AI cost · Mistral
Mistral API cost: what you actually pay
Model names don't track price tier here — the open-weight 'Large 3' is cheaper per token than the closed 'Medium 3.5' — and the large cache (~90%) and batch (50%) discounts require explicit opt-in, so default calls pay full price.
Public pricing
Token rates (USD per 1M tokens)
| Model | Input | Output | Cache read |
|---|---|---|---|
| Mistral Medium 3.5current top proprietary/closed general model | $1.5 | $7.5 | — |
| Mistral Large 3675B open-weight flagship; cheaper than Medium 3.5 despite the 'Large' name | $0.5 | $1.5 | — |
| Mistral Small 4 | $0.15 | $0.6 | — |
| Codestralcoding-specialized | $0.3 | $0.9 | — |
Batch processing is a flat 50% discount on both input and output; cached input tokens are priced about 90% below standard input — both require explicit API parameters/endpoints, not automatic. Open-weight ('Open' license) models like Large 3 and Small 4 can also be self-hosted, so the per-token price is effectively an inference convenience fee rather than a licensing fee.
Last updated: 2026-07-22. Sourced from Mistral AI API Pricing (official) ↗, Mistral AI Pricing overview (official) ↗.
Where the real Mistral API cost hides
The sticker is the smallest part of the story. For Mistral API, these are the line items that quietly inflate the bill:
- Mistral Medium 3.5 (closed, $1.5/$7.5) costs more per token than the open-weight 'Large 3' flagship ($0.5/$1.5) — picking by model name/size alone can land you on the pricier option.
- Magistral (reasoning) models bill chain-of-thought/thinking tokens as regular output tokens at the same $/M rate, so verbose reasoning traces multiply cost versus a non-reasoning model for the same task.
- The ~90% cached-input and 50% batch discounts don't apply unless explicitly invoked — uncached, non-batch calls (the default in most SDK quick-starts) pay full price.
What an audit finds on Mistral API
Teams default to Mistral Medium 3.5 assuming it's 'the best/priciest for a reason' without realizing the open-weight Large 3 is both cheaper per token and self-hostable later, so they overpay for a name rather than a needed capability.
Mistral API pricing FAQ
Is Mistral Large actually the most expensive model?
No — as of mid-2026 the open-weight 'Large 3' ($0.5/$1.5 per M) is cheaper than the closed 'Medium 3.5' ($1.5/$7.5 per M); check current per-model rates rather than assuming by name.
How much do batching or caching save?
Batch API cuts both input and output by 50%; prompt caching cuts repeated input tokens by about 90% — but only for calls that explicitly use the batch endpoint or caching parameters.
What is Mistral API really costing you?
Spendassay connects your Mistral API spend to actual usage and engineering output — a proof-level AI cost report with the wasted dollars named. Neutral across every AI vendor.
Free · read-only · no card
Other api platforms pricing
More: the complete AI cost management guide · every tool, with prices · cost per outcome · compare platforms · how we measure