Calibration
WideResearch → SkillsBench
SkillsBench is estimated from WideResearch with a Hill curve fitted on 6 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.80088^6.00 + x^6.00), R² = 0.81, cross-validated error 4.9 pp. It is used for 8 estimates.
| Estimated model | WideResearch | SkillsBench | Source |
|---|---|---|---|
| Kimi K2.6 | 80.8% | 61.6% | estimated ± 4.9 pp, medium confidence |
| Ornith-1.5-397B | 80.8% | 61.6% | estimated ± 4.9 pp, medium confidence |
| Claude Opus 4.5 | 76.4% | 51.6% | estimated ± 4.9 pp, medium confidence |
| Qwen3.6 Plus | 74.3% | 46.7% | estimated ± 4.9 pp, medium confidence |
| Qwen3.5 397B | 74.0% | 46.0% | estimated ± 4.9 pp, medium confidence |
| Kimi K2.5 | 72.7% | 43.1% | estimated ± 4.9 pp, low confidence |
| GLM-5 | 69.8% | 36.6% | estimated ± 4.9 pp, low confidence |
| Ornith-1.5-35B-A3B | 67.8% | 32.3% | estimated ± 4.9 pp, low confidence |
Fitted on
| Model | WideResearch | SkillsBench |
|---|---|---|
| Atria Dawn Preview | 81.9% | 66.4% |
| dots3-note Preview | 78.9% | 52.8% |
| Hy4 preview | 83.9% | 62.9% |
| Ling 3.0 Flash | 73.6% | 44.8% |
| Ling 3.1 Flash | 83.3% | 68.7% |
| Qwen3.8 Max | 81.9% | 70.2% |