benchgap
Calibration

GPQA → Vals GPQA Diamond

Vals GPQA Diamond is estimated from GPQA with a inverse Michaelis–Menten curve fitted on 25 models measured on both: y = 1.39188·(x − 0.1281) / (0.1281 + 2.0000 − x), R² = 0.62, cross-validated error 4.8 pp. It is used for 25 estimates.

Estimated modelGPQAVals GPQA DiamondSource
Claude 3.5 Sonnet59.4%42.3%estimated ± 4.8 pp, medium confidence
Claude Opus 4.587.0%82.1%estimated ± 4.8 pp, high confidence
DeepSeek V359.1%41.9%estimated ± 4.8 pp, medium confidence
Gemini 2.5 Pro83.0%75.3%estimated ± 4.8 pp, high confidence
Gemma 4 E2B43.4%25.1%estimated ± 4.8 pp, medium confidence
Gemma 4 E4B58.6%41.3%estimated ± 4.8 pp, medium confidence
GPT-4.166.3%50.8%estimated ± 4.8 pp, medium confidence
GPT-4.1 mini64.2%48.1%estimated ± 4.8 pp, medium confidence
GPT-4.1 nano50.3%32.1%estimated ± 4.8 pp, medium confidence
Granite 4.2 30B66.4%51.0%estimated ± 4.8 pp, medium confidence
Granite 4.2 3B54.8%37.0%estimated ± 4.8 pp, medium confidence
Granite 4.2 8B64.1%48.1%estimated ± 4.8 pp, medium confidence
Kimi K2.5 (Reasoning)87.6%83.1%estimated ± 4.8 pp, high confidence
LFM2.5-VL-450M25.7%9.6%estimated ± 4.8 pp, medium confidence
Ling 2.6 Flash59.0%41.8%estimated ± 4.8 pp, medium confidence
o175.7%63.8%estimated ± 4.8 pp, medium confidence
o1-pro79.0%68.8%estimated ± 4.8 pp, medium confidence
o3-mini77.2%66.1%estimated ± 4.8 pp, medium confidence
Qwen3 235B 250777.5%66.5%estimated ± 4.8 pp, medium confidence
Qwen3.5-122B-A10B86.6%81.4%estimated ± 4.8 pp, high confidence
Qwen3.5-27B85.5%79.5%estimated ± 4.8 pp, high confidence
Qwen3.5-35B-A3B84.2%77.3%estimated ± 4.8 pp, high confidence
Qwen3.5 397B88.4%84.6%estimated ± 4.8 pp, high confidence
Qwen3.6-27B87.8%83.5%estimated ± 4.8 pp, high confidence
Qwen3.6-35B-A3B86.0%80.3%estimated ± 4.8 pp, high confidence