benchgap
Calibration

SuperGPQA → HLE

HLE is estimated from SuperGPQA with a offset logistic curve fitted on 14 models measured on both: y = 0.0585 + (0.5113 − 0.0585) / (1 + exp(−19.17·(x − 0.6886))), R² = 0.94, cross-validated error 4.8 pp. It is used for 8 estimates.

Estimated modelSuperGPQAHLESource
Qwen 3.6 Max (preview)73.9%38.6%estimated ± 4.8 pp, high confidence
Seed 2.1 Pro70.8%32.6%estimated ± 4.8 pp, high confidence
Seed 2.1 Turbo67.4%25.3%estimated ± 4.8 pp, high confidence
Qwen3.5-122B-A10B67.1%24.7%estimated ± 4.8 pp, high confidence
Qwen3.5-27B65.6%21.6%estimated ± 4.8 pp, high confidence
Qwen3.5-35B-A3B63.4%17.6%estimated ± 4.8 pp, high confidence
Qwen3 235B 250762.6%16.3%estimated ± 4.8 pp, high confidence
MiniCPM5-1B23.1%5.9%estimated ± 4.8 pp, medium confidence

Fitted on

ModelSuperGPQAHLE
Claude Opus 4.570.6%30.8%
Claude Opus 4.695.0%53.0%
Claude Sonnet 4.695.0%49.0%
GLM-566.8%30.5%
Kimi K257.2%4.7%
Kimi K2.569.2%30.1%
MiniCPM5-2B40.8%8.9%
Qwen3 235B 2507 (Reasoning)64.9%18.2%
Qwen3.5 397B70.4%28.7%
Qwen3.6-27B66.0%24.0%
Qwen3.6-35B-A3B64.7%21.4%
Qwen3.6 Plus71.6%28.8%
Qwen3.7 Max73.6%41.4%
Qwen3.7 Plus71.4%34.7%