Whizi troubleshooting: fixing common problems

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The answer is out of date or invented	Web search is switched off in that conversation	Turn the search toggle on in the composer and ask again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. If no symptom here matches yours, how to contact Whizi support sets out which channel to use and what to include. One case that is not a fault at all is a dead connection, and what still works without one is listed in using Whizi offline.

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-6 Astra 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, and the exact refusal strings behind a locked model are on model unavailable.

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, and if the upload itself was refused with an error, file upload failed lists every refusal string.

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, and if the toggle is already on but no results ever come back, 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, and the exact refusal you see at zero is on out of 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.

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.