Benchmarks
Cybench
Models scored
22
Best score
0.93
Metric
Unguided % Solved
Cybench is an AI evaluation tracked by Epoch AI, with 22 scored model versions on record. The scores here are reported by third parties rather than produced by Epoch AI running the evaluation itself.
Scored models span February 2024 to February 2026. The best score recorded is 0.93, measured as Unguided % Solved. Models come from Anthropic, xAI, OpenAI, Mistral AI, Meta AI, Google DeepMind.
Full record
- Score metric
- Unguided % Solved
- Models scored
- 22
- Best score recorded
- 0.93
- Earliest model scored
- Feb 15, 2024
- Most recent model scored
- Feb 5, 2026
- Third-party reported
- Yes
Leaderboard
| 01 | claude-opus-4-6_unknown | Anthropic | 0.93 |
| 02 | claude-opus-4-5-20251101_unknown | Anthropic | 0.82 |
| 03 | claude-sonnet-4-5-20250929_unknown | Anthropic | 0.6 |
| 04 | claude-sonnet-4-5-20250929 | Anthropic | 0.55 |
| 05 | grok-4-0709 | xAI | 0.43 |
| 06 | claude-opus-4-1-20250805 | Anthropic | 0.42 |
| 07 | grok-4-1 | xAI | 0.39 |
| 08 | claude-opus-4-20250514 | Anthropic | 0.38 |
| 09 | claude-sonnet-4-20250514 | Anthropic | 0.35 |
| 10 | grok-4-fast | xAI | 0.3 |
| 11 | o3-mini-2025-01-31_medium | OpenAI | 0.23 |
| 12 | claude-3-7-sonnet-20250219 | Anthropic | 0.2 |
| 13 | gpt-4.5-preview-2025-02-27 | OpenAI | 0.17 |
| 14 | claude-3-5-sonnet-20240620 | Anthropic | 0.17 |
| 15 | gpt-4o-2024-11-20 | OpenAI | 0.13 |
| 16 | claude-3-opus-20240229 | Anthropic | 0.1 |
| 17 | o1-preview-2024-09-12 | OpenAI | 0.1 |
| 18 | o1-mini-2024-09-12_medium | OpenAI | 0.1 |
| 19 | Llama-3.1-405B-Instruct | Meta AI | 0.07 |
| 20 | gemini-1.5-pro-001-feb24 | Google DeepMind | 0.07 |
| 21 | Mixtral-8x22B-Instruct-v0.1 | Mistral AI | 0.07 |
| 22 | Meta-Llama-3-70B-Instruct | Meta AI | 0.05 |
More benchmarks
SourceEpoch AI, 'AI Benchmarking Hub'. Published online at epoch.ai. Retrieved 2026-07-29 from https://epoch.ai/benchmarks. Licensed under CC BY 4.0.
← All benchmarks