spreadsheet analysis goes wrong has nothing to do with AI. It is that the file contains 14 rows with a trailing space in the region name, two date formats, a subtotal row somebody left in the middle, and 300 blanks in the column you are about to average. Ask your question first and you will get a confident answer computed over garbage.

So the first prompt is never the question. It is always this one.

**Prompt: the profile**

`Profile this file before we analyse anything. Return: row count, column names with inferred type, the count and percentage of missing values per column, the number of exact duplicate rows, the distinct values for every column with fewer than 25 distinct values, the min and max of every numeric and date column, and any column where the values look inconsistent (mixed formats, trailing whitespace, mixed units, mixed date formats). Do not analyse or interpret yet. Just tell me what is in the file and what looks wrong with it.`

That single prompt catches most of what would otherwise ruin the analysis. The distinct-values list in particular is where you discover that "UK", "U.K.", and "United Kingdom" are three separate regions in your data.

**Prompt: fix what it found**

`Standardise the issues you identified: trim whitespace, unify [column] to a single format, and consolidate these variants: [list them]. Show me the mapping you applied as a table before applying it. Do not drop any rows without telling me which and why.`

#### Asking questions that produce checkable answers

Vague questions get vague answers. The prompts that work name the column, the operation, and the output format, and they ask for the working.

**Prompt: aggregation with the working shown**

`Group by [column] and calculate [metric]. Return a Markdown table with the group, the count of rows in it, and the metric. Below the table, state exactly how you computed the metric, which rows you excluded and why, and how you handled blanks. Sort descending by [column].`

**Prompt: the cohort question**

`For each [cohort dimension, for example signup month], calculate [metric] at [interval]. Show the cohort size alongside every figure. Flag any cohort with fewer than [n] rows as too small to interpret rather than reporting a percentage for it.`

That last instruction prevents the single most misleading output in spreadsheet analysis: a cohort of four users reported as "75% retention" and sitting in a table next to a cohort of nine thousand.

**Prompt: anomaly hunting**

`What is suspicious about this data? Look for: values outside a plausible range, sudden discontinuities in a time series, columns where the distribution changes partway through, rows that are exact or near duplicates, values that appear too round, and anything that suggests a change in how the data was collected. For each, quote the specific rows.`

That is the highest-value prompt on this page and it has no equivalent in a normal spreadsheet workflow. It routinely finds the day the tracking broke, the vendor who started reporting in a different currency, and the duplicate import that inflated a quarter.

**Prompt: the chart specification**

`Recommend the right chart for this question and this data shape, and explain why the obvious alternative is worse here. Then give me the specification: chart type, x, y, series, aggregation, sort order, and axis treatment. Do not use a dual axis. Do not truncate a bar chart axis.`

#### Where the arithmetic actually goes wrong

This deserves a straight answer, because "AI is bad at math" is both true and unhelpfully vague.

A language model reasoning over numbers in text is doing pattern completion, not calculation. It is reliable at describing structure, categorising rows, and identifying which computation you need. It is much less reliable at performing a long chain of arithmetic across hundreds of rows, and it fails silently, producing a well formatted table of wrong numbers with no distress signal.

The practical rules:

- **Ask for the method, not just the result.** "State exactly how you computed this and what you excluded" makes the error visible when there is one.
- **Spot check one group by hand.** Pick the smallest group in the output table and verify it in the actual spreadsheet. If it matches, the method is probably right; if it does not, nothing in the table can be trusted.
- **Cross-check anything consequential with a second model.** Two independent models arriving at the same number is meaningfully better evidence than one model repeating itself. Whizi's side-by-side view exists for exactly this.
- **Watch for double counting.** Subtotal rows left in the file and one-to-many joins are the two usual culprits, and the model will not know they are wrong.
- **For anything that must be exact, ask for the formula or the code.** `Give me the Excel formula` or `give me the pandas code` puts the arithmetic in a deterministic engine, and you keep the model for the part it is good at, which is knowing which computation to run.

That last one is the real answer for financial or reporting work. Use the model to design the analysis, use your spreadsheet or a script to execute it.

#### File size, formats, and getting the data in

CSV and Excel both upload directly. A few practical notes:

- **Give it a clean rectangle.** One header row, no merged cells, no blank spacer rows, no notes in column K. Multi-header formatted reports confuse extraction more than any file size limit.
- **Send the raw sheet, not the presentation sheet.** The version with the formatting and the subtotals is the one that produces double counting.
- **Very wide files benefit from a column list first.** Ask what each column means before analysing if the names are cryptic, and tell the model what it got wrong.
- **For very large files, use Gemini**, which carries the largest context window of the models in Whizi. Beyond that, sample deliberately: `analyse a random 5000 row sample and tell me the sampling error I should expect on each figure` is better than silently truncating.
- **Strip personal data first.** Names, emails, and identifiers are almost never needed for the analysis, and removing them is faster than arguing about whether you were allowed to upload them.

The mechanics of uploads are covered in [uploading documents](https://whizi.io/docs/upload-documents).

#### A worked sequence

The whole workflow on a real file, in order:

1. Upload. Run the profile prompt. Read the distinct values and missing counts carefully.
2. Fix what it found, reviewing the mapping table before it is applied.
3. Ask for a summary of what the data is about and the three questions it thinks you should be asking. This step is quick and often reframes the analysis.
4. Ask your actual question, with the method and exclusions required in the answer.
5. Run the anomaly prompt, always. This is where the surprises are.
6. Spot check the smallest group by hand.
7. Cross-check the headline number with a second model.
8. Ask for the chart specification, or the formula if the number needs to be exact and reproducible.

Steps 1, 5, and 6 are the ones people skip, and they are the ones that separate an analysis you can defend from a nicely formatted guess.

Key points:

- A profiling step that catches the data problems which would otherwise poison every answer
- Prompts for cleaning, aggregation, cohort questions, and anomaly hunting
- An honest account of where AI arithmetic goes wrong and how to catch it

Checklist:

- Profile the file before asking a single analytical question
- Read the distinct-value list to catch variant spellings and mixed formats
- Require the method and the exclusions alongside every number
- Flag small groups as too small rather than reporting a percentage for them
- Always run the "what is suspicious about this data" prompt
- Spot check the smallest group by hand before trusting the table
- Cross-check consequential figures with a second model
- Ask for the formula or code when the number has to be exactly right

Steps:

1. Upload the file: Drag a CSV or Excel file into the Whizi composer.
2. Profile the data first: Ask for row count, columns, types, missing values, and duplicates before asking any analytical question.
3. Ask the real question: Ask a specific question: "What percentage of rows have status = churned?" or "Which segments grew more than 10%?"
4. Request a table or chart spec: Ask for the answer as a clean Markdown table or a chart specification you can paste into a tool.
5. Verify with a second model: Re-run the important questions through Claude or Gemini to catch arithmetic errors.

FAQ:

Q: How big can my spreadsheet be?
A: It depends on the plan and the model, and the practical limit is the model context window rather than a file size cap. Gemini in Whizi handles the largest files. Beyond that, sample deliberately and ask the model to state the sampling error rather than letting the file be silently truncated, which is the failure mode that produces confident answers about a fraction of your data.

Q: Can AI catch my spreadsheet errors?
A: Yes, and it is one of the highest-leverage prompts available. Asking what is suspicious about the data reliably surfaces duplicate imports, the day tracking broke, mixed units or currencies, subtotal rows left inside the data, and distributions that shift partway through the file. Ask it to quote the specific rows so you can verify each finding rather than taking the list on trust.

Q: Can I trust the numbers it gives me?
A: Trust the structure, verify the arithmetic. Language models are strong at identifying which computation is needed and describing what is in a dataset, and weaker at long arithmetic chains across many rows, where they fail silently with a well formatted wrong answer. Spot check the smallest group by hand, cross-check headline figures with a second model, and for anything that must be exact, ask for the Excel formula or Python code and run it yourself.

Q: Which model is best for spreadsheet work?
A: GPT for structured reasoning, strict output formats, and clean tables or JSON. Claude as the cross-check on multi-step aggregation, because it tends to make different mistakes rather than the same ones. Gemini when the file is very large or when you want to analyse a spreadsheet alongside a PDF or report in the same conversation.

Q: Does it work with Excel formulas and multiple sheets?
A: It reads the values rather than executing your workbook, so formula results come through but the live formula logic does not. For multi-sheet workbooks, say which sheet you mean and describe how the sheets relate, or export the sheet you care about as CSV. Files built for presentation, with merged cells and subtotals inside the data, cause far more problems than large files do.

### The AI workspace for consultants: decks, frameworks, and research at partner speed

URL: https://whizi.io/use-cases/consultants/
Updated: August 2026

Quick answer: AI for consultants works best as a routing exercise rather than an answer machine. Structure problems and issue trees in GPT, draft slide headlines and executive summaries in Claude, run market and competitor scans in Gemini, then red team the recommendation in a model that did not write it. Redact client names before pasting.

Whizi for consultants: build frameworks with GPT, draft client-ready decks with Claude, and run competitive research with Gemini, from one workspace.

#### What AI actually replaces on an engagement

Consulting work is not one job. In a single week you structure a problem, interview stakeholders, clean a data extract, build a storyline, write the pages, and defend the recommendation. Those tasks reward completely different model behavior, which is why a single AI subscription always feels 70 percent useful and 30 percent frustrating.

The honest framing is this: AI does not produce the answer on an engagement. It removes the two hours you spend getting to a first draft of something you already know how to build, and it catches the gap you would have found on Thursday if you had more time. Used that way it is worth several hours a week. Used as an answer machine it produces exactly the generic output partners have learned to spot.

Here is the routing most consultants land on after a few weeks:

| Task | Best model | Why |
| --- | --- | --- |
| Problem structuring, issue trees, 2x2s | GPT | Follows explicit structural constraints, holds MECE logic, returns clean hierarchies |
| Slide headlines and storyline | Claude | Writes in a human register, resists filler adjectives, keeps a consistent argument thread |
| Executive summary and client email | Claude | Best at tone control and at saying difficult things diplomatically |
| Market and competitor scans | Gemini | Strongest at recent web material with sources you can check |
| Long document and transcript reading | Gemini | Largest context window, so a 120 page annual report fits in one pass |
| Data extraction into tables or JSON | GPT | Most reliable at strict output formats you can paste into Excel |
| Sanity checking a recommendation | Whatever you did not draft in | A second model is the cheapest red team you will ever run |

You do not need to memorize that. The practical version is: draft structure in GPT, draft prose in Claude, research in Gemini, and check the final logic in a model that did not write it.

#### Workflow 1: framework generation that is not generic

Asking for "a framework for this client problem" returns a Porter's Five Forces with the client name pasted in. The fix is to give the model the constraint set a manager would give you: the decision the client is actually making, the axes that are allowed to matter, and what has already been ruled out.

**Prompt: issue tree**

`You are structuring a problem for a consulting engagement. Client decision: [the specific decision, for example "whether to consolidate three regional distribution centres into one"]. Build a MECE issue tree, three levels deep. Level 1 must be the 3 to 5 questions that, if answered, resolve the decision. For each leaf, state the analysis required and the data source needed. Do not include branches we cannot test with available data. Flag any branch where the answer is likely already known.`

**Prompt: 2x2 that earns its place**

`Propose three candidate 2x2 matrices for this problem. For each, name both axes, explain what a position in each quadrant would imply for the client, and state the single decision the matrix drives. Reject any axis pair that is a restatement of the same variable. Then recommend one and explain why the other two are weaker here. Context: [paste your situation notes].`

**Prompt: pressure test a framework you already built**

`Here is a framework I built: [paste]. Act as a sceptical engagement manager. Identify overlapping branches, missing branches, branches that cannot be analysed in the time available, and any place where the structure presumes the answer. Do not rewrite it. List the problems only.`

The last one is the highest value prompt on this page. Models are far better at critiquing structure than at inventing it, and a critique pass takes 30 seconds. Run your own framework through it before you take it into a manager review, not after.

#### Workflow 2: deck narrative that survives a partner review

The failure mode of AI written slides is not grammar. It is that every headline is a topic rather than an assertion. "Market overview" is a topic. "The category grew 11 percent while our client lost 3 points of share, driven entirely by the discount channel" is an assertion, and a deck of assertions reads as thinking rather than as reporting.

**Prompt: storyline first, pages second**

`Build a horizontal storyline for a client deck. Situation: [2 sentences]. Complication: [2 sentences]. Question the client is asking: [1 sentence]. Evidence I have: [bullets]. Recommendation: [1 sentence]. Return 8 to 12 slide headlines, each a full assertion with a subject and a verb, that read in sequence as a single argument. No topic labels. After the headlines, list any assertion I do not yet have evidence for.`

That final instruction is the one that saves you. It gives you an evidence gap list before you build pages, which is exactly the list a partner will generate in the review anyway.

**Prompt: write the page**

`Write the body content for this slide. Headline: [paste headline]. Supporting data: [paste numbers or findings]. Audience: [for example "COO and two direct reports, operationally fluent, sceptical of consultants"]. Constraints: three supporting points maximum, each under 20 words, no adjectives that cannot be measured, no claim that goes beyond the data given. Add a one line "so what" for the client.`

**Prompt: the tone pass**

`Rewrite this to match the voice sample. Keep every number and claim identical. Remove any sentence that would appear unchanged in a deck for a different client. Voice sample: [paste two paragraphs from a deck your partner liked]. Draft: [paste].`

The voice sample matters more than any instruction about tone. Two paragraphs of your firm's actual writing will move the output further than a paragraph of adjectives describing how you want it to sound.

Before you send anything, run the deck through the reverse test: paste your headlines back into a fresh chat and ask `What is the argument being made here, and what is the weakest link in it?` If the model cannot restate your argument, neither can the client.

#### Workflow 3: market scans you can actually cite

Research is where AI is most useful and most dangerous on an engagement. A confident, fabricated market size number that reaches a client deck is a career problem, not an inconvenience. The discipline is simple: use a web connected model, demand sources, and treat every number as unverified until you have opened the source yourself.

**Prompt: category scan**

`Research the [category] market. Return: (1) market size estimates from the last 24 months, each with the publishing organisation, the year, and the URL; (2) the 5 to 8 largest players with any publicly reported revenue or share figures; (3) material changes in the last 12 months, meaning entries, exits, funding rounds, acquisitions, and regulatory changes; (4) the two or three structural forces most likely to change the competitive picture in the next 24 months. Where sources disagree, show both numbers rather than averaging them. Mark anything you cannot source as UNVERIFIED.`

**Prompt: competitor teardown**

`Build a positioning teardown of [competitor]. Cover: stated value proposition in their own words, target segments, pricing where public, distribution channels, recent product or messaging changes with dates, and the customer complaint themes visible in public reviews. Cite each claim. Separate what is stated by the company from what is observed by third parties.`

**Prompt: reading the long document**

`Read this document and extract only what matters for [the client decision]. Return a table with: claim, page or section reference, and whether it supports, contradicts, or is neutral to the hypothesis that [state the hypothesis]. Then list the three things in this document that most surprise you against industry norms.`

The last prompt is why a large context window matters in this job. Loading an entire annual report, a full regulatory filing, or a set of 15 interview transcripts and asking one question against all of it is a genuinely different capability from summarizing a page at a time. Gemini carries the largest context of the models in Whizi, so it is the default for that pass.

For a deeper walkthrough of sourcing and verification, see the guide on [AI for market research](https://whizi.io/resources/ai-for-market-research).

#### Client confidentiality: what to redact before you paste

Every consultant asks this before the second prompt, and the honest answer has two parts. Whizi does not train on your conversations, and each provider's data handling policy is available before you enable it. That covers the platform. It does not cover your client agreement, which is usually stricter than any vendor policy and is the document that actually governs you.

The workable habit is to strip identity without stripping structure. AI reasoning does not need the client name to be useful, and removing it costs you nothing:

- Replace the client and competitor names with tokens: Client A, Competitor B, Region 1.
- Round or index figures. "Revenue indexed to 100 in 2023, 94 in 2025" preserves every bit of the analytical shape.
- Remove names, emails, and titles from interview transcripts before uploading. Roles are enough, and roles are what the analysis uses.
- Keep anything covered by a specific NDA clause, personal data, or material non public information out of any AI tool entirely, including this one.
- If the engagement is under heightened restrictions, check with your firm's risk team before the engagement, not after.

A redacted prompt is usually a better prompt anyway. Forcing yourself to describe the situation in structural terms removes the incidental detail that was diluting the model's attention.

