| HLE(44.0 on text-only subset) | 42.6% |
| Apex Bench v1.1(Apex Bench v1.1 — not MathArena Apex) | 72.1% |
| Terminal-Bench 4.0(measured with the Claude Code framework) | 42.3% |
| DeepSWE v1.1(Mini-SWE-V2 framework) | 77% |
| ProgramBench | 29.1% |
| NL2Repo-Bench | 68.2% |
| Automation-Bench | 60.5% |
| Agents' Last Exam | 35.4% |
| Chartography (w/tools) | 83% |
| BabyVision (w/tools) | 92.5% |
| ZeroBench-main (w/tools)(Pass@5) | 51% |