| Benchmark | Score | Top model |
|---|---|---|
| Agents' Last Exam% | 27.6% | GPT 6 Astra |
| Automation Bench v1.0.6% | 52.3% | DeepSeek V4.1 Flash |
| CyberGym% | 95.1% | MiMo V2.6 Flash |
| DeepSWE v1.1% | 67.9% | DeepSeek V4.1 Flash |
| ExploitBench% | 25.3% | GPT 6 Astra |
| ExploitGym% | 6.0% | GPT 6 Astra |
| JobBench% | 61.2% | Claude Opus 5 |
| OSWorld-Verified% | 80.8% | Claude Fable 5 |
| ProgramBench% | 26.0% | Claude Opus 5 |
| SEC Bench Pro% | 47.5% | GPT 6 Astra |
| Terminal Bench 2.1% | 87.6% | Claude Fable 5.1 |
| Terminal Bench 4.0% | 28.8% | Claude Opus 5.5 |
| Toolathlon-verified% | 73.6% | Claude Opus 5 |
Figures are taken from public model cards, technical reports and vendor announcements, and are approximate where sources disagree. Results from different benchmark versions or tool-use settings never share a column: Terminal Bench 2.1, 3.0 and DeepSWE v1.1 are distinct benchmarks, and HLE with tools is separate from HLE without tools. “Top model” is the highest scoring model with a published result on that benchmark; when several models tie, all of them are listed. Benchmarks this model has no published result for are left out; the line above shows how many of the benchmarks are covered.