Mistral Large 4 drops to 25 points in Artificial Analysis 5 benchmark
In brief
- Artificial Analysis 5 benchmark now relies on 60 % private tests, up from 40 % previously.
- Mistral Large 4 drops from 38 to 25 points, tying with Meta’s Muse Spark Open.
- The gap between closed and open models widens from 12 to 26 points according to Artificial Analysis.
Artificial Analysis released on October 27, 2026 a new version of its benchmark where 60 % of evaluations now rely on private tests or secret answer formats. According to the publisher, version 5 replaces public tasks with inaccessible ones for labs, such as Cognition’s FrontierCode 2.0 or a subset of CritPt v2 in physics.
Twelve models with published or announced weights lose between 12 and 16 points, including Mistral Large 4, which drops from 38 to 25. The top closed model, Claude Opus 5.5, holds steady at 58 points, widening the gap with open models to 26 points. Artificial Analysis highlights a gap that grew from 3 to 12 points, but this calculation includes Kimi K3.1 Pro whose weights remain unpublished.
The benchmark’s redesign also changes the tests: GDPval-AA, SciCode and AA-LCR are removed, while Terminal-Bench-Science and FrontierCode 2.0 are added. European models with published weights remain rare, and Mistral Large 4, whose weights are expected by October 31, still lacks detailed per-task results.
This summary gives you the essentials. The full article, its figures and quotes are on ActuIA.
Read the original articleAI-generated summary of the source, checked by rules. How we write · Report an error