Mistral Small 3.2 24B pricing.

A page is about 600 tokens in and 60 out, a support reply 200 in and 300 out, an email 150 in and 250 out. Nothing else is counted: no cache discount, no batch rate, no free tier.

Each is a measurement of the stated version and configuration. Two benchmarks are not the same test, and a score is not the chance your own work succeeds.

Prices are per million tokens. The big number on each row is blended 3 in : 1 out, the way most work uses a model. A price below Mistral AI's own is listed as cheapest only when the provider says why.

Resells

The provider sells another company's endpoint, at a margin of its own or at the maker's price.

Runs it themselves

The provider runs the open weights on hardware of its own, so the price is its own rather than a resale.

Lower precision

The open weights, served at a lower precision than the maker's, so not quite the same product.

Credits

Priced in credits or points that convert to dollars, so the rate moves when the conversion does.

Same as maker

Sold by someone else at the price the maker charges.

Maker's price

What the lab that trained the model charges on its own API. Every other row is measured against it.

Costs more

Sold by someone else for more than the maker charges.

Free tier

Given away inside a quota. A limit is not a price, so it is never called cheapest.

Not ranked

Below the maker's price on a closed model with nothing to account for it, or from a catalog we could not read reliably. Listed, and never a pick.