Using Mistral in Whizi: models, credit costs and plan gates

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.

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.

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.