#### Why one workspace beats three subscriptions here

The specific advantage on an engagement is not price, although paying once instead of three times is real, and independents feel a sharper version of that arithmetic, worked through in [AI tools for freelancers](https://whizi.io/use-cases/freelancers). It is that context survives the model switch. You upload the client's data extract, structure the problem with GPT, draft the storyline with Claude, and pull the market context with Gemini, all in one thread that remembers what came before. Doing the same across three products means re-explaining the engagement three times and holding the version history in your head.

Side-by-side comparison is the second advantage, and it maps directly onto how consulting quality is judged. Running the same executive summary through two models and keeping the better paragraph is a two minute habit that measurably raises the floor of what you send. It also reduces the "this sounds like AI" reaction, because you are shipping the best output for each section rather than one model's house style stretched across an entire document.

The third is the red team pass. Ask a model that did not draft the recommendation to argue against it: `Argue the strongest case that this recommendation is wrong. What would have to be true for the opposite conclusion to hold?` Whatever it finds, the client's CFO would have found too.

#### What AI will not do for you

It will not tell you what the client actually wants, which is usually different from what the RFP says and is only learnable from the room. It will not know that the COO who has to implement your recommendation blocked the same idea two years ago. It will not carry accountability for a number, so every figure that reaches a page is yours to verify.

It is also weak at the specific thing consultants are paid for: judgment under incomplete information with reputational consequences. A model will happily give you a confident answer from thin evidence, because confidence is cheap for it and expensive for you. Treat every output as a well read analyst's first draft, produced by someone who has never met the client and cannot be blamed.

Key points:

- A model-per-task routing table for framework, narrative, analysis, and research work
- Copy-ready prompts for 2x2s, storylines, executive summaries, and competitor scans
- A redaction rule set for client material, plus the checks that keep a partner from rejecting the deck

Checklist:

- Save your five most used frameworks as reusable prompts instead of rebuilding them each engagement
- Run every framework you build through the sceptical engagement manager critique prompt before a manager review
- Draft storylines as full assertions, then ask the model which assertions lack evidence
- Keep two paragraphs of your firm's best writing on hand as a voice sample for every tone pass
- Demand sources and mark unverified numbers before any figure reaches a page
- Redact client names and index figures before pasting, as a default habit
- Red team the final recommendation with a model that did not write it

FAQ:

Q: Is client data safe?
A: Whizi does not train on your conversations, and each provider’s data policy is available for review before you enable that model. Your client agreement is usually the stricter constraint, so redact names, index figures, and keep material non public information out of any AI tool. Firms under heightened confidentiality obligations should clear AI use with their risk team before the engagement starts.

Q: Can I export to PowerPoint?
A: Whizi generates structured outlines, headline sets, and tables that paste cleanly into PowerPoint or Google Slides, so the storyline and page content transfer as text you can drop into your firm template. Native .pptx export is on the roadmap. In practice most consultants want the thinking rather than the file, since the template and formatting are already fixed by the firm.

Q: Which model should I default to for slide writing?
A: Claude, for headline and body copy. It holds a consistent argument across a long document and needs the least editing to sound like a person rather than a template. Use GPT when the output has to obey a rigid structure, such as an issue tree, a table, or a set of variants in a fixed format.

Q: Will a partner be able to tell the deck was AI assisted?
A: They will notice topic headlines instead of assertions, adjectives without evidence, and paragraphs that would fit any client. Those are prompt problems, not model problems. Give the model your real evidence, force assertion headlines, run a tone pass against a voice sample from your own firm, and edit the result like an editor rather than a passenger.

Q: How much time does this actually save?
A: The consistent wins are the first draft of a storyline, the first pass through a long document, and competitive scans, which together tend to be several hours a week. The wins do not come from asking a model for the answer. They come from compressing the distance between knowing what you want to say and having a draft of it in front of you.

Q: Do I still need ChatGPT Plus or Claude Pro alongside Whizi?
A: Not for the workflows on this page. Whizi includes the GPT, Claude, and Gemini model families in one subscription, so the reason to keep a separate plan is a product specific feature such as a shared Custom GPT your team depends on. Compare the coverage directly on the [Whizi vs ChatGPT Plus](https://whizi.io/vs/whizi-vs-chatgpt-plus) page.

### The best AI tools for developers: compare models before shipping code

URL: https://whizi.io/use-cases/developers/
Updated: August 2026

Quick answer: For developers the best setup is more than one model, because they fail differently. GPT is fast and idiomatic on well-trodden implementation and strict structured output, Claude is stronger on subtle reasoning and unfamiliar architecture, and Gemini holds the largest context for big codebases. Hand the first answer to a second model to find the hole.

Developers rarely stick to one model. Whizi gives them one place to route coding, reasoning, debugging, and research without stacking separate tools.

#### Chat models and coding agents are different tools

Worth separating up front, because the two get conflated. An agentic coding tool lives in your editor or terminal, reads your repository, and writes files. A chat workspace is where you think: you paste a stack trace, argue about an approach, review a diff, understand a library you have never used, and draft the design doc.

Most developers end up using both, and the chat side is where model choice matters most, because you are reading the reasoning rather than the diff. That is also where paying for three separate subscriptions to compare three models stops making sense.

| What you are doing | Model tendency | Notes |
| --- | --- | --- |
| Hard reasoning: concurrency, a subtle race, an architectural trade-off | Claude and GPT differ meaningfully | Ask both. This is the case where a second opinion pays for itself |
| Implementation speed on well-trodden ground | GPT | Fast, idiomatic, good at boilerplate and conversions |
| Reading a large unfamiliar codebase or a long spec | Gemini | Largest context window, so more of the system fits at once |
| Explaining an error or a concept | Whichever framing lands | Different models explain differently, and that is the point |
| Strict structured output: config, JSON, schema | GPT | Most reliable at obeying a format exactly |

#### Debugging prompts that beat pasting the stack trace

Pasting an error and asking what is wrong produces a guess. The guess is often right, and when it is wrong you lose twenty minutes chasing a plausible fix for a problem you do not have. These prompts change the shape of the answer.

**Prompt: hypotheses before fixes**

`Here is the error, the code, and what I have already ruled out. Do not give me a fix yet. List the four most likely causes ranked by probability, and for each, the single cheapest check that would confirm or eliminate it. Error: [paste]. Code: [paste]. Already ruled out: [list].`

**Prompt: the bug that only happens sometimes**

`This fails intermittently, roughly [frequency], under [conditions]. Here is the relevant code and what I know about the environment. Enumerate the categories of intermittent failure that could produce this specific symptom (timing, ordering, resource exhaustion, external dependency, state leakage between runs, clock or timezone, caching). For each, say what evidence in what I have given supports or contradicts it, and what I should log to distinguish them.`

**Prompt: explain the fix before I take it**

`Explain why this fix works, what it does not fix, and what it could break. If the underlying cause is somewhere else and this is a symptom patch, say so directly.`

That last prompt catches the most expensive class of AI assistance: a change that makes the symptom disappear while the actual defect stays in the codebase.

#### Two models on the same problem, which is not a gimmick

When the answer is obvious, one model is enough. The technique earns its keep on the problems where you are not sure, and it works because the models fail differently rather than identically.

The useful pattern is not to ask both and pick the one you like. It is to ask one, then hand its answer to the other:

**Prompt: adversarial review of an answer**

`Another engineer proposed this solution to this problem. Find what is wrong with it: correctness under edge cases, concurrency, error handling, performance at [scale], or a simpler approach that was missed. If it is actually sound, say so plainly rather than inventing objections. Problem: [paste]. Proposed solution: [paste].`

Two outcomes, both useful. Either the second model finds a real hole, which you now know before merging, or it agrees, which is genuine evidence given it had every incentive to disagree. Compare that to iterating with the same model, which tends to agree with itself.

The same pattern applies to design decisions:

**Prompt: argue the other side**

`I am choosing [approach A] over [approach B] for [context and constraints]. Make the strongest case for B. What would have to be true about our constraints for B to be the correct choice, and is any of it true here?`

Whizi's side-by-side comparison exists exactly for this, and it is documented in [compare models side by side](https://whizi.io/docs/compare-models-side-by-side).

#### Code review and reading unfamiliar code

**Prompt: review a diff like a demanding reviewer**

`Review this diff. Categories, in order: correctness bugs, security issues, unhandled failure modes, race conditions, then style. For each finding give severity, the specific line, and why it matters here rather than in general. Do not comment on formatting. If the diff is fine, say so. Context: this codebase uses [stack and conventions]. Diff: [paste].`

**Prompt: understand a codebase you just inherited**

`Here are the main source files. Produce: the entry points, the data flow from request to response, the state that is shared and where it is mutated, the external dependencies and what happens when each is unavailable, and the three parts most likely to contain bugs based on complexity and coupling. Say explicitly what you cannot determine from what I gave you.`

That final instruction matters more than it looks. Models will happily describe the behaviour of a file you did not paste, inferred from its name. Forcing an explicit list of unknowns tells you what to go read.

**Prompt: write the test you would not have thought of**

`Write test cases for this function, focusing on inputs I probably have not considered: boundaries, empty and null, unicode, very large values, concurrent calls, and any implicit assumption in the implementation. For each test, state the assumption it is checking. Function: [paste].`

#### The failure modes that actually cost time

**Invented APIs.** Models confidently produce method names, parameters, and configuration keys that do not exist, especially for libraries that changed recently or are less common. The signature will look right. Check the actual documentation before you build on top of anything unfamiliar.

**Confidently wrong fixes.** There is no signal in the tone. A fix that dissolves your problem and a fix that introduces a subtle new one are delivered with identical confidence. Always ask what the change could break.

**Outdated patterns.** Training data skews toward the volume of code written about a framework, which is often the previous major version. If the answer feels like it is from a few years ago, it probably is. Say which version you are on, in the prompt.

**Silent scope creep.** Ask for a fix and you often get a refactor. Add `change as little as possible, and list every line you changed and why` to keep the diff reviewable.

**Security theatre.** A model can name the vulnerability classes in your code, which is genuinely useful for a first pass, but it is not an audit. It does not know your threat model, your deployment, or your data sensitivity.

#### Where this fits with the rest of your tooling

It does not replace your editor integration or your agentic coding tool. It replaces the three browser tabs where you were comparing answers, plus the two subscriptions that were needed to have those tabs open at once.

The practical setup most developers land on: one default model for quick questions, a second one you switch to when the first answer is unconvincing, and Gemini when you need to put a large amount of code or a long specification in front of a model at once. All in one thread, so the context you already established carries across the switch instead of being re-pasted.

For deeper coverage see [AI for coding](https://whizi.io/resources/ai-for-coding), the [coding-focused alternatives comparison](https://whizi.io/resources/chatgpt-alternatives-for-coding), and the [Claude coding prompt pack](https://whizi.io/templates/claude-coding-prompts). The mechanics of running that setup inside Whizi are in [write and debug code with multiple models](https://whizi.io/docs/write-and-debug-code-with-multiple-models).

Key points:

- Prompts for debugging, code review, and understanding an unfamiliar codebase
- Why comparing two models on the same problem is a genuine technique, not a gimmick
- The failure modes that cost real time: invented APIs, plausible wrong fixes, and outdated patterns

Checklist:

- Ask for ranked hypotheses and cheap checks before asking for a fix
- Hand the first model's answer to a second one and ask it to find the hole
- Always ask what a proposed fix could break, and whether it is a symptom patch
- State your language, framework, and version in the prompt to avoid outdated patterns
- Verify any unfamiliar API against the real documentation before building on it
- Add "change as little as possible and list every change" to keep diffs reviewable
- Use the large-context model when the question spans more code than fits in a normal prompt

FAQ:

Q: Why not just stay with one coding model?
A: For routine work, one is fine. The value shows up on the problems where you are genuinely unsure, because the models fail in different places rather than the same one. Handing model A's proposed solution to model B and asking it to find the flaw either surfaces a real problem before you merge, or gives you meaningful confirmation. Iterating with a single model mostly produces agreement with itself.

Q: Is this a replacement for an agentic coding tool?
A: No, they solve different problems. An agent lives in your repository and edits files. A chat workspace is where you reason: stack traces, design arguments, diff review, understanding an unfamiliar library, and drafting the design doc. Most developers use both, and model choice matters more on the chat side because you are evaluating the reasoning rather than the resulting diff.

Q: Which model is best for coding?
A: It depends on the task, which is the honest answer and the reason this page exists. GPT tends to be faster and more idiomatic on well-trodden implementation work. Claude tends to be stronger on subtle reasoning, unfamiliar architecture, and explaining why something behaves the way it does. Gemini wins when the question requires holding a large amount of code or specification at once. Comparing them on your own real problems for a week beats any benchmark.

Q: Can I paste proprietary code?
A: Whizi does not train on your conversations, and each provider’s data policy is available for review before you enable that model. Your employer’s policy is usually the binding constraint and it varies widely, so check it. Where restrictions apply, a practical approach is to reproduce the problem in a minimal example that contains the structure but none of the business logic, which frequently produces a better answer anyway.

Q: How do I stop it from rewriting everything?
A: Instruct it explicitly: `change as little as possible, preserve the existing structure and naming, and list every line you changed with a one line reason`. Unprompted refactors are the main reason AI suggestions become unreviewable, and constraining the diff makes the difference between a change you can reason about and one you have to re-read from scratch.

### AI tools for founders: research, growth, and product work from one stack

URL: https://whizi.io/use-cases/founders/
Updated: August 2026

Quick answer: Founders need AI across more categories than one subscription covers: market research in Gemini, positioning and investor updates in Claude, pricing and financial structure in GPT. Paying separately for all three runs about $60 per person per month. Start with the stranger test on your homepage and your last quarter of sales call notes.

Founders use AI across too many categories to justify fragmented subscriptions. Whizi gives them one workspace for strategy, execution, and creative work.

#### The founder problem is context switching, not capability

A founder's day crosses more categories than any other role. Before lunch you might price a plan, write a customer email, review a contract, debug a deploy, and draft the paragraph that goes to investors. Each of those wants different assistance, and none of them gets a full hour.

That pattern is why single-model subscriptions frustrate founders specifically. You feel the gap several times a day rather than once a month. The research answer is worse than Gemini would have given you, the investor paragraph is worse than Claude would have written, and the structured pricing table is worse than GPT would have produced, all inside the same afternoon.

| What you are doing | Model | Why |
| --- | --- | --- |
| Market sizing, competitor scans, category research | Gemini | Recent web material with sources, and enough context for long reports |
| Positioning, investor updates, customer emails | Claude | Writes like a person, holds a consistent argument, best at difficult tone |
| Pricing models, financial structure, spreadsheets | GPT | Reliable structured output and arithmetic you can check |
| Code review, debugging, technical decisions | GPT or Claude | Compare both when the first answer is unconvincing |
| Landing pages and campaign visuals | Claude plus Flux | Copy and imagery without a second creative subscription |

#### Positioning: the highest value thing to run through a model

Founders are too close to the product to hear how it sounds. A model is a cheap stand-in for a stranger, which is exactly the perspective you have lost.

**Prompt: the stranger test**

`Here is our homepage copy. You have never heard of us. Answer as that stranger: what does this company do, who is it for, what does it replace, and what would you still need to know before trying it? Then list every sentence you did not understand or did not believe. Copy: [paste].`

**Prompt: positioning alternatives**

`Generate four distinct positioning statements for this product. Distinct means different categories or different reference points, not different adjectives. For each: the category we would be competing in, who the buyer is in that framing, what becomes our main competitor, the strongest proof we would need, and what we give up by choosing it. Product: [description]. Customers we have: [describe]. What they say when they explain us to a colleague: [paste real quotes if you have them].`

That last input is the one founders skip and the one that matters most. How a customer describes you to a colleague is your real positioning. Everything else is what you wish it were.

**Prompt: pricing structure**

`Propose three pricing structures for this product. For each: the value metric, the tier boundaries, who is priced out, who gets a bargain, the failure mode when a customer grows, and what it signals about who we are for. Do not give me a single recommendation until I have reacted to all three. Context: [what the product does, who buys, what it saves them, what competitors charge].`

#### Customer research when you do not have a research function

You do have research inputs. They are sitting in your inbox, your support queue, your churn notes, and your sales calls. Nobody has read them all together, which is precisely the job a model is good at.

