benchgap
Calibration

GPQA → CritPt

CritPt is estimated from GPQA with a offset logistic curve fitted on 61 models measured on both: y = 0.0056 + (0.4240 − 0.0056) / (1 + exp(−42.61·(x − 0.9340))), R² = 0.78, cross-validated error 4.1 pp. It is used for 21 estimates.

Estimated modelGPQACritPtSource
Claude 3.5 Sonnet59.4%0.6%estimated ± 4.1 pp, high confidence
Claude Mythos 594.1%24.6%estimated ± 4.1 pp, high confidence
Claude Sonnet 4.583.4%1.1%estimated ± 4.1 pp, high confidence
Interfaze Beta89.9%8.2%estimated ± 4.1 pp, high confidence
LFM2.5-230M25.4%0.6%estimated ± 4.1 pp, medium confidence
LFM2.5-VL-450M25.7%0.6%estimated ± 4.1 pp, medium confidence
MAI-Thinking-184.2%1.4%estimated ± 4.1 pp, high confidence
Mellum2-12B-A2.5B-Instruct40.9%0.6%estimated ± 4.1 pp, medium confidence
Mellum2-12B-A2.5B-Thinking57.6%0.6%estimated ± 4.1 pp, high confidence
Ornith-1.5-35B-A3B89.2%6.5%estimated ± 4.1 pp, high confidence
Ornith-1.5-397B92.8%18.8%estimated ± 4.1 pp, high confidence
Ornith-1.5-9B86.4%2.6%estimated ± 4.1 pp, high confidence
Qwen3 235B 250777.5%0.6%estimated ± 4.1 pp, high confidence
Qwen3.8 Max92.6%18.0%estimated ± 4.1 pp, high confidence
Qwen3.8-Omni-Flash91.0%11.6%estimated ± 4.1 pp, high confidence
Sakana Fugu95.5%30.3%estimated ± 4.1 pp, high confidence
Sakana Fugu-Ultra95.5%30.3%estimated ± 4.1 pp, high confidence
Soofi S 30B-A3B43.4%0.6%estimated ± 4.1 pp, high confidence
Ternary Bonsai 2 27B85.8%2.1%estimated ± 4.1 pp, high confidence
ZAYA1-74B-Preview57.3%0.6%estimated ± 4.1 pp, high confidence
ZAYA1-8B71.0%0.6%estimated ± 4.1 pp, high confidence