benchgap
Calibration

SimpleQA → GPQA

GPQA is estimated from SimpleQA with a Michaelis–Menten curve fitted on 14 models measured on both: y = 0.8069·x / (0.04361 + x), R² = 0.48, cross-validated error 9.1 pp. It is used for 0 estimates.

Fitted on

ModelSimpleQAGPQA
DeepSeek-R130.1%71.5%
DeepSeek-R1-052827.8%81.0%
DeepSeek V324.9%59.1%
Gemini 1.5 Flash8.6%50.0%
Gemini 2.0 Flash29.9%65.2%
Gemini 2.0 Flash-Lite16.5%50.5%
Gemini 2.5 Flash-Lite10.7%64.6%
GPT-4.562.5%69.5%
Grok 3 [Beta]43.6%75.4%
Kimi K231.0%75.1%
Mistral Small 3.110.4%46.0%
o1-mini7.6%60.0%
o4-mini20.2%77.6%
Qwen3 235B 250754.3%77.5%