benchgap
Calibration

CharXiv → MMMU-Pro

MMMU-Pro is estimated from CharXiv with a Michaelis–Menten curve fitted on 23 models measured on both: y = 1.2710·x / (0.50191 + x), R² = 0.64, cross-validated error 3.2 pp. It is used for 6 estimates.

Estimated modelCharXivMMMU-ProSource
Claude Sonnet 4.677.4%77.1%estimated ± 3.2 pp, high confidence
Gemini 3.1 Flash-Lite73.2%75.4%estimated ± 3.2 pp, high confidence
GLM-5.3-Flash89.4%81.4%estimated ± 3.2 pp, high confidence
Muse Spark 1.188.4%81.1%estimated ± 3.2 pp, high confidence
Sakana Fugu85.1%79.9%estimated ± 3.2 pp, high confidence
Sakana Fugu-Ultra86.6%80.5%estimated ± 3.2 pp, high confidence