benchgap
Calibration

Vals GPQA Diamond → CritPt

CritPt is estimated from Vals GPQA Diamond with a inverse Michaelis–Menten curve fitted on 46 models measured on both: y = 0.13275·(x − 0.7437) / (0.7437 + 0.3212 − x), R² = 0.50, cross-validated error 7.2 pp. It is used for 8 estimates.

Estimated modelVals GPQA DiamondCritPtSource
GLM-5.3-Flash86.4%7.9%estimated ± 7.2 pp, medium confidence
Qwen3.5 Flash82.8%4.7%estimated ± 7.2 pp, medium confidence
MiMo-V2.581.6%3.9%estimated ± 7.2 pp, medium confidence
Gemini 3.1 Flash-Lite81.1%3.5%estimated ± 7.2 pp, medium confidence
Claude Haiku 4.572.2%0.0%estimated ± 7.2 pp, medium confidence
GLM-4.572.2%0.0%estimated ± 7.2 pp, medium confidence
Laguna M.127.0%0.0%estimated ± 7.2 pp, low confidence
Laguna XS.255.1%0.0%estimated ± 7.2 pp, medium confidence

Fitted on

ModelVals GPQA DiamondCritPt
Claude Fable 593.2%28.6%
Claude Fable 5.193.4%29.7%
Claude Opus 4.790.2%5.1%
Claude Opus 4.892.4%20.9%
Claude Opus 593.4%29.1%
Claude Sonnet 4.685.6%0.9%
Claude Sonnet 588.9%16.9%
DeepSeek V4 Flash 073189.9%16.6%
DeepSeek V4 Pro 081392.4%18.0%
Gemini 3.1 Pro95.5%17.7%
Gemini 3.5 Flash92.7%13.1%
Gemini 3.5 Flash-Lite83.8%0.0%
Gemini 3.6 Flash93.4%10.6%
Gemini 3.7 Flash93.9%14.3%
Gemini 3.8 Flash94.4%18.3%
Gemini 3 Flash87.9%1.4%
GLM-4.674.5%0.0%
GLM-4.780.1%1.7%
GLM-5.184.5%4.6%
GLM-5.285.6%20.9%
GLM-5.388.1%19.1%
GPT-5.4 mini83.1%10.0%
GPT-5.4 nano77.5%9.3%
GPT-5.593.2%27.1%
GPT-5.6 Luna91.7%20.6%
GPT-5.6 Sol95.2%32.3%
GPT-5.6 Terra90.9%30.0%
Grok 4.391.4%8.0%
Grok 4.592.9%15.4%
Grok 4.694.7%17.1%
Inkling87.1%5.4%
Inkling-Small83.6%8.3%
Kimi K2.689.1%8.0%
Kimi K392.9%23.4%
Ling 3.0 Flash84.8%1.7%
MiMo-V2.5-Pro82.6%4.0%
MiMo-V2-Flash59.3%0.0%
MiniMax M2.786.6%0.6%
MiniMax M392.7%3.7%
Mistral Medium 3.5 128B34.8%0.0%
Muse Spark89.6%11.3%
Muse Spark 1.191.2%15.1%
Nemotron 3 Ultra86.1%3.1%
Qwen3.6 Plus87.4%2.9%
Qwen3.7 Max90.2%13.4%
Qwen3.8-27B88.9%5.4%