| Benchmark | Score | Top model |
|---|---|---|
| Agents' Last Exam% | 25.7% | GPT 6 Astra |
| Automation Bench v1.0.6% | 43.2% | DeepSeek V4.1 Flash |
| Codeforces (Rating) | 3348 | DeepSeek V4.1 Flash |
| CyberGym% | 83.3% | MiMo V2.6 Flash |
| DeepSWE v1.1% | 62.7% | DeepSeek V4.1 Flash |
| ExploitGym% | 5.4% | GPT 6 Astra |
| GPQA Diamond% | 92.4% | GPT 5.6 Sol |
| HLE (with tools)% | 60.0% | Claude Opus 5.5 |
| MathArena Apex% | 65.3% | DeepSeek V4.1 Flash, Kimi K3 |
| NL2Repo% | 61.5% | Claude Opus 5 |
| ProgramBench% | 15.5% | Claude Opus 5 |
| SEC Bench Pro% | 56.4% | GPT 6 Astra |
| Terminal Bench 2.1% | 87.9% | Claude Fable 5.1 |
| Terminal Bench 3.0% | 11.8% | Claude Opus 5 |
| Terminal Bench 4.0% | 12.4% | Claude Opus 5.5 |
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.