CyBench: leaderboard
40 professional-level Capture-the-Flag tasks across crypto, web exploits, reverse engineering, forensics, and privilege escalation. Models interact with Kali Linux via bash. Includes guided and unguided modes.
Metric: End-to-End % Solved. Source: cybench.github.io. Saturation forecast: Around January 2027. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Mythos Preview | 100 | #4 |
| 2 | Best Expert | 100 | |
| 3 | Claude Opus 4.7 | 96 | #49 |
| 4 | Claude Opus 4.6 | 93 | #63 |
| 5 | Claude Opus 4.5 | 82 | #83 |
| 6 | Muse Spark | 65.4 | #64 |
| 7 | Claude Sonnet 4.5 | 60 | #142 |
| 8 | Grok 4 | 43 | #181 |
| 9 | Claude Opus 4.1 | 42 | #140 |
| 10 | Grok 4.1 (Thinking) | 39 | #227 (Grok 4.1) |
| 11 | Claude Opus 4 | 38 | #163 |
| 12 | Claude Sonnet 4 | 35 | #199 |
| 13 | Grok 4 Fast | 30 | #254 |
| 14 | O3 Mini | 22.5 | #272 |
| 15 | Claude 3.7 Sonnet | 20 | #249 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cybench · How It Works · Data refreshed daily, snapshot 2026-10-20.