**Prompt: what the pipeline is telling you**

`Here are notes from [n] sales calls. Return: the objections in order of frequency with counts, the words customers use for the problem (verbatim, not paraphrased), what they compared us to, and every place where a deal stalled for a reason we did not record. Flag any theme appearing in fewer than three calls as a single observation rather than a pattern. Notes: [paste].`

**Prompt: churn autopsy**

`Read these cancellation reasons and support threads from churned accounts. Cluster them by root cause, not by stated reason, and explain the difference where they diverge. For each cluster: count, whether it was preventable at the point of sale, and the earliest signal we could have detected. Data: [paste].`

**Prompt: the interview guide**

`Write a 30 minute customer interview guide to test this hypothesis: [state it]. Rules: no leading questions, no questions about hypothetical future behaviour, and every question must be about something the person has actually done. Include the follow-up probe for each question.`

For the full research pipeline including sourcing and verification, see [AI for market research](https://whizi.io/resources/ai-for-market-research) and the [founder research stack](https://whizi.io/templates/founder-research-stack).

#### Investor updates and the writing you keep postponing

The monthly update is the highest-leverage document a founder writes and the one most often skipped, usually because the bad months are the hard ones to write. Claude is unusually good here, because the skill required is delivering difficult information without either burying it or catastrophising.

**Prompt: the monthly update**

`Draft our investor update. Numbers: [paste the metrics, including the bad ones]. What happened: [bullets]. What we learned: [bullets]. Asks: [list]. Rules: lead with the headline number whether it is good or bad, no adjectives that are not measured, explain the miss without excusing it, and make each ask specific enough that someone could act on it today. Length: under 500 words. Voice: direct, calm, not promotional.`

**Prompt: the hard paragraph**

`I need to tell investors [the bad news]. Write three versions: the shortest honest one, one that gives full context, and one that pairs it with the plan. For each, state what a reader would take away and what they would ask next. Do not soften the fact itself in any version.`

**Prompt: pre-mortem before a raise**

`Assume this fundraise failed. Write the five most plausible explanations using only what is in the deck and metrics below, ordered by likelihood. For each, state the diligence question that would surface it and whether we currently have a good answer. Materials: [paste].`

#### The cost argument, stated plainly

ChatGPT Plus, Claude Pro, and Gemini through Google One are roughly $20 each per month. A founder who wants all three capabilities is at $60 a month per person, and at the point where a second and third person need the same access, it becomes a real line item on a company that has no revenue.

Whizi covers the same model families in one subscription, with Starter below the price of any single one of those plans. The [savings calculator](https://whizi.io/tools/calculator) does the comparison against what you currently pay, and [AI subscription costs](https://whizi.io/resources/ai-subscription-costs) breaks down where the money actually goes across the market.

The less obvious saving is the image generation. A separate Midjourney or equivalent subscription is another line, and Flux and Stable Diffusion are included on Pro, which covers landing page imagery, social assets, and pitch visuals for most early stage companies.

#### Where founders get burned

**Treating a market size estimate as research.** A model will produce a confident TAM figure from nothing. Any number that goes in front of an investor needs a source you have opened, and investors are specifically good at asking where a number came from.

**Letting it write the vision.** Model output is the average of everything written about your category. Averaging is the opposite of a differentiated company. Use it to pressure test a position you hold, not to generate one you do not.

**Confusing speed with progress.** It is now possible to produce a deck, a landing page, a pricing model, and a launch plan in one day, none of which have been tested against a customer. The bottleneck was never document production.

**Pasting things you should not.** Term sheets, cap tables, unannounced acquisitions, and anything under NDA deserve care. Whizi does not train on your conversations and each provider's policy is reviewable, but describing a deal structurally without names and figures gets you the same analysis with none of the exposure.

Key points:

- The five founder jobs where AI is genuinely load-bearing, with the prompts for each
- Investor update, positioning, and customer research workflows you can run in an afternoon
- One subscription instead of three, which for a pre-revenue company is a real line item

Checklist:

- Run the stranger test on your homepage before you spend anything on acquisition
- Generate four positioning alternatives and react to all of them before choosing
- Feed six months of sales call notes into one thread and ask for objections with counts
- Draft the investor update on a schedule, including the bad months
- Pre-mortem the raise before diligence rather than after
- Source every number that reaches an investor
- Add up what you pay across AI subscriptions and compare it against one plan

FAQ:

Q: Is this only for technical founders?
A: No, and the non-technical case is arguably stronger. The workflows that pay off fastest are positioning, customer research synthesis, investor writing, and pricing structure, none of which are technical. Coding support is a bonus for founders who ship code themselves, but it is not where most of the value sits.

Q: What should I set up first?
A: Two things, in this order. Run the stranger test on your homepage, because it takes five minutes and frequently changes what you do next. Then paste your last quarter of sales call notes into one thread and ask for objections ranked by frequency with counts. Both use material you already have and both tend to produce something you did not know.

Q: Can it help with fundraising materials?
A: It is good at structure, at the narrative between slides, and at generating the diligence questions you have not prepared for, which is the most useful of the three. It is not good at your numbers. Every figure in a deck must come from your own model or a source you have opened, because investors ask where numbers come from and a fabricated market size is a memorable way to end a meeting.

Q: Why not just pay for ChatGPT Plus and be done?
A: Because the founder pattern is many small tasks across different categories, and the gap between the best model for a task and a general one shows up several times a day rather than once. Research is noticeably better in Gemini, investor writing in Claude, structured financial work in GPT. Paying separately for all three is around $60 per person per month, which is what Whizi consolidates.

Q: Is it safe to paste company confidential material?
A: Whizi does not train on your conversations and each provider’s data policy can be reviewed before you enable that model. Your own obligations are usually the tighter constraint, particularly around term sheets, employee matters, and anything covered by an NDA. Describing a situation structurally, with names removed and figures indexed, produces the same quality of analysis with far less exposure.

### AI tools for freelancers who bill by output, not by AI subscriptions

URL: https://whizi.io/use-cases/freelancers/
Updated: August 2026

Quick answer: Freelancers get the most from AI on the unpaid work: proposals, scoping, and chasing late invoices. Route client writing and difficult emails to Claude, research to Gemini, and structured output to GPT. Three separate plans at roughly $20 each cost $720 a year, which one workspace consolidates.

Freelancers need fast production, clear cost control, and one workspace that can switch between research, writing, code, and image tasks.

#### The freelance problem: too many task types, one person

A freelancer switches context more often than anyone inside a company. A single Tuesday can include a proposal, a client research call, a copy deliverable, an invoice chase, a data cleanup, and a mockup concept. None of those is big enough to justify a specialist tool, and together they are more than one subscription covers well.

The other constraint is margin. If you bill fixed scope or retainer, every recurring subscription comes directly out of your take. Three AI plans at roughly $20 each is $720 a year before you have earned anything, and it is the kind of cost that quietly stops being questioned.

| Client work | Model | Why |
| --- | --- | --- |
| Proposals, scope documents, difficult client emails | Claude | Best tone control, which is the entire skill in client communication |
| Deliverable copy, long form writing, editing | Claude | Least generic prose, holds a client voice across a document |
| Client and category research, competitor scans | Gemini | Recent web material with sources you can hand over |
| Data cleanup, spreadsheets, structured output | GPT | Reliable formats you can paste straight into a deliverable |
| Code, scripts, small automations | GPT or Claude | Compare both when the first answer looks fragile |
| Concepts, moodboards, campaign visuals | Flux, Stable Diffusion | Included, so no separate image subscription |

#### The unpaid hours: proposals, scoping, and chasing

Freelance economics are decided by the work you do not bill for. Proposals, scoping calls, revision negotiations, and payment chasing are all unpaid, all recurring, and all highly templatable, which makes them the best possible target.

**Prompt: the proposal**

`Draft a proposal for this project. Client brief in their words: [paste]. What I think they actually need: [your read]. My relevant experience: [bullets]. Structure: my understanding of the problem, what I would deliver, what is explicitly out of scope, timeline with dependencies on them, price presented as a range with what moves it, and next step. Rules: name the risk they have not mentioned, no flattery, no adjectives about my own work. Under 700 words.`

The out of scope section is the one that pays for itself. Most scope disputes are not disagreements, they are things nobody wrote down, and a model is good at listing the adjacent work a client will assume is included.

**Prompt: the scope guard**

`Read this brief and list everything a client would plausibly assume is included that I have not committed to. Then list the dependencies on the client that, if late, would delay delivery. Brief: [paste].`

**Prompt: the difficult email**

`Write an email that says [the difficult thing: the deadline moved, the scope grew, the invoice is 30 days late]. Tone: warm, direct, not apologetic, no hedging that invites negotiation on the fact itself. Keep the relationship. End with one specific next step and a date. Context: [paste].`

That prompt category alone is worth the subscription for most freelancers. The hard part of a late-invoice email is never the words, it is that you rewrite it four times because you are annoyed. A first draft you did not write emotionally is easier to send.

#### Producing client work faster without producing generic work

Clients notice AI copy immediately, and they notice it for the same reason their own team's AI copy is unusable: it has no specific knowledge in it. The fix is a per-client context block that you paste into the top of every prompt for that client.

A client context block contains: three paragraphs of their existing best writing, their banned words, their actual proof points with numbers, who their reader is and what that reader is sceptical of, and any claim their legal team has blocked. It takes 20 minutes to build per client and it improves everything you do for them afterwards.

**Prompt: build it from their material**

`Read these samples of the client's existing content. Describe their voice operationally: sentence length, formality, how they open, how they handle claims, what they never do. List ten words that appear repeatedly and ten that never appear. Samples: [paste].`

**Prompt: the deliverable**

`Write [deliverable] for [client]. Voice: [paste client context block]. Brief: [paste]. Evidence available: [paste]. Rules: no claim beyond the evidence given, every benefit names its mechanism, nothing that would fit their competitor unchanged. Then list any place you had to assume something I did not provide.`

**Prompt: the pass you must not skip**

`Before rewriting, name the three weakest things here: a claim with no proof, a sentence that would fit any company in this category, a benefit with no mechanism, a tone mismatch against the voice sample. Quote each one. Do not rewrite.`

Then edit it yourself. The billable skill is the judgment about what the client actually needs to say, and that has not been automated. What has been automated is the two hours between knowing that and having a draft.

#### What to tell clients about AI use

This comes up more every year and the answer that holds up is straightforward: be able to say what you did without discomfort.

- **Check the contract.** An increasing number of client agreements now include AI clauses, some of which prohibit it outright and some of which require disclosure. Read it before the project, not when a dispute starts.
- **Do not paste what you are not allowed to paste.** Client data under NDA, personal data belonging to their customers, and unreleased material deserve the same care as any other third party tool. Describing the situation structurally, with names removed, usually gets you the same result.
- **Own the output completely.** "The AI wrote it" is not a defence for a factual error, a fabricated statistic, or a plagiarised phrase in a deliverable you invoiced for. Verify every number and every claim.
- **Be careful with generated imagery in paid placements.** Usage rights, recognisable faces, and brand safety are the client's exposure as well as yours.

The reputational risk is not using AI. It is being unable to explain a claim in your own deliverable.

#### The margin argument

The direct saving is simple: ChatGPT Plus, Claude Pro, and a Gemini plan run around $20 each per month, and a freelancer who wants writing quality, research quality, and structured output quality genuinely wants all three. Whizi covers those model families in one subscription starting below the price of any single one, with image generation included on Pro instead of being another line.

Run your own numbers with the [savings calculator](https://whizi.io/tools/calculator), or read the market-wide breakdown in [AI subscription costs](https://whizi.io/resources/ai-subscription-costs).

The indirect saving is larger and harder to invoice for. Reclaiming two unpaid hours a week from proposals and client email is roughly 100 hours a year. At any freelance rate, that dwarfs the subscription question entirely.

Key points:

- Proposal, scoping, and client communication prompts that shorten the unpaid part of the job
- A model routing table for writing, research, code, and asset work across mixed clients
- The disclosure and confidentiality rules that keep AI use from becoming a client problem

Checklist:

- Build a client context block for each retained client and paste it into every prompt
- Template the proposal, including an explicit out of scope section
- Run the scope guard prompt on every brief before you quote
- Draft difficult client emails from a neutral first pass rather than from irritation
- Run the criticism prompt before any rewrite of a deliverable
- Read the AI clause in the contract before the project starts
- Verify every number in anything you invoice for
- Add up your current AI subscriptions and compare against a single plan

FAQ:

Q: Does this replace every specialist tool?
A: No, and it is not trying to. What it usually removes is the need to keep two or three general AI subscriptions active at once, plus a separate image generation plan. Specialist tools with deep integrations into a specific workflow, such as a dedicated design suite or an accounting product, stay where they are.

Q: Is this better for agencies or solo operators?
A: Solo operators and small teams get more out of it, because the value comes from covering many task types with one subscription and from reclaiming unpaid hours. An agency with dedicated researchers, writers, and designers has less overlap to consolidate, although the per-seat cost comparison still holds.

Q: Should I tell clients I use AI?
A: Check the contract first, since AI clauses are now common and some prohibit or require disclosure explicitly. Beyond that, the standard most freelancers land on is that you should be able to describe what you did without discomfort, and you remain fully accountable for every claim in the deliverable. Being unable to explain a number in work you invoiced for is the actual reputational risk.

Q: How do I stop deliverables from sounding generic?
A: Build a client context block: three paragraphs of their real writing, their banned words, their proof points with numbers, and who their reader is. Paste it into every prompt for that client. That single habit does more than any prompt technique, because generic output is almost always the result of a prompt that contained nothing specific to the client.

Q: Which model should I default to for client writing?
A: Claude, for anything the client or their audience has to be persuaded by, which covers proposals, deliverable copy, and difficult emails. Switch to GPT when you need many variants in an exact format or structured output such as tables. When a piece really matters, run both and keep the stronger half of each, which takes about two minutes.

### The AI workspace for marketers: briefs, drafts, and image generation in one tool

URL: https://whizi.io/use-cases/marketers/
Updated: August 2026

Quick answer: AI for marketers only produces usable copy once you give the model a brand voice file: real examples of your writing, banned phrases, proof points, and claims boundaries. Draft long form copy and email in Claude, briefs and variant sets in GPT, competitor scans in Gemini, and campaign imagery in Flux.

Whizi for marketers: write briefs in Claude, generate campaign imagery with Flux, analyze competitor positioning with Gemini, and ship faster with one subscription.

#### Why most AI marketing copy is unusable

It is not the model. It is that "write a landing page for our product" gives the model nothing that distinguishes your product from a thousand others, so it returns the average of all of them. That average has a recognisable texture: inflated adjectives, benefits with no mechanism, and a headline that would fit any competitor if you swapped the logo.

Everything on this page is downstream of one fix. Give the model three things before you ask for copy: how you sound, what is true, and what is forbidden. Everything else is technique.

| Task | Model | Why |
| --- | --- | --- |
| Long form copy, brand voice, email | Claude | Warmest register, least prone to hype, holds a voice across a long piece |
| Briefs, variant sets, structured output | GPT | Follows format constraints, generates n clean variants without drifting |
| Competitor and category research | Gemini | Recent web material with sources you can verify |
| Campaign imagery and social assets | Flux, Stable Diffusion | Included in Whizi, so no second image subscription |
| Choosing between two drafts | Both, side by side | The cheapest quality lift available |

#### Build the brand voice file once

This is a 30 minute task that improves every prompt you write afterwards. Create one document and paste it into the top of any copy prompt. It should contain:

- Three to five paragraphs of your actual best copy, unedited. Real examples beat any description of tone.
- A banned list: the specific words and constructions you never use. Most teams have them. "Revolutionary", "seamless", "game-changing", "unlock", "in today's fast-paced world".
- Your proof points, written as facts with numbers. "Cuts review time from three days to one for two pilot teams" is a proof point. "Saves time" is not.
- Your claims boundary: what legal will not let you say, and how you phrase the near version instead.
- Two or three sentences on the reader, specifically what they already believe and what they are sceptical of.

**Prompt: extract the voice from your own archive**

`Read these examples of our writing. Describe the voice in operational terms a writer could follow: sentence length pattern, level of formality, how we open, how we handle claims, what we never do. Then list ten words or phrases that appear repeatedly and ten that never appear. Examples: [paste 3 to 5 pieces].`

Save the output. That description plus the raw examples is the most valuable prompt asset a marketing team can own, and it is the difference between AI copy that needs a rewrite and AI copy that needs an edit.

