Write and debug code with multiple models in Whizi

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, 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, Mistral, Qwen and GLM. 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.

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.

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. 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.

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.

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 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.
• Parallel. Comparing 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 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 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, which apply to every chat in that project automatically.