Medium is the new large.
Mistral AI ● Covered by 2 sources
Mistral AI dropped a new model called Mistral Medium 3 today. It matches most of Claude Sonnet 3.7's benchmarks but costs way less to run.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral AI just launched Mistral Medium 3, and the pitch is refreshingly blunt: near-frontier performance at a fraction of the price. The company says it hits roughly 90% of Claude Sonnet 3.7's scores across its benchmark suite while charging $0.4 per million input tokens and $2 per million output tokens, a gap Mistral is calling an 8x cost advantage over comparable frontier models.
The numbers matter less than the positioning. Medium 3 isn't trying to be the biggest model on the leaderboard. It's trying to be the model finance teams actually approve. Mistral claims it beats Llama 4 Maverick and Cohere Command A on performance, and undercuts DeepSeek v3 on price, whether you're hitting an API or running it yourself. Coding and STEM tasks are where the model reportedly shines brightest, closing much of the gap with far larger, slower systems, and third-party human evals apparently back that up in real-world coding scenarios rather than just academic tests.
What's notable is how hard Mistral is leaning into deployability. Medium 3 runs on self-hosted setups with as few as four GPUs, and the company is offering hybrid, on-premises, and in-VPC deployment alongside custom post-training. That's a direct pitch to regulated industries — Mistral name-checks beta customers in financial services, energy, and healthcare who are already using the model for customer service, business process work, and dense dataset analysis. The message is clear: this is built for companies that can't just call an API and call it done.
Availability starts today through Mistral's own La Plateforme and Amazon SageMaker, with IBM WatsonX, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex support coming soon. And Mistral tucked in a teaser at the end worth noting: after shipping a Small model in March and now a Medium model, a Large release is apparently coming within weeks, with hints it might be open-sourced given how well Medium 3 already stacks up against open flagship models like Llama 4 Maverick.
My take — AI-written commentary, not fact-checked reporting
This is Mistral doing what it does best — undercutting on price while quietly outperforming bigger names, and I think that's exactly the right strategy for a company that can't out-compute OpenAI or Anthropic. The real story here isn't the benchmarks, it's the four-GPU self-hosting and VPC deployment options, because that's the stuff European enterprises and regulators actually care about. If the teased Large model ships open-weight, Mistral will have built the most credible non-American alternative in the room, and that's worth more to me than another point on a leaderboard.
Read more about this at: Mistral AI