Cybench
Agents · 2024-08-15
A cybersecurity agent benchmark measuring autonomous vulnerability discovery and exploitation across sandboxed challenges.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Mythos Preview | 100.0 |
| Opus 4.6 | 93.0 |
| Opus 4.5 | 82.0 |
| GPT-5 | 73.5 |
| Sonnet 4.5 | 60.0 |
| gpt-oss-120b | 49.5 |
| Opus 4 | 46.9 |
| Grok 4 | 43.0 |
| Opus 4.1 | 42.0 |
| DeepSeek-V3.1 | 40.0 |
| Grok 4.1 | 39.0 |
| DeepSeek-R1-0528 | 35.5 |
| Sonnet 4 | 35.0 |
| Grok 4 Fast | 30.0 |
| o3-mini | 22.5 |
| Claude 3.7 Sonnet | 20.0 |
| Claude 3.5 Sonnet (June 2024) | 17.5 |
| GPT-4.5 | 17.5 |
| DeepSeek-R1 | 16.7 |
| GPT-4o (2024-11-20) | 12.5 |
| Opus 3 | 10.0 |
| o1-mini | 10.0 |
| o1 Preview | 10.0 |
| Gemini 1.5 Pro | 7.5 |
| Llama 3.1 405B Instruct | 7.5 |
| Mixtral 8x22B Instruct | 7.5 |
| Llama 3 70B Instruct | 5.0 |