Mistral Medium 3.1

Mistral Medium 3.1 was the August 2025 refresh of the model with which the French lab made its clearest commercial argument: that a company does not need a frontier model to do most of its work. Mistral pitched the Medium 3 line as delivering at or above 90% of Claude Sonnet 3.7's benchmark performance while charging 0.40 US dollars per million input tokens and 2.00 per million output — roughly an eighth of what the model it was compared against cost at the time. It also claimed to beat leading open models such as Llama 4 Maverick and enterprise systems such as Cohere Command A. The 3.1 revision, catalogued as mistral-medium-2508, kept that pricing and added the practical refinements enterprise buyers had asked for, working with a context window of roughly 131,000 tokens and accepting images alongside text. The more consequential feature was where it could run: Mistral offered hybrid, on-premises and in-VPC deployment, and stated that the model would run self-hosted on four GPUs and above. For European organisations that could not send customer data to an American API, that made it one of very few realistic options in its performance class. Its life was short. Mistral released Medium 3.5 on 29 April 2026, deprecated Medium 3.1 on 22 May 2026 and set 31 August 2026 as the date the model stops answering. The successor is several times more expensive per token but adds open weights and a far larger context window — a trade that reads as Mistral moving its mid-tier upmarket and leaving the very cheap segment to its Small and Ministral families.

▸Videos

No videos yet.

▸Reviews