Mistral AI launches Mistral Medium 3

What happened

On 7 May 2025, Mistral AI launched Mistral Medium 3, a model placed in the middle of its lineup rather than at the very top of it^[1]. The announcement was not about a model alone: on the same day, Mistral introduced Le Chat Enterprise^[2]. Developers could reach the new model through the Mistral Medium 3 API on two routes at once — Mistral's own La Plateforme and Amazon SageMaker^[1].

Price, and the comparison it invites

Mistral priced Mistral Medium 3 at $0.40 per million input tokens and $2 per million output tokens^[1]. A token is the unit a model reads and writes in — roughly a short word, or a fragment of a longer one — so per-million-token prices are the usual way to compare what a workload costs to run. On Mistral's own account, the model reached at least 90% of the benchmark performance of Claude Sonnet 3.7 while costing about 8 times less^[1]. Both halves of that sentence matter. The claim is a ratio against a named competitor, measured on benchmarks, and it is the vendor's own claim rather than an independent measurement.

What it runs on

Mistral said the model can be self-deployed on four or more GPUs^[1]. Self-deployment means running the model's weights on hardware the customer controls — its own data centre or its own cloud account — instead of sending requests to a vendor's API. The GPU figure is the practical constraint that decides who can do it. A handful of accelerators is a plausible purchase or rental for a single team, whereas a model that needs a far larger cluster is effectively API-only for most users. For an organisation with data it will not send outside its own boundary, that difference is often the deciding factor.

Capability claims

Mistral also said Medium 3 comes close to much larger and slower competitors on coding and STEM tasks^[1]. Coding and STEM (science, technology, engineering, mathematics) problems are common tests because their answers can be checked: code either runs or it does not, a maths result is either right or wrong. That makes them easier to score than open-ended writing. The claim being made is about a trade-off — not that a mid-sized model beats the biggest ones, but that it gets near them while being cheaper and faster to serve.

Why it matters

Two things stand out. First, the price per token sets the floor for anything built on top of the model: at $0.40 per million input tokens, a workload that reads a lot of text costs little enough that applications which were previously marginal become worth running^[1]. Second, shipping the same model on Amazon SageMaker as well as Mistral's own platform, and offering self-deployment on four GPUs, gives it three quite different routes to users — a vendor API, a third-party cloud service, and customer-owned hardware^[1]. That matters to large organisations, which is also why the same-day Le Chat Enterprise announcement fits the pattern^[2].

The wider point is about where capability sits. Mistral's pitch is that a medium-sized model is enough for a large share of real work, and that the right comparison is not benchmark score alone but score per dollar and per unit of response time^[1]. Whether that holds up is for users and independent evaluations to settle; the launch itself states the bet clearly.