benchgap
Calibration

WideResearch → HLE w/ tools

HLE w/ tools is estimated from WideResearch with a offset logistic curve fitted on 6 models measured on both: y = 0.3041 + (0.5702 − 0.3041) / (1 + exp(−35.54·(x − 0.7374))), R² = 1.00, cross-validated error 2.4 pp. It is used for 9 estimates.

Estimated modelWideResearchHLE w/ toolsSource
Atria Dawn Preview81.9%55.6%estimated ± 2.4 pp, medium confidence
Claude Opus 4.576.4%49.6%estimated ± 2.4 pp, medium confidence
GLM-569.8%35.7%estimated ± 2.4 pp, medium confidence
Kimi K2.680.8%55.0%estimated ± 2.4 pp, medium confidence
Kimi K2.572.7%41.3%estimated ± 2.4 pp, medium confidence
Ling 3.0 Flash73.6%43.4%estimated ± 2.4 pp, medium confidence
Qwen3.5 397B74.0%44.3%estimated ± 2.4 pp, medium confidence
Qwen3.6-35B-A3B60.1%30.6%estimated ± 2.4 pp, medium confidence
Qwen3.6 Plus74.3%45.0%estimated ± 2.4 pp, medium confidence