#### Campaign brief to first draft

**Prompt: the brief**

`Draft a campaign brief. Product: [what it does, in plain language]. Audience: [specific segment, not "businesses"]. The belief we need to change: [what they currently think]. Evidence we can use: [proof points]. Channels: [list]. Constraint: [budget, timing, legal]. Return: the single message, three supporting angles, the objection each angle handles, what success looks like as a measurable outcome, and what we are deliberately not saying.`

The last field is the useful one. Naming what the campaign will not claim keeps the brief from becoming a list of everything good about the product, which is the standard way a campaign loses its edge before a word of copy is written.

**Prompt: messaging angles before polish**

`Generate five distinct positioning angles for this campaign. Distinct means they rest on different reasons to believe, not different wordings of the same reason. For each: the angle in one sentence, who it lands hardest with, the proof required to support it, and the strongest objection to it. Do not write finished copy yet. Brief: [paste].`

**Prompt: ad variants**

`Write [n] ad variants for [channel]. Character limits: [exact limits]. Voice: [paste brand voice file]. Angle: [chosen angle]. Each variant must lead with a different mechanism, not a different adjective. No claim beyond the proof points listed. Return as a table: variant, hook type, the proof point it uses, and the objection it pre-empts.`

Asking for the hook type and objection in the table is what makes the set testable. Twelve variants that vary only in phrasing teach you nothing when the test finishes. Twelve variants built on different mechanisms teach you which mechanism works.

#### Landing pages, email, and the edit that actually matters

**Prompt: landing page**

`Write a landing page for [product] targeting [segment]. Structure: hero (headline, subhead, one CTA), problem section in the reader's own language, how it works in three steps, proof, objection handling, and final CTA. Rules: the headline must state a specific outcome, every benefit must name its mechanism, no statistic that is not in the proof points, no more than one exclamation of enthusiasm in the whole page. Voice: [paste voice file]. Proof points: [paste].`

**Prompt: lifecycle email**

`Write a [type] email for [segment] whose last action was [behaviour]. Goal: [single action]. Tone: [from voice file]. Constraints: under [n] words, one call to action, subject line under 45 characters, no fake urgency, no "just checking in". Return three subject line options with the trade-off each one makes.`

**Prompt: the criticism pass, before any rewrite**

`Before rewriting, identify the three weakest things about this draft: a claim with no proof, a sentence that would fit any competitor, a benefit with no mechanism, or a tone mismatch against the voice file. Quote each one. Do not rewrite yet.`

That criticism prompt is the highest leverage habit in this whole workflow. Models will polish forever if you let them, producing smoother copy that says nothing new. Forcing an evaluation step before the rewrite gets you a different draft rather than a shinier one.

Then compare. Run the same brief through Claude and GPT side by side and keep the better opening, the better proof section, and the better close. That is two minutes of work and it reliably beats iterating on a single draft.

#### Campaign imagery without a second subscription

Image generation is included in Whizi on Pro and above, which for most marketing teams removes the Midjourney line item. The models that matter are Flux for photographic and product-adjacent work, and Stable Diffusion variants for stylised or illustrative treatments.

The prompt pattern that produces usable assets rather than pretty noise:

`[subject], [action or state], [setting], [lighting], [composition and framing], [lens or medium], [colour treatment], [mood]. Negative: [what must not appear].`

Worked example: `A mid-thirties operations manager reviewing a tablet in a warehouse mezzanine, morning light through high windows, shot from slightly below at eye level, 35mm, muted blues and warm greys, calm and competent. Negative: stock-photo smiling, cluttered background, visible logos, text.`

Practical notes from real campaign work:

- Generate at the aspect ratio you actually need. Cropping a square into a banner wastes the composition you asked for.
- Iterate one variable at a time. Change the lighting or the framing, not both, or you will not know what fixed it.
- Keep text out of generated images and add it in your design tool. Rendered text is still the weakest area across every image model.
- Check your usage rights and your brand rules before anything generated goes into a paid placement. Faces, recognisable locations, and anything resembling a real person deserve particular caution.
- Save the prompts that worked. A prompt library of eight reliable house styles is worth more than any single image.

#### Competitive and category monitoring on a schedule

This is the task most teams intend to do monthly and actually do never. It takes fifteen minutes with a web connected model and a saved prompt.

**Prompt: the monthly scan**

`Research what changed for [competitor list] in the last [period]. Cover: messaging and homepage positioning changes, pricing or packaging changes, new product or feature announcements with dates, funding or acquisition news, and any shift in the segments they appear to be targeting. Cite every claim with a URL and a date. Where you cannot verify something, say so rather than inferring.`

**Prompt: the positioning map**

`Using the scan above, place each competitor on two axes that genuinely differentiate this category (propose the axes and justify them). Then state where we sit, what space is currently uncontested, and what would have to be true for that space to be worth owning.`

For a deeper treatment of sourcing and verification, see the guide on [AI for market research](https://whizi.io/resources/ai-for-market-research).

#### Does the AI copy actually perform?

Worth being honest about. AI accelerates production, and production was rarely the bottleneck. What it changes is how many genuinely different angles you can afford to put into a test, which is where the performance gain comes from. Twelve variants of the same idea will not beat your control. Four variants resting on four different mechanisms might.

Two habits keep this from becoming volume for its own sake. First, tag every variant with the mechanism it uses, so the test result tells you something transferable. Second, feed the winner back into the voice file as a new example, so the next generation starts from what worked rather than from the average of the internet.

Key points:

- A brand voice file you build once and reuse in every prompt, which fixes most generic output
- Copy-ready prompts for briefs, ad variants, landing pages, lifecycle email, and competitive monitoring
- Image generation included, so campaign visuals do not need a second subscription

Checklist:

- Build the brand voice file once: real examples, banned words, proof points, claims boundary
- Paste the voice file into every copy prompt instead of describing your tone in adjectives
- Generate angles before copy, and make the angles differ by mechanism rather than wording
- Run the criticism prompt before any rewrite so you get a different draft, not a shinier one
- Compare Claude and GPT on the same brief and keep the better section from each
- Generate imagery at the final aspect ratio and add text in your design tool, not in the model
- Save a monthly competitor scan prompt and actually run it
- Feed winning variants back into the voice file

FAQ:

Q: Can I share templates with my team?
A: Yes, on Pro and team plans. In practice the highest value shared asset is not a clever prompt, it is the brand voice file: your real examples, your banned words, and your proof points. Share that and every prompt anyone on the team writes improves at once, including the ones they invent themselves.

Q: Does Whizi replace a separate image generation tool?
A: For most campaign work, yes. Flux and Stable Diffusion in Whizi cover product-adjacent photography, social assets, and illustration, which is the bulk of what a marketing team generates, and they are included rather than being a second subscription. A dedicated tool still wins for very specific stylistic control or for teams with a deep existing prompt library there.

Q: How do I stop AI copy from sounding like AI?
A: Three things, in order of impact. Paste real examples of your own writing instead of describing your tone. Ban the specific phrases that give it away, in the prompt. Run a criticism pass that names claims without proof and sentences that would fit any competitor, before you ask for a rewrite. Most of the tells are prompt problems, and they disappear when the model is given something specific to be faithful to.

Q: Is it safe to paste unreleased campaign material?
A: Whizi does not train on your conversations, and each provider’s data policy is available before you enable that model. The stricter constraint is usually your own agreements, particularly anything under embargo or covered by a client NDA. A workable habit is to describe the offer structurally and withhold names and dates until the material is public.

Q: Which model should write the actual copy?
A: Claude for anything a person has to be persuaded by, meaning long form, email, and brand voice work. GPT when you need many variants in an exact format, or structured output such as a table of headlines with metadata. The strongest habit is to run both on the same brief and keep the better half of each, which takes about two minutes and consistently beats iterating on a single draft.

### The AI workspace for product managers: specs, research, and narrative in one tool

URL: https://whizi.io/use-cases/product-managers/
Updated: August 2026

Quick answer: AI for product managers pays off most on reading, not writing: theming hundreds of pieces of feedback into counted themes takes minutes instead of a half day. Use Claude for PRD narrative, GPT for structured extraction and acceptance criteria, and Gemini for long transcripts and discovery. Ask a different model to critique the spec.

Whizi for PMs: write specs with Claude, run discovery with Gemini, analyze feedback with GPT, and stop juggling three AI subscriptions.

#### Where AI actually fits in a PM week

Product management is four different jobs sharing one calendar. You read a lot (feedback, tickets, transcripts, analytics exports), write a lot (specs, updates, briefs), analyze a bit (funnels, cohorts, survey results), and persuade constantly. Each of those rewards a different model, which is why a single AI subscription covers roughly two thirds of the work and leaves the rest feeling like a fight.

| PM task | Best model | Why |
| --- | --- | --- |
| PRD narrative, problem statements, product updates | Claude | Holds a long argument, writes prose an engineer will actually read |
| Feedback theming, ticket clustering, structured extraction | GPT | Reliable at strict output formats and consistent category labels |
| Discovery research, competitor scans, market context | Gemini | Best on recent web material, returns sources you can open |
| Long transcripts, research decks, 100 page reports | Gemini | Largest context window, so the whole corpus fits in one pass |
| Spec critique and edge case hunting | Any model that did not write the spec | An independent reader catches what the author cannot |

None of this replaces judgment about what to build. It compresses the distance between having the inputs and having something written down, which is where most PM weeks actually leak time.

#### Writing a PRD an engineer will not send back

Most AI written specs fail on the same thing: they describe a feature rather than a decision. Engineering does not need a paragraph about why the customer matters. It needs the states, the edge cases, and what happens when the call fails. Prompt for that explicitly and the output changes character.

**Prompt: problem statement first**

`Write the problem statement section of a PRD. Evidence I have: [paste support tickets, analytics, interview quotes]. Do not propose a solution. Return: who has the problem, how often, what they currently do instead, what it costs them, and what we would expect to change if it were solved. Mark any claim not supported by the evidence I pasted as ASSUMPTION.`

**Prompt: the spec body**

`Turn this into a specification for an engineering team. Feature: [description]. User states to cover: [list]. Return: user stories with acceptance criteria, every state including empty, loading, error, and permission denied, the behaviour when a dependency is unavailable, analytics events with their properties, and open questions. Do not invent requirements I did not state. List anything you had to assume in a separate section at the end.`

**Prompt: the critique pass**

`Act as a staff engineer reviewing this spec before estimation. List only the problems: undefined behaviour, missing states, requirements that conflict, hidden migration work, and anything that will produce a follow-up question in refinement. Do not rewrite the spec.`

Run that third prompt in a model that did not write the spec. It reliably surfaces the three questions your team would otherwise raise in refinement, and answering them beforehand is the difference between a 20 minute grooming session and a 50 minute one.

#### Turning raw feedback into something you can prioritize

The highest leverage AI task in product management is not writing. It is reading 400 pieces of feedback in a form you can act on. Done manually this is a half day. Done well with a model it is 20 minutes, and the quality depends almost entirely on whether you force stable categories.

**Prompt: first pass theming**

`Here is raw customer feedback. Cluster it into themes. For each theme return: a label, the count of items, the severity implied by the language, a representative verbatim quote copied exactly, and whether the theme is a bug, a missing capability, a usability problem, or an expectation mismatch. Do not merge themes that have different root causes even if the wording is similar. Do not paraphrase quotes. Feedback: [paste].`

**Prompt: second pass against a fixed taxonomy**

`Re-classify the same feedback using only these categories: [paste your existing taxonomy]. Anything that does not fit goes into UNCLASSIFIED with an explanation. Return a table of category, count, and percentage.`

The two pass structure matters. The first pass tells you what is actually in the data. The second makes the result comparable to last quarter, which is what makes it usable in a prioritization conversation rather than just interesting.

| What to ask for | What you get | What it is good for |
| --- | --- | --- |
| Themes with counts | A ranked list of problem areas | Roadmap input, quarterly planning |
| Verbatim quotes only | Unedited customer language | Copy, positioning, exec persuasion |
| Severity and frequency split | A 2x2 of pain against volume | Deciding what to fix first |
| Contradictions | Where segments want opposite things | Catching a false consensus early |

That last row is worth a standing prompt: `Where in this feedback do different users want incompatible things? Name the segments and the tradeoff.` A theme list flattens disagreement, and disagreement is usually the most useful thing in the data.

#### Discovery, competitors, and the research you never have time for

Discovery is the work that gets cut first when a release is late, which is precisely when a bad decision is most expensive. Model assisted scanning does not replace talking to customers, but it does replace the excuse for going into a decision blind.

**Prompt: competitor teardown**

`Build a teardown of how [competitor] handles [job to be done]. Cover: their stated positioning in their own words, the flow as documented in their help centre, pricing where it is public, what changed in the last 12 months with dates, and the complaint themes visible in public reviews. Cite every claim with a URL. Separate what the company states from what third parties observe.`

**Prompt: interview synthesis**

`Read these interview transcripts. Return: the jobs users are trying to accomplish, the workarounds they have built, the moments where they expressed frustration with exact quotes, and any place where what a user said contradicts what they described doing. Do not generalise beyond the transcripts. If a pattern appears in fewer than three interviews, label it as a single observation rather than a pattern.`

That last constraint is the one PMs most often forget. Models are eager to produce clean patterns, and a clean pattern from two interviews is how a roadmap ends up serving a customer who does not exist. Ask for counts alongside every claim.

#### Executive narrative and launch comms

The same content has to exist at four altitudes: a spec for engineering, an update for the team, a paragraph for the leadership review, and a launch note for customers. Rewriting between altitudes is the most mechanical work in the job and the easiest to hand off.

**Prompt: altitude change**

`Rewrite this for [audience]. They care about [specific concerns]. They have [level] context on this product area. Keep every factual claim identical. Length: [constraint]. Lead with the decision or outcome, not the background. Draft: [paste].`

**Prompt: the leadership paragraph**

`Compress this update into 120 words for an executive who will read it once. Structure: what changed, what it means for the metric we committed to, what we need from them, and the single risk worth their attention. No adjectives that are not measured.`

**Prompt: the pre mortem**

`Assume this launch failed six months from now. Write the three most plausible explanations, ordered by likelihood, using only what is in the plan below. For each, state the early signal we could watch for. Plan: [paste].`

Keep all four altitudes in the same Whizi thread. The launch note inherits the context from the spec and the feedback analysis, so you stop re-explaining the feature every time you change audience.

#### Where AI misleads product managers specifically

Three failure modes matter more in this job than in most others.

**Fabricated verbatims.** If you ask for representative quotes without pinning the model to the source text, you will sometimes get a plausible sentence no customer said. Always instruct `copy quotes exactly, do not paraphrase`, and spot check three of them against the raw data before any quote reaches a slide.

**False confidence from small samples.** A model will theme eight support tickets with exactly the same assurance it themes eight hundred. Ask for counts on every theme, and treat anything under a handful of instances as an observation rather than a signal.

**Roadmap theater.** Asking a model to prioritize your backlog produces a confident ranking derived from nothing but the words in your tickets. It has no access to your strategy, your capacity, your technical debt, or the deal that closes next quarter. Use it to structure the tradeoff, never to make the call.

Key points:

- A model-per-task map for specs, feedback analysis, discovery, and executive narrative
- Copy-ready prompts for PRDs, feedback theming, competitive teardowns, and launch briefs
- The failure modes that matter for PMs, including fabricated user quotes and false confidence in small samples

Checklist:

- Save a PRD prompt template in Claude and a critique prompt to run in a different model
- Save a feedback theming template in GPT with your existing taxonomy pasted in
- Save a competitor scan template in Gemini that demands a URL for every claim
- Always ask for counts alongside themes, and treat small counts as observations
- Instruct the model to copy verbatims exactly, then spot check three against the source
- Run a pre mortem on every launch plan before the launch review, not after
- Keep spec, feedback, and launch comms in one thread so context carries across

FAQ:

Q: Can I paste customer interviews?
A: Yes. For long transcripts use Gemini, which has the largest context window of the models in Whizi, so a full set of interviews fits in one pass rather than being chunked. Strip names, emails, and company identifiers first. Roles and segments are all the analysis needs, and removing the rest keeps you clear of most internal data policies.

Q: Does Whizi integrate with Jira or Linear?
A: Not natively yet. In practice the workflow is to generate the structured output in Whizi (user stories with acceptance criteria, a table of themes with counts) and paste it into your tracker, which takes seconds because the format is already what the tracker expects. Ask for the output as a Markdown table or as one issue per block if you want to paste them individually.

