Calibration
ResearchClawBench → Claw-Eval
Claw-Eval is estimated from ResearchClawBench with a Hill curve fitted on 12 models measured on both: y = 0.5322 + (1.2000 − 0.5322)·x^6.00 / (0.23818^6.00 + x^6.00), R² = 0.80, cross-validated error 3.2 pp. It is used for 7 estimates.
| Estimated model | ResearchClawBench | Claw-Eval | Source |
|---|---|---|---|
| Claude Opus 4.7 | 20.7% | 73.3% | estimated ± 3.2 pp, medium confidence |
| Claude Opus 4.8 | 21.1% | 75.0% | estimated ± 3.2 pp, medium confidence |
| Gemini 3.5 Flash | 18.0% | 63.7% | estimated ± 3.2 pp, high confidence |
| GLM-5.2 | 20.7% | 73.3% | estimated ± 3.2 pp, medium confidence |
| GPT-5.5 | 17.0% | 61.0% | estimated ± 3.2 pp, high confidence |
| Grok 4.1 | 13.5% | 55.4% | estimated ± 3.2 pp, high confidence |
| Grok 4.3 | 12.4% | 54.5% | estimated ± 3.2 pp, medium confidence |