Mistral Medium 3 lands at $0.40/$2 per million tokens — and ships without open weights
Mistral released Mistral Medium 3, which it says performs at or above 90% of Claude Sonnet 3.7 on its benchmark set while costing 8x less: $0.40 per million input tokens and $2 per million output tokens. The company claims it surpasses open models including Llama 4 Maverick and enterprise models like Cohere Command A, and says it undercuts DeepSeek v3 on price both in API and self-deployed form, with self-hosting from four GPUs upward, continuous pretraining and full fine-tuning for enterprise customers. It launched on Mistral's La Plateforme and Amazon SageMaker, with IBM watsonx, NVIDIA NIM, Azure AI Foundry and Google Cloud Vertex listed as coming.
The details a license reviewer needs are absent: Mistral published no parameter count, context window or training-cutoff figure, and unlike Mistral Small 3.1 in March, Medium 3 shipped as an API and enterprise deployment product rather than open weights. The blog closes by teasing a large model in the coming weeks instead.
WHY IT MATTERS
This is what the price of capability looks like when a mid-tier model is sold as frontier-adjacent: a single-digit multiple of nothing versus Sonnet 3.7 forces flagship API pricing and per-token budgeting down across the board, and a four-GPU self-host floor makes on-prem or VPC routing — and data-residency constraints — viable for teams that cannot send code to a hosted endpoint. Because the weights were not released, treat it as a contract negotiation, not a download: you need the deployed model version pinned in the agreement, plus custom post-training terms, before you build a dependency on that cost line.
