Calibration

Vals GPQA Diamond → Vals MMLU-Pro

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

Vals MMLU-Pro is estimated from Vals GPQA Diamond with a Hill curve fitted on 55 models measured on both: y = 0.7083 + (0.9866 − 0.7083)·x^6.00 / (0.83901^6.00 + x^6.00), R² = 0.82, cross-validated error 2.2 pp. It is used for 0 estimates.

Fitted on

Model	Vals GPQA Diamond	Vals MMLU-Pro
Claude Fable 5	93.2%	91.5%
Claude Fable 5.1	93.4%	92.4%
Claude Haiku 4.5	72.2%	78.7%
Claude Opus 4.7	90.2%	89.9%
Claude Opus 4.8	92.4%	89.6%
Claude Opus 5	93.4%	91.6%
Claude Sonnet 4.6	85.6%	87.3%
Claude Sonnet 5	88.9%	87.5%
DeepSeek V4 Flash 0731	89.9%	86.2%
DeepSeek V4 Pro 0813	92.4%	87.0%
Gemini 3.1 Flash-Lite	81.1%	86.2%
Gemini 3.1 Pro	95.5%	91.0%
Gemini 3.5 Flash	92.7%	89.5%
Gemini 3.5 Flash-Lite	83.8%	85.8%
Gemini 3.6 Flash	93.4%	89.3%
Gemini 3.7 Flash	93.9%	90.1%
Gemini 3.8 Flash	94.4%	90.2%
Gemini 3 Flash	87.9%	88.6%
GLM-4.5	72.2%	81.2%
GLM-4.6	74.5%	82.2%
GLM-4.7	80.1%	82.7%
GLM-5.1	84.5%	86.9%
GLM-5.2	85.6%	86.7%
GLM-5.3	88.1%	86.8%
GLM-5.3-Flash	86.4%	86.1%
GPT-5.4 mini	83.1%	84.6%
GPT-5.4 nano	77.5%	77.2%
GPT-5.5	93.2%	88.1%
GPT-5.6 Luna	91.7%	86.0%
GPT-5.6 Sol	95.2%	89.1%
GPT-5.6 Terra	90.9%	86.7%
Grok 4.20	88.6%	86.3%
Grok 4.3	91.4%	85.8%
Grok 4.5	92.9%	89.2%
Grok 4.6	94.7%	89.4%
Inkling	87.1%	86.3%
Inkling-Small	83.6%	85.6%
Kimi K2.6	89.1%	87.6%
Kimi K3	92.9%	88.0%
Laguna M.1	27.0%	68.8%
Laguna XS.2	55.1%	69.1%
Ling 3.0 Flash	84.8%	82.0%
MiMo-V2.5	81.6%	82.9%
MiMo-V2.5-Pro	82.6%	84.6%
MiniMax M2.7	86.6%	80.4%
MiniMax M3	92.7%	84.2%
Mistral Medium 3.5 128B	34.8%	75.3%
Muse Spark	89.6%	87.3%
Muse Spark 1.1	91.2%	88.7%
Nemotron 3 Ultra	86.1%	85.8%
Qwen3.5 Flash	82.8%	84.1%
Qwen3.6 Plus	87.4%	87.7%
Qwen3.7 Max	90.2%	89.3%
Qwen3.8-27B	88.9%	84.3%
Qwen3.8 Max	93.7%	88.6%