benchgap
Calibration

WideResearch → SkillsBench

SkillsBench is estimated from WideResearch with a Hill curve fitted on 5 models measured on both: y = 0.0000 + (1.2000 − 0.0000)·x^6.00 / (0.80235^6.00 + x^6.00), R² = 0.79, cross-validated error 5.6 pp. It is used for 10 estimates.

Estimated modelWideResearchSkillsBenchSource
Claude Opus 4.576.4%51.2%estimated ± 5.6 pp, low confidence
GLM-569.8%36.3%estimated ± 5.6 pp, low confidence
Kimi K2.680.8%61.3%estimated ± 5.6 pp, low confidence
Kimi K2.572.7%42.8%estimated ± 5.6 pp, low confidence
Ornith-1.5-35B-A3B67.8%32.0%estimated ± 5.6 pp, low confidence
Ornith-1.5-397B80.8%61.3%estimated ± 5.6 pp, low confidence
Ornith-1.5-9B59.5%17.1%estimated ± 5.6 pp, low confidence
Qwen3.5 397B74.0%45.7%estimated ± 5.6 pp, low confidence
Qwen3.6-35B-A3B60.1%18.0%estimated ± 5.6 pp, low confidence
Qwen3.6 Plus74.3%46.4%estimated ± 5.6 pp, low confidence