Q: Which model writes the best PRD?
A: Claude for the narrative sections, meaning the problem statement, the rationale, and anything a human has to be persuaded by. GPT for the structured sections, meaning user stories, acceptance criteria, state tables, and analytics event definitions. Splitting the document between the two takes one extra model switch and noticeably reduces the editing pass.

Q: Is it safe to paste internal roadmap or revenue data?
A: Whizi does not train on your conversations, and each provider’s data policy is available before you enable that model. Your company policy is usually the stricter constraint. A reliable habit is to index sensitive figures rather than pasting absolutes, since analysis of relative movement works identically and the numbers stop being sensitive.

Q: Can AI prioritize my backlog?
A: It can structure the tradeoff, which is genuinely useful: score items against criteria you define, surface where two items depend on each other, and show which segments a given choice serves. It cannot make the call, because it has no visibility into your strategy, your team capacity, or the commercial context. Treat any ranking it produces as a prompt for discussion.

### The AI workspace for researchers: long-context reading plus cited writing

URL: https://whizi.io/use-cases/researchers/
Updated: August 2026

Quick answer: AI for researchers splits into three jobs no single model does best: Gemini holds 15 to 30 papers in one context window, GPT extracts structured evidence tables, and Claude drafts prose that preserves academic hedging. Upload every source rather than relying on model memory, and resolve every DOI yourself, because fabricated references are the serious failure mode.

Whizi for researchers: upload long PDFs to Gemini, draft lit reviews in Claude, and get structured data back from GPT, all in one workspace.

#### The three model jobs in a research pipeline

Research is the clearest case for a multi-model workspace, because the pipeline has three genuinely different demands. You need a model that can hold an enormous amount of text at once, a model that writes carefully enough to survive peer review, and a model that returns rigid structured output without drifting. No single product is best at all three, and the switching cost between three products is paid in re-uploading the same corpus.

| Stage | Model | What it is doing |
| --- | --- | --- |
| Reading a corpus of papers | Gemini | Largest context window, so 15 to 30 PDFs sit in one conversation |
| Building an evidence table | GPT | Strict schemas, consistent column values, no editorial drift |
| Drafting prose and synthesis | Claude | Careful hedged academic register, faithful to the source |
| Finding recent work and checking references | A web connected model | Retrieval against live sources rather than memory |
| Adversarial review of your own argument | Any model that did not draft it | Independent reader, no sunk cost |

The single most important thing to understand before you start: a language model asked about a paper from memory will invent details, and it will do so fluently. A model reading a PDF you uploaded is working from the actual text. Those are different activities with different reliability, and almost every horror story about AI in research comes from the first being mistaken for the second. Upload the source. Always.

#### Reading 20 papers without losing the thread

The mistake is asking for a summary. A summary of 20 papers is a paragraph that could describe any 20 papers in the field. What you actually want is an evidence table, because a table forces the model to commit to a specific value per paper and makes the gaps visible.

**Prompt: the evidence table**

`I have uploaded [n] papers. Build a table with one row per paper and these columns: citation (authors, year), research question, study design, sample size and population, primary outcome measured, headline finding with the effect size or key statistic, stated limitations, and funding source if disclosed. If a field is not reported in a paper, write NOT REPORTED. Do not infer values. Do not include any paper I did not upload.`

**Prompt: the synthesis pass**

`Using only the table you just built, answer: where do these studies agree, where do they disagree, and what explains the disagreement (design, population, measurement, or period)? Identify the questions in this literature that no uploaded study addresses. For every claim, cite the specific rows it rests on.`

**Prompt: methodological critique**

`For each study in the table, note the most serious threat to validity: selection, measurement, confounding, statistical power, or generalisability. Be specific to what the paper reports, not generic. If a study is methodologically strong, say so rather than inventing a criticism.`

That last instruction is doing real work. Ask a model to find problems and it will always find problems, including in a well conducted study, because the request implies they exist. Giving it explicit permission to say a paper is sound keeps the critique honest.

For the mechanics of getting long documents into a chat cleanly, see [how to chat with a PDF](https://whizi.io/docs/chat-with-pdf) and the guide to [summarizing PDFs with AI](https://whizi.io/resources/summarize-pdf-with-ai).

#### Extraction: getting structured data out of unstructured text

Systematic extraction is the task where AI saves the most defensible time, because the output is verifiable. You can check a number against page 7. You cannot check a vibe.

**Prompt: strict schema extraction**

`Extract every reported statistic from this paper as JSON. Schema: {"value": number, "unit": string, "measure": string, "population": string, "location_in_paper": string, "confidence_interval": string or null, "p_value": string or null}. Include only figures that appear in the text, tables, or figure captions. Do not compute derived values. Do not round. If a figure appears in both a table and the text with different values, return both entries and flag the discrepancy.`

The `location_in_paper` field is the part people leave out, and it is the part that makes the output usable. It turns a verification pass from re-reading the paper into a targeted check, which means the verification actually gets done.

**Prompt: comparing across papers**

`Across the uploaded papers, find every reported estimate of [the quantity you care about]. Return a table of: paper, estimate, unit, population, method of estimation, and year of data collection. Do not average them. Note explicitly where the estimates are not comparable because of differing definitions or populations.`

Instructing the model not to average is not pedantry. Averaging incomparable estimates is the single most common way an AI assisted literature review produces a number that is confidently wrong, and it is the kind of error a reviewer catches immediately.

#### Drafting the review without borrowing anyone else's sentences

Claude is the best drafter here, but the prompt has to enforce the two rules that matter in academic writing: claims stay attached to sources, and hedging is preserved. Models trained to be helpful tend to strengthen careful language, turning "was associated with" into "caused" without being asked.

**Prompt: section draft**

`Draft the [section name] of a literature review using only the evidence table above. Rules: every claim must be attributable to a specific row, preserve the original hedging (do not upgrade an association to a causal claim), use [citation style] format, and mark any sentence where you were tempted to generalise beyond the evidence with [CHECK]. Audience: [field] researchers. Length: approximately [n] words.`

**Prompt: the hedging audit**

`Review this draft against the source material. List every sentence that states more certainty than the underlying study supports, and every sentence that attributes a finding to the wrong source. Quote the sentence and the conflicting source text side by side. Do not rewrite.`

Run the hedging audit on your own writing too, not just on model output. It catches the same drift in a human draft written at 11pm.

#### Citation verification, which is not optional

This is the section that matters most, so it is deliberately blunt. Language models fabricate references. They produce citations with real author names, plausible journal titles, correct-looking volume and page numbers, and DOIs that resolve to nothing or to an unrelated paper. Retracted papers have entered the literature this way. Reputations have been damaged by it.

The rules that prevent it:

- Never ask a model to "find sources" for a claim from memory. Ask a web connected model to retrieve them, then open every link yourself.
- Verify every DOI by resolving it. A DOI that does not resolve is a fabrication, not a typo.
- Check that the cited paper actually says what the citation claims it says. A real paper attached to a claim it does not support is the harder error to catch, and it is more common than an invented paper.
- Never cite something you have not opened. This rule existed before AI and AI has made it load bearing.
- Keep the uploaded PDFs as the source of truth. Anything the model tells you about a paper you did not upload is unverified by construction.

**Prompt: reference audit**

`For each reference in this list, state: whether you can confirm it from a source in this conversation, or whether it comes from your training data and is therefore unverified. Do not guess. Mark every unverified entry clearly. I will check them manually.`

That prompt does not make the model reliable. It makes it explicit about which claims are anchored to something you provided, which is the distinction you need in order to know what to check.

#### What this does not replace

It does not replace a systematic review protocol. If your work requires PRISMA, the search strategy, screening, and inclusion decisions remain yours to document and defend. AI can help you draft the protocol and speed up extraction from included studies; it cannot be the search.

It does not replace your reference manager. Keep Zotero, Mendeley, or EndNote as the record of what you have actually read. The chat is working memory, not an archive.

It does not replace statistical analysis. A model can explain a method, write the R or Python for it, and sanity check your interpretation, but running your analysis inside a chat window and trusting the arithmetic is not defensible. Generate the code, run it in your own environment, and check the output.

And it does not replace disclosure. Journals and institutions increasingly require a statement of how AI tools were used. Write down what you used at each stage as you go, because reconstructing it at submission is unpleasant.

Key points:

- A reading pipeline for 10 to 30 papers that produces an evidence table, not a vague summary
- Extraction prompts that return structured rows you can paste into a spreadsheet
- A citation verification routine, because fabricated references are the failure mode that ends careers

Checklist:

- Upload the papers rather than asking the model what it remembers about them
- Build an evidence table with one row per paper before asking for any synthesis
- Require a location_in_paper field on every extracted figure so verification is targeted
- Instruct the model to preserve hedging and never upgrade association to causation
- Resolve every DOI and open every source before it enters your reference list
- Ask the model to flag which claims are anchored to uploaded text and which are from memory
- Keep a running note of how AI was used at each stage for your disclosure statement

FAQ:

Q: Can Whizi handle very long papers?
A: Yes. Gemini in Whizi carries the largest context window of the commercial models, which in practice means a long dissertation, a full regulatory filing, or a set of 15 to 30 papers can sit in one conversation and be queried as a whole. That is different from summarizing a document a chunk at a time, because cross-document questions such as "where do these disagree" only work when everything is present at once.

Q: Does Whizi cite sources?
A: When you use a web connected model it returns links, and when you upload documents it can point to where in your source a claim came from. Neither is a substitute for verification. Resolve every DOI, open every link, and confirm that the cited work supports the claim attached to it. Fabricated and misattributed references are the most serious failure mode of AI assisted research, and the only reliable defence is opening the source yourself.

Q: Why not just use one model for everything?
A: Because the three stages want opposite behaviour. Reading a large corpus needs an enormous context window. Extraction needs rigid, boring adherence to a schema. Drafting needs careful prose that preserves academic hedging. Running all three inside one workspace means the corpus is uploaded once and the context carries across the switch, instead of being re-explained to three separate products.

Q: Is it acceptable to use AI in academic work?
A: Policy varies by journal, funder, and institution, and it is changing quickly, so check the specific requirements that apply to you. The broad pattern is that using AI for extraction, drafting, and editing is generally acceptable with disclosure, while listing a model as an author is not, and you remain fully accountable for every claim and citation in the finished work. Record what you used as you go.

Q: Can I upload unpublished or confidential data?
A: Whizi does not train on your conversations and each provider’s policy is reviewable before you enable that model, but your ethics approval and any data sharing agreement are the binding constraints. Human subjects data almost always needs to stay out unless your approval explicitly covers third party processing. De-identified extracts and your own drafts are the normal safe cases.

### The AI workspace for students: Claude, GPT, and Gemini for one student price

URL: https://whizi.io/use-cases/students/
Updated: August 2026

Quick answer: AI for students helps most when the model reacts to your thinking instead of doing it. Ask for study guides with questions and no answers, explain concepts from memory and get the gaps named, and request rubric feedback rather than rewrites. Use Gemini to read, GPT to structure, and Claude to critique.

Whizi for students: one subscription that covers note summarization, essay drafts, research help, and coding projects across Claude, GPT, and Gemini.

#### The honest version of what AI is good for at university

Two things are true at once. AI is the most useful study tool that has ever existed, and using it the obvious way will make you worse at your subject. The difference is entirely in whether the model does the thinking or interrogates yours.

Asking for an essay produces a passable essay and zero learning. Asking a model to quiz you on a chapter until you can explain it without notes produces no artifact at all and a genuine grasp of the material. Both take the same amount of time. Only one of them helps in the exam, where you will not have it.

Everything below is built around that distinction. The workflows that hold up are the ones where you generate first and the model reacts, rather than the reverse.

| What you need | Model | Why |
| --- | --- | --- |
| Summarizing readings, slides, long PDFs | Gemini | Largest context, so a whole textbook chapter or lecture set fits at once |
| Essay structure, counterarguments, outlines | GPT | Strong at explicit structure and at holding an argument to a brief |
| Prose, feedback on your writing, tone | Claude | Best writing quality and the most useful critique of a draft |
| Debugging code assignments | GPT or Claude | Both explain errors well; compare them when one is unconvincing |
| Active recall and practice questions | Any of them | Quality depends on the prompt far more than the model |

#### Turning readings into something you will actually revise from

A summary you did not write is a summary you will not remember. The fix is to make the model produce revision material that requires you to do the retrieval.

**Prompt: the study guide**

`I have uploaded [the reading]. Produce: (1) the five to eight core concepts with a one sentence definition each, in the source's own terminology; (2) how the concepts connect to each other; (3) the three points a first-time reader is most likely to misunderstand; (4) ten questions that test understanding rather than recall, without answers. Do not include anything not present in the document.`

Withholding the answers is the whole trick. Answer the ten questions yourself first, then paste your answers back with: `Mark these against the source. For each, say what is correct, what is missing, and what is wrong. Do not be generous.`

**Prompt: the Feynman check**

`I am going to explain [concept] in my own words. Identify every place where my explanation is vague, circular, or wrong, and ask me one follow-up question that would expose the gap. Do not explain it for me. My explanation: [type it out from memory].`

That prompt is worth more than any summary on this page. Explaining from memory and getting a targeted objection back is close to what a good tutorial with a supervisor does, and you can run it at midnight the day before.

**Prompt: exam-shaped practice**

`Generate [n] exam questions on [topic] at the level of [course and year]. Match the style of these past questions: [paste two or three]. Do not provide answers. After I answer, mark against the criteria that would apply, state the grade band, and give the single change that would move it up one band.`

#### Essays: where the line actually is

Submitting model-written prose as your own is plagiarism at essentially every institution, and it is also the least useful thing you can do with the tool. Everything short of that is a spectrum, and most departments have published where they draw it. Read your own department's policy, because they differ more than students assume.

What tends to be both permitted and genuinely useful:

- **Interrogating your thesis before you commit.** `Here is my argument. What is the strongest objection to it, what evidence would I need to defeat that objection, and is there a version of this claim that is more defensible?`
- **Structure from your own material.** Give it your notes and your argument, and ask for two or three possible structures with the trade-off each makes.
- **Feedback on your draft.** `Mark this against these criteria: [paste rubric]. For each criterion, quote the strongest and weakest passage and say specifically why. Do not rewrite anything.`
- **Finding the gap.** `Which claims in this draft are asserted without support? Quote each one.`
- **Explaining a source you are stuck on.** Upload it and ask for the argument structure, then check the explanation against the text yourself.

What to avoid regardless of policy: generating paragraphs and editing them into your voice (that is still not your argument), and asking for sources without opening every one. Models fabricate citations that look completely real, with plausible authors, journals, and page numbers. A fabricated reference in a bibliography is caught easily and treated seriously.

**Prompt: the rubric pass**

`Here is my essay and the marking criteria. For each criterion: the grade band this draft currently sits in, the exact passage that is holding it there, and the single most efficient change to move it up. Be blunt. Do not rewrite my sentences. Essay: [paste]. Criteria: [paste].`

#### Coding assignments, without breaking the point of the assignment

The temptation is to paste the problem and copy the answer. That works until the exam or the technical interview, at which point the gap is unpleasant. The alternative is barely slower and actually teaches.

**Prompt: explain the error, do not fix it**

`Here is my code and the error. Explain what the error means, why it is happening in this specific code, and what category of mistake it is. Do not give me the corrected code. Ask me what I think the fix is.`

**Prompt: review after you have it working**

`This works. Review it as a demanding TA would: correctness edge cases, complexity, naming, and anything that would lose marks on style for [language and course]. Rank the issues by how much they matter. Do not rewrite it.`

**Prompt: the concept behind the bug**

`I keep making this mistake: [describe]. What is the underlying concept I have not understood? Explain it, then give me three small exercises that would test whether I have it, without solutions.`

When one model's explanation does not land, switch and ask the other. Claude and GPT explain the same concept with genuinely different framings, and the second framing is often the one that clicks. That single habit is the most concrete argument for having more than one model available.

#### What it costs, and the honest comparison

ChatGPT Plus is about $20 a month for the GPT family alone. Claude Pro is another $20. Gemini through Google One AI Premium is another $20 again. Paying for all three as a student is not reasonable, and picking one means losing the specific thing the other two are better at, which for coursework is exactly the point.

Whizi's Starter plan sits below the price of a single one of those and covers GPT, Claude, Gemini, and open models together, with Pro adding image generation and higher limits. The [savings calculator](https://whizi.io/tools/calculator) will do the arithmetic against whatever you are currently paying.

