benchgap
Calibration

SuperGPQA → HLE

HLE is estimated from SuperGPQA with a linear curve fitted on 12 models measured on both: y = 0.7717·x + -0.2155, R² = 0.68, cross-validated error 7.8 pp. It is used for 7 estimates.

Estimated modelSuperGPQAHLESource
Qwen3 235B 250762.6%26.8%estimated ± 7.8 pp, medium confidence
Qwen3.5-122B-A10B67.1%30.2%estimated ± 7.8 pp, medium confidence
Qwen3.5-27B65.6%29.1%estimated ± 7.8 pp, medium confidence
Qwen3.5-35B-A3B63.4%27.4%estimated ± 7.8 pp, medium confidence
Qwen 3.6 Max (preview)73.9%35.5%estimated ± 7.8 pp, medium confidence
Seed 2.1 Pro70.8%33.1%estimated ± 7.8 pp, medium confidence
Seed 2.1 Turbo67.4%30.5%estimated ± 7.8 pp, medium confidence