Calibration

Vals MMLU-Pro → CritPt

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap
License: https://creativecommons.org/licenses/by-nc/4.0/

CritPt is estimated from Vals MMLU-Pro with a inverse Michaelis–Menten curve fitted on 46 models measured on both: y = 0.10573·(x − 0.7403) / (0.7403 + 0.2501 − x), R² = 0.40, cross-validated error 7.5 pp. It is used for 0 estimates.

Fitted on

Model	Vals MMLU-Pro	CritPt
Claude Fable 5	91.5%	28.6%
Claude Fable 5.1	92.4%	29.7%
Claude Opus 4.7	89.9%	5.1%
Claude Opus 4.8	89.6%	20.9%
Claude Opus 5	91.6%	29.1%
Claude Sonnet 4.6	87.3%	0.9%
Claude Sonnet 5	87.5%	16.9%
DeepSeek V4 Flash 0731	86.2%	16.6%
DeepSeek V4 Pro 0813	87.0%	18.0%
Gemini 3.1 Pro	91.0%	17.7%
Gemini 3.5 Flash	89.5%	13.1%
Gemini 3.5 Flash-Lite	85.8%	0.0%
Gemini 3.6 Flash	89.3%	10.6%
Gemini 3.7 Flash	90.1%	14.3%
Gemini 3.8 Flash	90.2%	18.3%
Gemini 3 Flash	88.6%	1.4%
GLM-4.6	82.2%	0.0%
GLM-4.7	82.7%	1.7%
GLM-5.1	86.9%	4.6%
GLM-5.2	86.7%	20.9%
GLM-5.3	86.8%	19.1%
GPT-5.4 mini	84.6%	10.0%
GPT-5.4 nano	77.2%	9.3%
GPT-5.5	88.1%	27.1%
GPT-5.6 Luna	86.0%	20.6%
GPT-5.6 Sol	89.1%	32.3%
GPT-5.6 Terra	86.7%	30.0%
Grok 4.3	85.8%	8.0%
Grok 4.5	89.2%	15.4%
Grok 4.6	89.4%	17.1%
Inkling	86.3%	5.4%
Inkling-Small	85.6%	8.3%
Kimi K2.6	87.6%	8.0%
Kimi K3	88.0%	23.4%
Ling 3.0 Flash	82.0%	1.7%
MiMo-V2.5-Pro	84.6%	4.0%
MiniMax M2.7	80.4%	0.6%
MiniMax M3	84.2%	3.7%
Mistral Medium 3.5 128B	75.3%	0.0%
Muse Spark	87.3%	11.3%
Muse Spark 1.1	88.7%	15.1%
Muse Spark 1.2	88.3%	17.7%
Nemotron 3 Ultra	85.8%	3.1%
Qwen3.6 Plus	87.7%	2.9%
Qwen3.7 Max	89.3%	13.4%
Qwen3.8-27B	84.3%	5.4%