The free tiers of each product are a legitimate option too, and worth saying plainly: if your usage is light, free accounts across two providers will cover a lot. What you hit is rate limits during exactly the weeks you need it most, which is the week before a deadline.

#### Six habits that separate students who get better from students who get faster

- Always attempt first, then ask for critique. Generating before the model does is the difference between practice and consumption.
- Ask for questions rather than answers, and answer them before revealing anything.
- Upload the source instead of asking what the model remembers about it. Memory-based answers about specific papers, cases, and texts are unreliable.
- Never submit a citation you have not opened.
- Switch models when an explanation does not land, rather than re-reading the same one.
- Write down what you actually did. Many courses now ask for an AI use statement, and reconstructing it later is worse than noting it as you go.

Key points:

- Study workflows that build understanding instead of outsourcing it, with the prompts written out
- One subscription across essays, revision, research, and coding assignments
- A straight answer on academic integrity, and where the real line sits

Checklist:

- Pick your hardest module and upload the core reading for it this week
- Ask for a study guide with questions and no answers, then answer them from memory
- Run the Feynman check on the concept you are least confident about
- Get rubric-based feedback on your next essay draft instead of asking for a rewrite
- Ask for explanations of code errors rather than corrected code
- Check your department's AI policy before the deadline, not after
- Open every source before it enters a bibliography

FAQ:

Q: Is there a student discount?
A: Whizi’s Starter plan is already priced below a single ChatGPT Plus seat while covering GPT, Claude, and Gemini together, which is where most of the saving comes from for a student. Check the [pricing page](https://whizi.io/pricing) for current promotions and any student offer running at the moment, and the [calculator](https://whizi.io/tools/calculator) if you want the comparison against what you already pay.

Q: Can I use Whizi without internet for exam prep?
A: You need a connection to chat with the models, since the models run on the providers’ infrastructure rather than on your device. The practical workaround is to generate what you need in advance and export it: study guides, practice question sets, and worked explanations all copy out as text you can revise from offline.

Q: Will professors know I used AI?
A: Detection tools exist and they are unreliable in both directions, producing false accusations and missing real cases, which is why most institutions have moved toward disclosure requirements rather than detection. That is the practical answer, but not the important one. The important one is that submitting model-written work is plagiarism at nearly every institution regardless of whether it is detected, and it leaves you without the understanding the assessment was measuring. Whizi is not a detection-evasion tool and there is nothing on this page that is.

Q: Which model is best for essays?
A: Claude, for feedback on your writing and for the clearest critique of an argument. GPT is better when you want structure, counterarguments, or an outline held tightly to a brief. Use Gemini for the reading stage, since it can hold an entire set of course materials in one conversation. The workflow most students settle on is Gemini to read, GPT to structure, Claude to critique.

Q: Can it read my lecture slides and textbook chapters?
A: Yes. Upload PDFs, slide decks, and documents directly, and Gemini in particular can handle a very large amount at once, so a full set of lecture slides or several chapters can sit in a single conversation. Always upload the source rather than asking what the model already knows about a text, because answers drawn from memory are much less reliable than answers drawn from the document in front of it.

### How to summarize meetings with AI in Whizi

URL: https://whizi.io/use-cases/summarize-meetings-with-ai/
Updated: August 2026

Quick answer: To summarize a meeting with AI, stop asking for a summary and ask for the four things a meeting produces: decisions, commitments, open questions, and disagreements. Run the recap in Claude, then extract action items in GPT as a table with owner, due date, and the exact supporting quote.

Paste a meeting transcript into Whizi, get a clean summary in Claude, and extract action items with GPT. A ten-minute workflow that replaces notetakers.

#### Why the default summary is useless

Paste a transcript, ask for a summary, and you get a paragraph that says the team discussed the project and agreed to follow up. It is accurate and it is worthless, because it strips out the only two things anyone opens meeting notes for: what was decided and who owes what.

The fix is to stop asking for a summary and start asking for the four things a meeting actually produces. Decisions. Commitments. Open questions. Disagreements. That last one gets lost the most often and matters the most later.

Two models, two passes. Claude reads the room and writes the narrative, GPT extracts the structure without editorialising. Splitting the job takes one model switch and produces notes people actually read.

#### The two prompts

**Prompt 1: the recap, in Claude**

`Here is a meeting transcript. Produce: (1) a one-line headline stating what this meeting actually resolved; (2) the decisions made, with who made each one; (3) five to eight bullets covering the substance, in the order it matters rather than the order it was said; (4) open questions that were raised and not answered; (5) any point where people disagreed, stating both positions without resolving them. Do not include pleasantries, scheduling talk, or anything discussed and abandoned. If a decision was implied but never stated explicitly, list it separately under UNCONFIRMED. Transcript: [paste].`

**Prompt 2: the action items, in GPT**

`From the same transcript, extract every commitment as a table: task, owner, due date, and the exact quote it came from. Rules: only include items where someone actually committed, not things that were merely suggested. If the owner is ambiguous, write UNASSIGNED rather than guessing. If no date was stated, write NO DATE rather than inferring one. Then list separately any task that was raised and never picked up by anyone.`

The quote column is what makes this trustworthy. It turns a disputed action item from an argument into a lookup, and it makes fabricated commitments immediately obvious because there is nothing to quote.

**Prompt 3, optional: the follow-up email**

`Draft a follow-up email from the recap and action items above. Audience: [who was and was not in the room]. Lead with the decision. Keep it under 200 words. State each person's action item explicitly so nobody has to interpret. Tone: [your voice]. Flag anything I should confirm before sending.`

#### Templates for the meetings you actually have

Different meetings produce different artifacts. Adapt the extraction, not just the wording.

| Meeting type | Ask for | The thing people forget |
| --- | --- | --- |
| Customer or sales call | Objections raised, verbatim language for the problem, what they compared you to, next step and who owns it | Copy the customer's exact words rather than paraphrasing them |
| Decision meeting | The decision, the alternatives rejected and why, who has to be told | Recording the rejected options, which is what prevents relitigating it in six weeks |
| Standup or status | Blockers with owners, changes to previous estimates, anything that slipped silently | Slippage that nobody said out loud |
| User interview | Jobs they described, workarounds they built, exact frustration quotes, contradictions between what they said and did | Contradictions, which are the most useful part |
| Interview or hiring panel | Evidence against each competency, with quotes, and questions left unasked | Separating evidence from impression |
| Board or investor meeting | Commitments made by you, questions you could not answer, follow-ups requested | The questions you could not answer |

Keep whichever of these you run weekly as a prompt you paste, with a `{{transcript}}` placeholder marking where the transcript goes, so the whole workflow becomes paste, run, skim.

#### Getting a usable transcript

Whizi does not join or record your calls. You bring the transcript, and every major platform will produce one: Zoom, Teams, and Meet all have built-in transcription, and dedicated notetakers export plain text.

A few things that materially improve the output:

- **Keep speaker labels.** Attribution is most of the value. A transcript with speakers stripped can still be summarized, but every action item becomes UNASSIGNED.
- **Do not clean it up first.** Filler words and false starts cost you nothing and the model ignores them. Time spent tidying a transcript is wasted.
- **Upload long transcripts rather than pasting them.** A two hour meeting is a lot of text. Uploading it as a file keeps the conversation readable, and Gemini handles the longest ones comfortably.
- **Include the agenda if you have one.** It tells the model what was supposed to happen, which makes the "raised and never resolved" section far more accurate.

#### What to check before you send it

Three specific failure modes, all of which have a cheap check.

**Invented commitments.** A model will occasionally convert "someone should probably look at that" into an owned action item with a name attached. The quote column catches this instantly, which is why it is in the prompt.

**Flattened disagreement.** Summaries drift toward consensus, because most text in the world resolves. If two people disagreed and the notes read as though everyone agreed, the notes are wrong in the way that causes trouble two weeks later. Ask explicitly: `Where did people disagree, and was it resolved?`

**Wrong attribution on a crosstalk-heavy call.** Transcription tools mislabel speakers when people talk over each other. Spot check the attribution on anything consequential before it goes into a document with someone's name on it.

A useful last pass: `What is in this transcript that a busy person would need to know but would not find in the summary above?` It regularly surfaces one thing worth adding.

Key points:

- A two-prompt workflow that produces a recap people read and a task list they can act on
- The extraction rules that stop a model from inventing commitments nobody made
- Templates for standups, customer calls, interviews, and decision meetings

Checklist:

- Keep speaker labels in the transcript, since attribution is most of the value
- Ask for decisions, commitments, open questions, and disagreements rather than a summary
- Require the exact supporting quote on every extracted action item
- Use UNASSIGNED and NO DATE instead of letting the model guess
- Check the disagreement section, since summaries drift toward false consensus
- Keep a reusable prompt for your most frequent meeting type

Steps:

1. Paste or upload the transcript: Drop a Zoom, Teams, or Otter transcript into Whizi.
2. Ask Claude for a summary: Claude produces the warmest, most accurate narrative summary. Ask for 5 to 8 bullets and a one-line headline.
3. Ask GPT for action items: GPT is strongest at structured extraction. Ask for a table of action items with owner and due date.
4. Ask for a follow-up email draft: Claude writes the nicest follow-up emails. Ask it to draft one in your voice.

FAQ:

Q: Does Whizi record meetings?
A: No. Whizi does not join calls or capture audio. You bring the transcript from whatever you already use, whether that is Zoom, Teams, Google Meet, or a dedicated notetaker, and paste or upload it. Keeping speaker labels intact matters more than the source, because attribution is where most of the value in a meeting summary comes from.

Q: Is this safe for sensitive meetings?
A: Whizi does not train on your conversations and each provider’s data policy can be reviewed before you enable that model, but a transcript is one of the most sensitive documents a company produces. For legal, HR, board, and personnel matters, check your internal policy first, and consider summarizing from your own notes rather than the raw transcript. Enterprise plans add further controls for teams with formal requirements.

Q: Which model should I use?
A: Claude for the narrative recap, because it reads intent well and writes notes people finish reading. GPT for the action item extraction, because it holds a strict output format without editorialising. Gemini when the transcript is very long, since it can take a multi-hour meeting in one pass without chunking.

Q: How long can the transcript be?
A: Longer than most meetings produce. A typical hour-long call is well within range for any of the models. For multi-hour sessions, all-day workshops, or a batch of several meetings at once, upload the files and use Gemini, which carries the largest context window of the models in Whizi and can hold the whole set at once.

Q: Can it tell me what was decided if nobody said it clearly?
A: It can flag it, and that is the right behaviour. The prompt on this page asks the model to list implied but unstated decisions separately under UNCONFIRMED, which surfaces them for you to confirm rather than recording a decision that was never actually made. Meetings where nothing was explicitly decided are extremely common, and notes that pretend otherwise cause the problem later.

### How to write blog posts with AI (that do not read like AI)

URL: https://whizi.io/use-cases/write-blog-posts-with-ai/
Updated: August 2026

Quick answer: Writing a blog post with AI goes wrong when you draft first. Research existing coverage in Gemini, find the gap nobody has argued, outline in GPT, then draft section by section in Claude with three paragraphs of your own writing pasted in as a voice sample. Audit every claim and open every source.

A realistic workflow for writing blog posts with AI: GPT for structure, Claude for the actual prose, and a final human pass that takes ten minutes.

#### The mistake is drafting first

Almost everyone starts by asking for a draft. That is why almost every AI blog post reads the same: the model has nothing to work from except the average of everything written on the topic, so it returns exactly that. Fluent, structurally correct, and indistinguishable from the twelve posts already ranking.

Writing a blog post is three jobs, and only the third one is drafting. First you find out what has already been said. Then you decide what you are adding. Only then does anyone write a sentence. Each job also has a different best model, which is what makes a single-model workflow feel like it is fighting you at one stage or another.

| Stage | Model | What it contributes |
| --- | --- | --- |
| Research and existing coverage | Gemini | Recent web material, with sources you can open |
| Angle and thesis | Any, used adversarially | The gap in existing coverage, and the objection to your take |
| Structure and outline | GPT | Holds a brief, produces clean hierarchy, no drift |
| Prose | Claude | Warmest register, least filler, sustains a voice |
| Fact check | Gemini or a web-connected model | Retrieval against live sources, not memory |
| Final edit | You | The part that is not automatable and never was |

Because Whizi keeps the thread when you switch models, the drafting model sees the research and the outline. That continuity is most of why the finished post is coherent rather than a set of well written paragraphs that do not add up.

#### Step 1 and 2: research, then find the gap

**Prompt: what already exists**

`Research current coverage of [topic]. Return: the top pieces currently ranking with their URLs and publication dates, the argument each one makes, the points every single one of them makes (the consensus), any recent data or developments from the last 12 months with sources, and the questions readers ask that these pieces do not answer. Cite everything. Where you cannot verify a claim, mark it UNVERIFIED.`

**Prompt: the gap**

`Based on that coverage, what is missing? Specifically: what does every piece assume without arguing for it, what would an experienced practitioner find naive about the consensus, what question do they all avoid because it is hard, and what would be true for a reader for whom this standard advice fails?`

That second prompt is where a post stops being interchangeable. The output is not your angle, it is a list of candidates, and the one you pick has to be something you actually know to be true. A model can find the gap; it cannot fill it with experience it does not have.

**Prompt: pressure test your angle**

`My thesis is: [state it in one sentence]. Argue against it as forcefully as you can. What is the strongest counterexample, what would I need to demonstrate for this to hold, and is there a narrower version that is more defensible? Do not agree with me.`

#### Step 3 and 4: outline, then draft

**Prompt: the outline**

`Build an outline for a post arguing [thesis]. Audience: [who they are and what they already know]. Return H2s and, under each, the specific point it makes and the evidence it uses. Rules: every section must advance the argument rather than covering a subtopic, no "what is X" section unless the audience genuinely does not know, no conclusion that only restates. Flag any section where I have not given you evidence. Research: [reference the thread above].`

The evidence gap flag is what stops you writing 800 words before discovering you cannot support the central claim.

**Prompt: draft a section, not the post**

`Write the [section name] section. Point it makes: [from the outline]. Evidence: [paste]. Voice: [paste 3 paragraphs of your own writing]. Rules: open with the point rather than context, no sentence that would appear unchanged in a post by a competitor, every claim traceable to the evidence given, no adjective that cannot be measured, no transitional filler. Length: roughly [n] words.`

Draft section by section rather than asking for the whole post. Full-post generation drifts toward summary, whereas a section with a defined job and its own evidence stays specific. It also means you catch a bad section at 200 words instead of at 1500.

**Prompt: criticism before rewriting**

`Before revising, name the three weakest things in this draft: a claim with no support, a paragraph that says nothing new, a sentence that could open any post on this topic, or a place where the argument skips a step. Quote each. Do not rewrite yet.`

#### The tells, and how to remove them

Readers recognise AI writing quickly, and it is a short list of specific habits rather than a vague quality. Each one has a fix.

| The tell | What it looks like | The fix |
| --- | --- | --- |
| The throat clear | "In today's fast-paced digital landscape" | Instruct: open with the point, no scene setting |
| Symmetrical everything | Every section the same length, every list exactly three items | Ask for uneven emphasis: some points deserve a sentence, some a page |
| Unearned balance | "While X has benefits, it also has drawbacks" applied to everything | Require a position, and require the strongest objection to be answered rather than noted |
| Benefits without mechanism | "Improves efficiency" | Require every claim to name how it works |
| The restating conclusion | A final section that summarises what you just read | Ask for a conclusion that tells the reader what to do or what changes |
| Nobody's voice | Correct, smooth, could be anyone | Paste real samples of your writing rather than describing your tone |

The single highest-impact fix is the last one. Three paragraphs of your own real writing pasted into the prompt does more than any list of tone adjectives, because the model can imitate an example and cannot imitate a description.

Then add what a model cannot: the specific number from your own work, the thing you tried that failed, the objection your customer actually raised. One concrete detail from real experience does more for credibility than an entire pass of stylistic editing.

#### Step 5: the fact check you do not get to skip

A fabricated statistic on your blog is permanent, quotable, and yours. Models produce plausible numbers attached to plausible sources with total confidence, and the failure is invisible in the output.

**Prompt: audit the claims**

`List every factual claim, statistic, date, and attributed quote in this draft as a table: the claim, whether it came from a source in this conversation or from your general knowledge, and the source URL if there is one. Mark everything from general knowledge as UNVERIFIED. Do not fill gaps with plausible values.`

Then open every source yourself. Not the ones you doubt, all of them. The most common error is not an invented statistic, it is a real statistic attached to the wrong claim, or a 2019 figure presented as current, and neither is visible without opening the page.

