Calibration
ResearchClawBench → Toolathlon
Toolathlon is estimated from ResearchClawBench with a offset logistic curve fitted on 9 models measured on both: y = 0.0000 + (0.5240 − 0.0000) / (1 + exp(−200.00·(x − 0.1385))), R² = 0.68, cross-validated error 6.9 pp. It is used for 3 estimates.
| Estimated model | ResearchClawBench | Toolathlon | Source |
|---|---|---|---|
| MiMo-V2.5 | 16.9% | 52.3% | estimated ± 6.9 pp, medium confidence |
| MiMo-V2-Pro | 15.3% | 49.7% | estimated ± 6.9 pp, medium confidence |
| Grok 4.1 | 13.5% | 17.3% | estimated ± 6.9 pp, low confidence |
Fitted on
| Model | ResearchClawBench | Toolathlon |
|---|---|---|
| Claude Opus 4.8 | 21.1% | 59.9% |
| Gemini 3.5 Flash | 18.0% | 56.5% |
| GLM-5.2 | 20.7% | 48.2% |
| GPT-5.4 | 15.3% | 54.6% |
| GPT-5.5 | 17.0% | 55.6% |
| Kimi K2.6 | 18.0% | 50.0% |
| Kimi K2.5 | 14.0% | 27.8% |
| Qwen3.5 397B | 14.2% | 36.3% |
| Qwen3.6 Plus | 18.0% | 39.8% |