benchgap
Calibration

SuperGPQA → AA-HLE

AA-HLE is estimated from SuperGPQA with a offset logistic curve fitted on 17 models measured on both: y = 0.0801 + (0.2669 − 0.0801) / (1 + exp(−70.54·(x − 0.6260))), R² = 0.44, cross-validated error 7.6 pp. It is used for 2 estimates.

Estimated modelSuperGPQAAA-HLESource
Seed 2.1 Pro70.8%26.6%estimated ± 7.6 pp, low confidence
Seed 2.1 Turbo67.4%26.1%estimated ± 7.6 pp, low confidence

Fitted on

ModelSuperGPQAAA-HLE
Claude Opus 4.570.6%13.2%
Claude Opus 4.695.0%19.1%
Claude Sonnet 4.695.0%13.3%
GLM-566.8%29.3%
Kimi K257.2%7.4%
Kimi K2.569.2%30.7%
MiniCPM5-2B40.8%8.9%
Qwen3.5-122B-A10B67.1%25.2%
Qwen3.5-27B65.6%23.9%
Qwen3.5-35B-A3B63.4%21.0%
Qwen3.5 397B70.4%29.0%
Qwen3.6-27B66.0%23.1%
Qwen3.6-35B-A3B64.7%22.2%
Qwen 3.6 Max (preview)73.9%30.8%
Qwen3.6 Plus71.6%27.8%
Qwen3.7 Max73.6%40.5%
Qwen3.7 Plus71.4%35.6%