Two things to check specifically: that the number still refers to what you say it refers to, and that the source is the original rather than someone else citing it. Statistics degrade as they get passed along, and a figure three citations deep is often unrecognisable at the origin.

#### What this realistically saves

A 1500 word post that took three hours takes about 45 minutes with this workflow. The saving is concentrated in research, outlining, and getting to a first draft, which were always the slow parts. If you write posts for clients rather than for yourself, [AI tools for freelancers](https://whizi.io/use-cases/freelancers) covers the proposals, scoping, and client voice work that surrounds the writing.

What it does not compress is the final edit, and you should not try. Expect to keep perhaps 80 percent of the generated prose and rewrite the rest, particularly the opening, the transitions between arguments, and anything drawing on your own experience. That last pass is where the post stops being competent and starts being worth reading.

On search: Google's guidance targets low quality content, not AI assistance as such. A post that is accurate, genuinely useful, and adds something the existing coverage does not will do fine. A post generated in one prompt and published unedited will not, and increasingly it will not even be indexed properly, because there are already several thousand of it.

Key points:

- A six-step workflow that puts research and angle before drafting, which is where quality is decided
- The specific tells that make a post read as AI written, and the prompt fixes for each
- A fact-checking pass, because a fabricated statistic on your blog is a permanent liability

Checklist:

- Research existing coverage before writing anything, and note what every post already says
- Find the gap, then pressure test your angle by asking the model to argue against it
- Outline with evidence attached to each section, and flag sections lacking it
- Draft section by section rather than asking for a whole post
- Paste three paragraphs of your real writing as a voice sample in every drafting prompt
- Run the criticism prompt before any rewrite
- Audit every claim and open every source, including the ones you believe
- Add one concrete detail from your own experience per section

Steps:

1. Research the topic: Ask Gemini for recent angles, stats, and competitor coverage with citations.
2. Find the angle nobody else has: Ask what every existing post on this topic says, then write the thing they all leave out.
3. Outline the structure: Ask GPT to build a structured outline with H2s and key points.
4. Draft the prose: Ask Claude to write each section, referencing the outline. Claude is best at warm, human tone.
5. Verify and edit: Re-run claims through Gemini to check facts; tighten the draft with your own voice.
6. Publish with confidence: You will keep maybe 80% of the AI output and rewrite the rest. That is the realistic workflow.

FAQ:

Q: Will Google penalize AI-assisted posts?
A: Google’s stated position targets low quality content rather than AI assistance specifically, and its guidance emphasises usefulness, originality, and demonstrable expertise regardless of how the text was produced. In practice the posts that fail are the ones generated in a single prompt and published unedited, because they add nothing to what already ranks. A post with a real angle, verified facts, and something only you could have written does fine.

Q: Is Claude really better than GPT for writing?
A: For prose that a person reads end to end, yes, consistently. Claude uses less filler, sustains a voice across a long piece, and needs less editing to sound human. GPT is better for structured content such as outlines, comparison tables, and posts built around code or step lists. The strongest workflow uses GPT for the outline and Claude for the drafting, which is one model switch inside the same thread.

Q: How much of the output should I expect to rewrite?
A: Roughly 20 percent, concentrated in predictable places: the opening, the transitions between arguments, and anywhere the post should be drawing on your own experience. If you find yourself rewriting far more than that, the problem is almost always upstream, meaning the angle was generic or the drafting prompt contained no real evidence or voice sample.

Q: Can I just ask for the whole post in one prompt?
A: You can, and the result will be the average of everything already published on the topic, which is exactly what you are competing against. Full-post generation drifts toward summary because no section has a defined job. Drafting section by section, with the point and the evidence specified for each, produces something specific and lets you catch a weak section at 200 words rather than at 1500.

Q: Do I need to disclose that AI was used?
A: Search engines do not require it. Some publications, clients, and employers do, and an increasing number of contracts now address it explicitly, so check the ones that apply to you. The standard worth holding regardless is that you are accountable for every claim published under your name, which means opening every source and being able to explain any number in the piece.

## Prompt templates

### Claude coding prompt pack for debugging, refactors, and architecture reviews

URL: https://whizi.io/templates/claude-coding-prompts/
Updated: August 2026

Quick answer: This Claude coding prompt pack collects twelve copy-ready prompts for debugging, code review, refactoring, architecture, and tests. Four constraints make all of them work: state your framework and language version, constrain the diff, ask for ranked hypotheses before solutions, and require the failure mode. The prompts work with GPT and Gemini too.

A practical prompt library for developers who want to move faster with Claude-style long-context coding workflows inside Whizi.

#### The four constraints that make every prompt below work

Before the prompts, the rules they all share. Adding these to any coding prompt improves the output more than changing model.

**State your version.** Training data skews toward whichever major version has the most written about it, which is often not the one you are on. `We are on [framework] [version], [language] [version]` prevents most outdated answers.

**Constrain the diff.** Ask for a fix and you frequently get a refactor. `Change as little as possible, preserve existing structure and naming, and list every line you changed with a one line reason` is the single most useful sentence in this document.

**Ask for hypotheses before solutions.** A model asked what is wrong gives you a guess delivered as a conclusion. A model asked for ranked causes and cheap checks gives you a debugging plan.

**Require the failure mode.** `What could this break, and is this fixing the cause or the symptom?` catches the most expensive class of AI assistance, which is a change that makes the symptom disappear while the defect stays.

#### Debugging

**1. Ranked hypotheses**

`Here is the error, the relevant code, and what I have already ruled out. Do not give me a fix yet. List the four most likely causes ranked by probability, and for each the single cheapest check that would confirm or eliminate it. Error: [paste]. Code: [paste]. Already ruled out: [list]. Stack: [language, framework, versions].`

**2. The intermittent bug**

`This fails intermittently, roughly [frequency], under [conditions]. Enumerate the categories of intermittent failure that could produce this specific symptom: timing, ordering, resource exhaustion, an external dependency, state leaking between runs, clock or timezone, caching. For each, say what in the code supports or contradicts it, and exactly what I should log to distinguish them. Code: [paste].`

**3. It works locally**

`This works locally and fails in [environment]. List every category of environment difference that could cause this specific symptom: configuration, environment variables, versions, filesystem and case sensitivity, timezone and locale, network and DNS, permissions, resource limits, and build or bundling differences. Rank by likelihood given the symptom, and give me the diagnostic command for each.`

**4. Explain the fix before I take it**

`Explain why this fix works, what it does not fix, and what it could break. If the real cause is elsewhere and this is a symptom patch, say so directly.`

#### Code review

**5. Review a diff**

`Review this diff as a demanding reviewer. Categories in priority order: correctness bugs, security issues, unhandled failure modes, race conditions, then style. For each finding give severity, the specific line, and why it matters in this codebase rather than in general. Do not comment on formatting. If the diff is sound, say so rather than manufacturing findings. Conventions: [describe]. Diff: [paste].`

**6. The security pass**

`Review this code for security issues specifically: injection, authentication and authorisation gaps, unsafe deserialisation, secrets in code or logs, unvalidated input reaching a sensitive operation, and dependency risk. For each, give the attack path concretely rather than naming the category. State clearly what you cannot assess without seeing [deployment, auth layer, data sensitivity].`

**7. The failure mode audit**

`For each external call in this code, state what happens when it is slow, when it fails, when it returns unexpected data, and when it succeeds but partially. Which of those are currently unhandled, and which would be silent?`

That last one finds more real production issues than a general review does, because it asks about the paths nobody wrote a test for.

#### Refactoring and architecture

**8. The refactor plan**

`Propose a sequenced plan to refactor [description]. Constraints: the public API of [x] cannot change, we deploy continuously so every step must be independently shippable, and tests must pass after each step. For each step give the change, the risk, how to verify it, and how to roll it back. Order by risk, lowest first. Do not write the code yet.`

**9. Argue the other side**

`I am choosing [approach A] over [approach B] for [context and constraints]. Make the strongest case for B. What would have to be true about our constraints for B to be correct, and is any of it true here? Do not conclude that both are valid.`

**10. Understand what you inherited**

`Here are the main source files. Produce: the entry points, the data flow from request to response, the state that is shared and where it is mutated, external dependencies and what happens when each is unavailable, and the three areas most likely to contain bugs based on complexity and coupling. State explicitly what you cannot determine from what I provided.`

That final instruction matters. Models will describe the behaviour of a file you did not paste, inferred from its name. Forcing an explicit list of unknowns tells you what to go read.

#### Tests

**11. The tests you would not have written**

`Write test cases for this function, focusing on inputs I probably have not considered: boundaries, empty and null, unicode, very large values, concurrent calls, and any implicit assumption in the implementation. For each test, state the assumption it is checking. Function: [paste].`

**12. Test the test suite**

`Here is a function and its existing tests. What behaviour is not covered? Specifically: error paths, boundary values, interactions between parameters, and anything the implementation does that no test asserts. Do not rewrite the existing tests.`

The second is the higher-value prompt and it is rarely run. Coverage percentages tell you which lines executed, not which behaviours are actually pinned down, and the gap between those two is where regressions live.

#### The second-opinion pattern

The highest-leverage habit in this whole pack, and the one that requires more than one model.

Get an answer from one model. Then switch and hand it over:

`Another engineer proposed this solution to this problem. Find what is wrong with it: correctness under edge cases, concurrency, error handling, performance at [scale], or a simpler approach that was missed. If it is genuinely sound, say so plainly rather than inventing objections. Problem: [paste]. Proposed solution: [paste].`

Two outcomes and both are useful. Either the second model finds a real hole, which you now know before merging, or it agrees despite being pushed to disagree, which is meaningful confirmation. Iterating with the same model gives you neither, because a model reviewing its own output mostly agrees with itself.

Use it on the decisions that would be expensive to get wrong: a schema change, a concurrency fix, anything touching auth or money. Not on routine work. See [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side), [switching models mid-conversation](https://whizi.io/docs/switch-models-mid-conversation), and [write and debug code with multiple models](https://whizi.io/docs/write-and-debug-code-with-multiple-models) for the whole workflow in one place.

#### What to watch for

**Invented APIs.** Confident method names, parameters, and config keys that do not exist, especially for libraries that changed recently. The signature will look right. Check the real documentation before building on anything unfamiliar.

**No confidence signal.** A correct fix and a subtly wrong one arrive with identical certainty. Tone tells you nothing.

**Silent scope creep.** This is what constraint two exists for.

**Security theatre.** Naming vulnerability classes in your code is a useful first pass. It is not an audit, and the model does not know your threat model, deployment, or data sensitivity.

Keep the prompts you use weekly somewhere you can paste from, and put standing constraints in a [project's instructions](https://whizi.io/docs/use-projects) so they apply to every chat in that project automatically.

Key points:

- Twelve copy-ready prompts for debugging, review, refactoring, architecture, and tests
- The constraints that keep output reviewable, including the one that stops unrequested refactors
- A second-opinion pattern that catches the plausible wrong answer before you merge

Checklist:

- State your language, framework, and version in every coding prompt
- Add the constrain-the-diff sentence to any prompt that produces code
- Ask for ranked hypotheses and cheap checks before asking for a fix
- Always ask what a fix could break and whether it treats the symptom
- Run the second-opinion pattern on anything expensive to get wrong
- Ask what the existing test suite does not cover, not just for more tests
- Verify unfamiliar APIs against the real documentation
- Keep the prompts you use weekly where you can paste them

FAQ:

Q: Do these prompts only work with Claude?
A: No. They are written for the long-context, careful-reasoning style Claude does well, and they work directly with GPT and Gemini too. In fact several of them are better used across models: the second-opinion prompt requires two, and the "argue the other side" prompt is more useful when the model arguing did not make the original choice.

Q: Which model should I use for which prompt?
A: As a starting point: Claude for subtle reasoning, unfamiliar architecture, and explaining why something behaves as it does; GPT for fast implementation on well-trodden ground and strict structured output; a large-context model when the question spans more code than fits comfortably in a normal prompt. Then override that with a week of your own comparisons, since the right answer depends on your stack more than on any benchmark.

Q: Is this a replacement for an agentic coding tool?
A: No, they solve different problems. An agent lives in your repository and edits files. These prompts are for the reasoning layer: understanding an error, reviewing a diff, planning a refactor, arguing about an approach. Most developers use both, and model choice matters more here because you are evaluating the reasoning rather than the resulting diff.

Q: How do I stop it rewriting code I did not ask about?
A: Add this to the prompt: change as little as possible, preserve the existing structure and naming, and list every line you changed with a one line reason. Unrequested refactoring is the main reason AI suggestions become unreviewable, and constraining the diff is the difference between a change you can reason about and one you have to re-read from scratch.

Q: Can I paste proprietary code?
A: Whizi does not train on your conversations and each provider’s data policy is reviewable before you enable that model, but your employer’s policy is the binding constraint and it varies widely. Where restrictions apply, reproducing the problem as a minimal example that preserves the structure and drops the business logic is usually both permitted and a better prompt, since it removes the detail that was competing for attention.

### Founder research stack for positioning, market maps, and pricing analysis

URL: https://whizi.io/templates/founder-research-stack/
Updated: August 2026

Quick answer: This founder research stack is nine copy-ready prompts covering market maps, competitor teardowns, pricing structure, positioning, and customer language mining. Run them in one thread so the landing page draft sees the actual research. Two rules decide the quality: open every source you cite, and keep customer language verbatim.

A reusable prompt and workflow set for founders using Whizi to speed up research and messaging work without tool sprawl.

#### How to use this pack

Every prompt below is meant to be copied, filled in, and kept somewhere you can paste from, so you run it again next quarter rather than rewriting it. Bracketed sections are yours to replace. If your research is academic rather than commercial, [the researcher workspace](https://whizi.io/use-cases/researchers) covers the reading and citation side of the same setup.

Two rules make the difference between useful output and a nicely formatted guess. First, every factual claim needs a source you have opened, because a fabricated market size in front of an investor is a memorable way to end a meeting. Second, paste real customer language wherever you have it. The gap between research that reads like a consultant's template and research that is about your company is almost entirely the presence of real quotes.

| Job | Model | Why |
| --- | --- | --- |
| Market map and category scan | Gemini | Recent web material with links you can check |
| Long reports, filings, transcripts | Gemini | Largest context window, so the whole thing fits |
| Competitor teardown | Gemini, then Claude | Retrieve first, then synthesise |
| Pricing structure | GPT | Reliable structured output and arithmetic you can verify |
| Positioning and messaging | Claude | Writes like a person, holds a consistent argument |
| Customer language synthesis | GPT for extraction, Claude for the write-up | Two different jobs |

#### Market map

**Prompt 1: the category scan**

`Research the [category] market. Return: (1) size estimates from the last 24 months, each with the publishing organisation, year, and URL; (2) the 8 to 12 most relevant players with any public revenue, funding, or share figures; (3) material changes in the last 12 months, meaning entries, exits, funding, acquisitions, and regulatory changes, with dates; (4) the two or three structural forces most likely to reshape the category in 24 months. Where sources disagree, show both figures rather than averaging. Mark anything unsourced as UNVERIFIED.`

**Prompt 2: the segmentation cut**

`Using the scan above, propose three different ways to segment this market: by buyer, by job to be done, and by company characteristic. For each segmentation, say which segment is currently underserved and what evidence in the scan supports that. Reject any segmentation where the segments would buy the same product for the same reason.`

The second prompt is the one that produces something usable. A market map that only lists companies tells you who exists; a segmentation tells you where to aim.

