Mistral API

What Mistral actually meters

Mistral meters tokens for text, pages for OCR, minutes for audio and characters for speech, and the token card is unusually regular. Cached input lands at exactly 10% of standard input on every row, from Large 4 down to Ministral 3 3B, and writing to the cache costs nothing.

Output multiples are where models diverge. The three Ministral models charge the same for output as input, a 1x spread that's almost unheard of, while Large 3 runs 3x, Small 4 runs 4x and Medium 3.5 runs 5x.

Three service modifiers sit on top, each all-or-nothing by model family. Batch takes 50% off, Priority Tier multiplies by 1.75, regional inference adds 10%. Toggle any of them and only the seven flagship rows move: Large 3 goes to $0.25 batched, $0.875 on Priority, $0.55 regional. OCR, Voxtral, Codestral, moderation and the Z.ai GLM 5.3 model Mistral resells hold standard rates throughout. Mistral doesn't say whether Priority and regional stack.

One lifecycle detail carries a price. Mistral says plainly that calling a model by its -latest alias opts you into silent updates in behaviour and pricing.

Dollars or euros, and why the answer changes

Mistral serves both currencies and converts them by two different rules. On the API rate card, every euro figure is the dollar figure times 0.85: $0.50 becomes €0.425, $7.50 becomes €6.375, $4 per 1,000 OCR pages becomes €3.40. That's a fixed internal rate, not a live one, and the round numbers sit on the dollar side, so dollars are the base.

The subscription page works the other way. Pro costs $14.99 a month excluding tax, and euro mode shows 17.99€ including tax, which is 14.99 plus 20% French VAT. Team does the same, $24.99 per user becoming 29.99€. The number never converts, so a European team pays 14.99 of a more valuable currency for the same plan. The included credit stays in dollars either way: Pro grants $25.50 a month in API credits, 1.7 times what the plan costs.

What happens when you hit the limit

Usage stops, and which limit stops it depends on settings you have to log in to see. Included monthly usage is shared across Studio, the API and Vibe Code, and it drains first. With pay-as-you-go on, everything past it bills per token. Without it, usage can halt until the next billing period or until an admin changes the setting.

Spending limits are the harder stop. Organisation and workspace monthly caps both exist, workspace caps can't exceed the organisation's, and Mistral warns that reaching one can suspend API access until the month rolls over. Failed invoice payments do the same.

The real gap is upstream. Mistral documents that completion limits exist as tokens per minute and requests per second, audio limits in seconds per minute, OCR limits in pages per minute, then publishes a number for none of them, nor for what Free mode includes.