Mistral Large 4 Preview - Cybench: leaderboard
Mistral Large 4 public-preview post (2026-10-06): Cybench, 40 professional capture-the-flag security challenges; share of challenges solved, as reported by Mistral for five models (the post's text says the preview solved 93%). Unlike the neighbouring CyberGym chart it is not attributed to Artificial Analysis, and attempts, agent scaffold and who ran the comparators are not stated; none of these rows are on the Cybench leaderboard. Challenges solved (%), higher is better.
Metric: Challenges Solved (%). Source: mistral.ai. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Mistral Large 4 | 93 | #106 |
| 2 | Kimi K3 | 90 | #35 |
| 3 | DeepSeek V4 Pro (0813) | 88 | #58 |
| 4 | GLM-5.3 | 85 | #44 |
| 5 | GLM-5.2 | 73 | #80 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mistral-large-4-preview-cybench · How It Works · Data refreshed daily, snapshot 2026-11-05.