Mistral fights back: 38 points at $1.13/task, best open model in the West, top cyber score
rickasaurus · x · 2026-10-06
A detailed numbers breakdown of Mistral's new model:
- Dollar-per-intelligence: 38 points at $1.13/task vs MiMo v2.6 Pro's 46 at $0.13; ties GPT 6 Luna but at 16x the cost ($0.07/task)
- Best open model out of the US or Europe, ahead of Nemotron 3 Ultra (23), Gemma 4 (17), GPT-OSS (12)
- Ranked #64 of 225 on Artificial Analysis; 1T parameters still lands 3 points under GLM 5.3 Flash (320B, $0.25)
- Open weights (Oct 27) though AA lists it as proprietary; uses a new "custom Mistral license", less open than Apache 2.0 / modified MIT
- SOTA on legal (15%); highest cyber score of any model at 82% (Opus 5.5 and Astra refuse and score near zero)
- Beats GLM 5.3 on DeepSWE 62–61 per Mistral's own chart (live leaderboard has GLM at 69)
Related event: Mistral Large 4 Review: France Rises to Third in Frontier Model Rankings(9 posts)→
More from Models
- Dev Claims Opus 5.5 Was Nerfed: Model Names Are a Trick, Only Margins Are Real — StewartalsopIII · 2026-10-07
- Mistral launches 1T-param Large 4 "Le Chonk": 49B active, open weights by end of October — Kyrannio · 2026-10-07
- Kardashev-0.7: a trained swarm of 32 models claims frontier-level performance at 1% of inference cost — ZeroStateReflex · 2026-10-07
- Cohere Labs' Tiny Aya L2-Thinker reasons natively across 60 languages via data mixing — Cohere_Labs · 2026-10-07
- Ramp: Record 8% of firms switched top AI model provider in September — annbordetsky · 2026-10-07
- Would an LLM push back on a Marxist user, or stay sycophantic? Investor ponders AI psychosis — StewartalsopIII · 2026-10-07