Mistral Medium 3: a proprietary model at 0.40/2 dollars per million tokens, self-hostable from four GPUs

Think of AI models like cars: there are the expensive high-end sports cars (powerful, costly) and the cheap economy models (affordable but limited). Mistral Medium 3 is the premium mid-range sedan — near top-tier performance at a fraction of the price.

Mistral AI, the French startup that became Europe's flagship AI company, introduced its new mid-size model on 7 May 2025. The pitch is cost-performance: the company does not claim to beat the largest models on every test, but to get close while charging far less. Published pricing is 40 cents per million input tokens and 2 dollars per million output tokens (tokens are the fragments of text a model breaks language into).

One clarification matters, because it is where Mistral changes strategy: Medium 3 is not an open model. The releases that made the company famous — Mistral 7B and the Mixtral family — could be downloaded and run freely. This one cannot. You use it through an API, or you install it on your own servers under a commercial agreement: the weights are not public and there is no download to grab.

That said, for a European company the deployment option is still the interesting part. Medium 3 can run on Mistral's own cloud, on Amazon's cloud, or inside the corporate perimeter on a relatively small group of accelerators — four or more, per the stated requirement. In every case sensitive data never has to leave the chosen jurisdiction, which makes GDPR compliance a design decision rather than an afterthought.

In short: a mid-size company can adopt a high-tier model without necessarily routing data through an American provider and without paying frontier-model prices, accepting in exchange a dependency on a commercial licence rather than on free weights.