Mistral's mid-size frontier model: near-flagship quality at a fraction of Large pricing.
Mistral Medium 3 occupies the sweet spot between Mistral Large and Small: it beats Claude Sonnet 3.7 on most benchmarks at roughly one-eighth the API cost, and supports multimodal input plus function calling out of the box.
Who it's for: Cost-conscious teams running high-volume production workloads who still need near-flagship quality.
MMLU, HumanEval and MATH scores within a few points of flagship models at ~$0.40/M input tokens.
Understands images, documents and charts alongside text — no separate vision endpoint needed.
Structured outputs and tool use are first-class, so agents and RAG pipelines work out of the box.
Available via La Plateforme, Azure AI, Amazon Bedrock, and self-hosted weights for enterprise.
Especially good on European languages versus US-centric competitors.
Fast token generation suitable for chat UX and real-time copilots.
If your bill is scaling faster than your traffic, Mistral Medium 3 is the first place to look. It delivers 90% of frontier quality for a tenth of the price, with real multimodal and function-calling support. For the last 10% of hard reasoning, keep a flagship on standby — but route the default path here.