The company says the model beats its open rivals at coding, and the first independent testing suggests it is fast and fairly cheap. On Terminal-Bench 4.0, extra reinforcement learning training lifted its score from 28 to 41 percent, just ahead of the Chinese open-weight model GLM-5.3 at 40 percent. On Artificial Analysis' Coding Agent Index, Mistral claims 48 percent, effectively tied with GLM-5.3 while using 47 percent fewer tokens. Anthropic's Claude Opus 5.5 still tops that index at 62 points.

The overall picture is less flattering: in Artificial Analysis' Intelligence Index, the final version scores 26 points, only one point more than the preview, and it trails Chinese models in the overall rankings. Mistral's claim to be the leading open model for cyber defence also deserves a caveat. Its own chart shows DeepSeek V4.1 Flash at 57 percent and MiMo-V2.6-Pro at 56 percent on the Cyber Index, putting Large 4, at 51 percent, third among open models.

There is a real dual-use tension. Mistral says the model refuses 96.6 percent of malicious cyber requests on average, more than any other open model, but admits it could make real attacks easier and plans to tighten safeguards in the coming months. A special version for penetration testing is already being tested by select partners such as Orange Cyberdefense, with access requiring a review and contractual restrictions.

Pricing is $0.70 per million input tokens and $2.10 per million output tokens, down from $1.36 and $4.18 during the preview. The release uses a modified MIT licence: use, modification and commercial deployment are free, but companies reselling Large 4 as a commercial cloud service need a separate licence from Mistral.