Calibration
Toolathlon → ResearchClawBench
ResearchClawBench is estimated from Toolathlon with a Michaelis–Menten curve fitted on 9 models measured on both: y = 0.2682·x / (0.25066 + x), R² = 0.40, cross-validated error 2.3 pp. It is used for 0 estimates.
Fitted on
| Model | Toolathlon | ResearchClawBench |
|---|---|---|
| Claude Opus 4.8 | 59.9% | 21.1% |
| Gemini 3.5 Flash | 56.5% | 18.0% |
| GLM-5.2 | 48.2% | 20.7% |
| GPT-5.4 | 54.6% | 15.3% |
| GPT-5.5 | 55.6% | 17.0% |
| Kimi K2.6 | 50.0% | 18.0% |
| Kimi K2.5 | 27.8% | 14.0% |
| Qwen3.5 397B | 36.3% | 14.2% |
| Qwen3.6 Plus | 39.8% | 18.0% |