#### Competitor teardown

**Prompt 3: the teardown**

`Build a teardown of [competitor]. Cover: stated positioning in their own words, target segments, pricing where public, distribution channels, the flow for [the job to be done] as documented in their help centre, product or messaging changes in the last 12 months with dates, and complaint themes visible in public reviews. Cite every claim with a URL. Separate what the company states from what third parties observe.`

**Prompt 4: the positioning map**

`Place each competitor from the scan on two axes that genuinely differentiate this category. Propose the axes and justify why they matter to a buyer, then reject any axis pair where one axis is a restatement of the other. Show where we sit, which space is currently uncontested, and what would have to be true for that space to be worth owning.`

**Prompt 5: the vulnerability read**

`For each major competitor, identify the customer their positioning implicitly excludes and the complaint theme they have not addressed in 12 months. Which of those gaps could we credibly serve given [our capabilities]? Be specific about what we would need that we do not have.`

#### Pricing

**Prompt 6: structure before number**

`Propose three pricing structures for [product]. For each: the value metric, the tier boundaries, who is priced out, who gets a bargain, the failure mode when a customer grows, and what the structure signals about who we are for. Do not recommend one until I have reacted to all three. Context: what it does [x], who buys [y], what it saves them [z], what competitors charge [paste].`

**Prompt 7: pressure test the price**

`We are considering [price] for [tier]. Argue that this is too high, then argue that it is too low. For each, state what evidence would settle it and what we could measure in the next 30 days to find out. Do not hedge by concluding it is about right.`

The value metric is the part founders most often get wrong and the hardest to change later. Ask specifically: `What are we charging for, and does the customer's use of that thing grow at the same rate as the value they get?` If those two diverge, the pricing breaks as customers succeed.

#### Customer language

The most underused research input a founder has is the pile of things customers have already said. Sales notes, support tickets, churn reasons, and review text are all sitting there unread as a set.

**Prompt 8: mine the language**

`Here are [sales call notes / support tickets / review text]. Return: (1) the exact words customers use for the problem, copied verbatim, never paraphrased, with counts; (2) the objections in order of frequency; (3) what they compared us to; (4) the outcome they say they want, in their words. Flag any theme appearing in fewer than three sources as a single observation rather than a pattern.`

Verbatim is not a stylistic preference. The words customers use are the words that should appear on your landing page, and any paraphrase converts their language into yours, which is exactly the thing you were trying to remove.

**Prompt 9: from research to copy**

`Using only the verbatim language above and the positioning we chose, draft a landing page: hero headline stating a specific outcome, subhead, the problem section in the customer's own words, how it works in three steps, proof, the strongest objection handled, and a final call to action. No claim beyond the evidence provided. Flag anything you had to invent.`

#### The one-afternoon sequence

Run in this order, in one Whizi thread so each step sees the previous ones.

1. Category scan and segmentation, in Gemini. About 40 minutes including checking the sources.
2. Teardowns of your three closest competitors, in Gemini. About 30 minutes.
3. Positioning map and vulnerability read. About 20 minutes.
4. Customer language mining from whatever notes you already have, in GPT. About 20 minutes.
5. Pricing structures, in GPT. About 20 minutes.
6. Landing page draft, in Claude, which now has all of the above in context. About 20 minutes.
7. Open every source. This is not optional and it is the step people skip.

Keeping it in one thread is what makes step 6 work. The drafting model can see the actual competitor language and the actual customer quotes rather than your summary of them, which is the difference between copy that reflects the research and copy that reflects a memory of it.

Key points:

- Nine copy-ready prompts covering market maps, competitors, pricing, positioning, and customer language
- A suggested model for each, since these five jobs want different things
- A one-afternoon sequence that turns the outputs into landing page copy

Checklist:

- Keep each prompt so next quarter is a re-run rather than a rewrite
- Paste real customer quotes wherever you have them
- Demand a URL for every claim and open every one
- Reject segmentations where the segments would buy for the same reason
- Decide the value metric before the number
- Keep customer language verbatim, never paraphrased
- Run the whole sequence in one thread so the copy step sees the research

FAQ:

Q: Is this only for startup founders?
A: No. Consultants doing client positioning work, operators launching a new line, and solo professionals defining an offer all use the same five jobs: map the category, understand the competitors, decide the price, define the position, and use the customer’s own words. The prompts work unchanged; only the bracketed context differs.

Q: How much can I trust the market size numbers?
A: Treat every figure as unverified until you have opened the source. Models produce confident market size estimates from very little, and investors are specifically good at asking where a number came from. The prompts here demand a publishing organisation, a year, and a URL for exactly this reason, and they instruct the model to show disagreeing figures rather than averaging them, since averaging incomparable estimates is the most common way to end up confidently wrong.

Q: Why use three different models for this?
A: Because the five jobs want different things. Retrieval and long documents need the largest context window and live web access. Pricing structures and extraction need strict, reliable formatting. Positioning and copy need prose that sounds like a person rather than a category description. Running them in one thread means each stage can see the previous ones rather than starting from your summary.

Q: What should I do with the output?
A: The landing page draft is the immediate artifact, but the more durable output is the saved templates and the verbatim customer language. Re-run the competitor teardown and category scan quarterly, since both go stale, and keep adding real customer quotes to the language file as they arrive. That file becomes the input that makes every future piece of copy specific to you.

## Release notes

### Whizi release notes: August 2026

URL: https://whizi.io/releases/august-2026/
Updated: August 2026

Quick answer: Whizi shipped 11 user-visible changes in August 2026. This page is the dated log: one section per day, newest first, with what each change means for the way you actually use it. Nothing appears here until it is live for every account, so the list is a record of what shipped rather than a roadmap.

Everything that shipped to Whizi in August 2026, dated, with what each change means if you use it.

#### 20 August 2026

- **Import your ChatGPT and Claude history.** A sidebar entry takes an export file from either service and turns it into ordinary Whizi chats. The file is read in your browser and only the transcript is sent, so a multi-megabyte export never has to be uploaded to be discarded. ChatGPT stores a conversation as a tree of every branch it has ever had, so the import walks back from the node that was last on screen and reconstructs the conversation as you actually left it, rather than replaying discarded drafts and regenerated answers.

- **A reference page for every model in the catalogue.** The /models family gives each metered model its own page: the list rate per million tokens, what one standard answer costs, the context window, the credits it spends in Whizi, and the plan that unlocks it. The pages regenerate from the price data, so they cannot drift from what you are actually charged.

- **Docs rebuilt around the questions people ask.** The documentation is now a reference surface rather than a feature tour. Every price and quota sits in one table, the full credit scale is published with all fourteen rungs, and supported file types, billing, cancellation and data handling each have a page that answers the question directly.

- **AI glossary and context window comparison.** Two new reference pages: a glossary defining the terms that show up in model documentation, and a table comparing context windows across the catalogue so you can see what actually fits in one conversation.

#### 19 August 2026

- **Every reply is credited to the model that wrote it.** Each assistant message now records its own model instead of borrowing whatever the conversation was set to. Before this, switching models partway through relabelled every earlier reply as if the new model had written it, and stepping back through versions credited the wrong model for answers it never produced. The name, avatar and chip on a reply are now the model that actually generated it.

- **The AI Model Cost Index.** A published dataset of what a standard answer costs across the whole catalogue, with the spread between the cheapest and most expensive model stated outright. It is free to cite and refreshes whenever list prices move.

- **Deep research models retired from the catalogue.** sonar-deep-research billed an entire multi-search run per message rather than per turn, so a single message came back anywhere between $0.73 and $2.05 against a rung priced for roughly what one ordinary message costs. The spread between a cheap run and an expensive one was wider than the gap between credit rungs, so no rung could price it honestly and the model was removed rather than repriced. Conversations already pinned to it keep answering.

- **Pricing page opens on monthly billing.** The plan comparison now defaults to the monthly rate, so the first number you see is the one you pay if you do not commit to a year.

#### 18 August 2026

- **Replies survive a closed tab.** Generations now run independently of the request that started them, so a reply keeps going if you close the tab or lose connection mid-answer. Reopening the chat replays what you missed and re-attaches to the live stream. Stop still works on a resumed reply, and the partial answer that gets saved is the one the server actually produced.

- **Reasoning shows above the answer.** On models that expose a reasoning trace, the trace now renders above the reply in both the live stream and the saved conversation. It used to sit below, which looked correct while the answer was still empty and wrong after a refresh, when the finished answer pushed the reasoning to the bottom.

#### 16 August 2026

- **Pin, star, rename and delete chats.** Each chat in the sidebar has a menu: rename it, pin it to a section at the top, star it in one of five colours, or delete it behind a confirmation. Pins and stars persist between sessions, and titles are held to a single line so rows keep their height. Available in all 25 languages.

Key points:

- Import your ChatGPT and Claude history
- A reference page for every model in the catalogue
- Docs rebuilt around the questions people ask

Checklist:

- 11 user-visible changes in August 2026
- Areas touched: Chat, Reference, Models, Pricing
- Every entry is something already live, not something merged or planned

FAQ:

Q: What changed in Whizi in August 2026?
A: 11 user-visible changes shipped in August 2026, across chat, reference, models, pricing. The most recent was on 20 August 2026: import your chatgpt and claude history.

Q: Does this list include everything?
A: It includes every change a person using Whizi could notice. Internal work, infrastructure and documentation plumbing are left out, and nothing appears here until it is actually live rather than merged or waiting on a deploy.

Q: How do I get new features?
A: Nothing to install on the web. Changes are live for every account as soon as they are listed here, on every plan unless the entry says a tier is required.

### How to tell whether a new AI model is actually better for your work

URL: https://whizi.io/releases/how-to-evaluate-ai-models/
Updated: August 2026

Quick answer: To evaluate an AI model, test it on your own work rather than reading launch benchmarks. Collect five real tasks from the last two weeks, write down what a good answer contains before you run anything, run both models in parallel, and score on editing effort rather than how the output reads. Keep the log.

A practical guide to evaluating two frontier model releases against your own tasks, instead of trusting launch benchmarks and vibes.

#### Why the launch benchmarks will not answer your question

Every frontier release arrives with a chart showing it ahead on a set of standardised evaluations. Those numbers are real and they are also close to useless for deciding what you should use on Monday, for three reasons.

**The margins are small and the tasks are not yours.** A two point difference on a reasoning benchmark tells you nothing about whether one model writes a better client email or holds a schema more reliably across two hundred rows.

**Benchmarks measure the tasks that are easy to score.** Things with a right answer. Most professional work does not have one: tone, structure, judgment about what to leave out. Those are exactly where models differ most and where nothing is measured.

**Launch comparisons are made by the vendor.** Not dishonestly, necessarily, but nobody publishes the evaluation where their model came second.

The question worth answering is narrower: on the specific tasks you do twenty times a week, which of these two is better? That has no published answer, and it takes about an hour to find out.

#### What actually tends to change between releases

Across recent frontier releases, the improvements that matter in daily use are consistently in the same areas, and they are rarely the ones featured in the announcement.

| What improved | How you notice | Whether the benchmark shows it |
| --- | --- | --- |
| Instruction adherence | It stops ignoring your third constraint | Rarely |
| Negative instructions | "Do not use analogies" is actually obeyed | No |
| Long-context recall | It finds the thing on page 140, not just page 3 | Partially |
| Formatting discipline | The table has the columns you asked for, every time | No |
| Calibrated uncertainty | It says it does not know instead of inventing | No |
| Tone control | Fewer rewrites before something is sendable | No |
| Reasoning depth | It catches the edge case you missed | Yes, this is the one they measure |

Five of those seven are invisible in a benchmark chart and are the reason a new model feels better or worse to work with. Instruction adherence in particular is the difference between a first draft you edit and a first draft you throw away.

#### The one-hour evaluation

Do this rather than reading launch coverage. Run both models side by side on your own material.

1. **Collect five real tasks** from the last two weeks. Real ones, with your actual context pasted in, not toy prompts. Include at least one writing task, one structured output task, and one where you needed reasoning about something unfamiliar.
2. **Write down what a good answer contains** for each, in one sentence, before you run anything. This is the step that prevents you from preferring whichever output is longer and more confident, which is a strong and largely unconscious bias.
3. **Run each task against both models** in parallel so neither answer is anchored by the other.
4. **Score on editing effort**, meaning how much work between the output and something you would send. Not on how it reads.
5. **Note the gap size**, not just the winner. Interchangeable results tell you to stop thinking about model choice for that task, which is genuinely useful.
6. **Keep the log.** In three months, when the next release lands, you re-run the same five tasks and get a real answer in twenty minutes.

That final point is the compounding one. A saved evaluation set is the only thing that makes each subsequent release cheap to assess.

#### Six prompts that separate frontier models

General questions produce similar answers from any capable model. If you want to see a difference, apply pressure to a specific capability.

- **Tone under difficulty.** `Write a note telling a client we missed the deadline. Take responsibility without over-apologising and without excuses. Under 120 words.` Register differences show up immediately.
- **Negative constraint.** `Explain [concept] without using any analogy or metaphor.` Compliance with negative instructions varies far more than you would expect.
- **Strict extraction.** `Extract every date, amount, and party into a JSON array with exactly these keys. If a field is absent use null. Do not infer.` Tests format discipline and the tendency to fill gaps.
- **Long-context recall.** Upload a long document and ask about something in the middle. Reveals usable context, which is not the same as advertised context.
- **Admitting ignorance.** Ask about something genuinely obscure or very recent. The best answer is a clear "I do not know" or a sourced retrieval. Fabrication here is disqualifying regardless of any benchmark.
- **Multi-constraint compliance.** Give six constraints at once and count how many survive. This single test predicts day-to-day satisfaction better than anything else on the list.

#### The answer is usually "both, for different things"

People approach a release wanting a verdict, and the honest finding after running the evaluation is nearly always split. One model wins on writing and tone. The other wins on strict structure and speed. They are close enough on general reasoning that it does not decide anything.

That is not a fudge, it is the actual result, and it has a practical implication. If you can only use one model, you are picking which category of task to be worse at. If you can use both, the release question stops being "should I switch" and becomes "which tasks move", which is a much smaller and lower-stakes decision.

It also changes what a release means to you. When new models land in a workspace that already has several, you re-run your five tasks, adjust your routing, and carry on. There is no migration, no cancelled subscription, and no month of using something worse because you committed before testing.

#### What to do on release day

**Do not switch your defaults immediately.** Launch-week impressions are dominated by novelty and by whichever examples circulated first.

**Re-run your evaluation set.** Twenty minutes if you kept one from last time.

**Check the boring things.** Context window, whether your existing prompts still behave the same way, and whether anything you relied on has changed. A model that is better in general can be worse on your specific template, and that is worth knowing before you move production work onto it.

**Update routing per task, not wholesale.** Move the categories where the new model clearly won and leave the rest.

**Wait a fortnight for the limitations.** The failure modes of a new model surface within about two weeks of wide use, and they are rarely in the announcement.

For the general framework see [how to choose an AI model](https://whizi.io/resources/how-to-choose-ai-model), and for the mechanics of running two models on one prompt see [comparing models side by side](https://whizi.io/docs/compare-models-side-by-side).

Key points:

- Why launch-day benchmarks systematically fail to predict which model is better for you
- A one-hour evaluation you can run on your own real tasks
- The six prompts that reliably separate two frontier models

Checklist:

- Build a set of five real tasks from your own work and keep it
- Write down what a good answer contains before you read either output
- Run both models in parallel so neither anchors the other
- Score on editing effort, not on how the output reads
- Record the gap size, since interchangeable results are useful information
- Test negative constraints and multi-constraint compliance specifically
- Wait a fortnight before moving production work to a new model
- Update routing per task rather than switching wholesale

FAQ:

Q: Should I switch to the newest model?
A: Not on launch day, and not wholesale. Re-run a small set of your own real tasks against both, score on how much editing each output needs, and move only the categories where the new model clearly wins. Newer is reliably better at some things and occasionally worse at others, particularly for existing prompts tuned against the previous version.

Q: Why do benchmarks not match my experience?
A: Because they measure what can be scored automatically, which means tasks with a right answer. Most professional work has no single right answer, and the qualities that decide day-to-day satisfaction, meaning instruction adherence, tone control, formatting discipline, and knowing when to say "I do not know", are largely unmeasured. A model can lead every published chart and still be more annoying to work with.

Q: How often should I re-evaluate?
A: Whenever a model you use gets a significant release, which currently means every few months, plus once a quarter regardless. Keeping a fixed set of five real tasks makes this a twenty minute job rather than an afternoon, and it is the only way to notice that the routing you set six months ago is now wrong.

Q: Can I just use whichever model wins overall?
A: You can, and you will be accepting worse results on a predictable slice of your work. The consistent finding when people evaluate on their own tasks is that one model wins on writing and tone while another wins on strict structure and speed, with general reasoning close enough not to decide anything. If both are available to you, the choice becomes per task rather than per subscription.

Q: Does it matter which product I access the model through?
A: The model is the model, so output quality is broadly the same wherever you reach it. What differs is whether you can compare, whether context carries when you switch, and whether a new release costs you a migration or is just another option in the picker. A workspace with several models turns each release from a decision into an adjustment.