Models and cost estimates
Why the model list is always current and what the costs mean.
The model catalog refreshes from Mistral’s live model list, so newly released models appear and retired ones disappear without any package update. Startup works offline: the last fetched snapshot restores automatically, and a static fallback applies on the very first start.
Rules for which models appear: only chat-capable models; internal and archived models are dropped; deprecated models are dropped — unless that would drop every eligible model. Vision-capable models accept image input; reasoning-capable models expose thinking levels.
Where prices come from
Section titled “Where prices come from”The live model list carries no prices, so costs come from a sourced family rate card: the first entry whose match string is a substring of the model id applies, with a catch-all so every id gets a price. Each built-in rate was verified against Mistral’s pricing pages at implementation time. Your own models.rateCard entries are checked before the built-in card.
What the costs mean
Section titled “What the costs mean”Every cost shown in pi is a catalog estimate, not a billed provider total. Mistral’s wire formats report token counts only — billed totals live behind an admin-only endpoint — so the provider marks every usage cost block as an estimate. Model ids without a rate-card entry get fall-through pricing (the catch-all rates) and are flagged approximate; models.fallThrough: false swaps that for a zero-cost placeholder, still flagged approximate.
Output limits
Section titled “Output limits”A model’s maximum output tokens derive from Mistral’s wire-enforced completion ceilings and documented output limits — the API rejects a larger max_tokens with 400 code 3058. The mapping tables live in Max-output mapping.