benchgap
Calibration

MMMU-Pro → PerceptionBench

PerceptionBench is estimated from MMMU-Pro with a linear curve fitted on 5 models measured on both: y = 2.9882·x + -1.8258, R² = 0.85, cross-validated error 2.6 pp. It is used for 34 estimates.

Estimated modelMMMU-ProPerceptionBenchSource
GPT-6.1 Terra85.5%72.9%estimated ± 2.6 pp, low confidence
Gemini 3.1 Pro83.9%68.1%estimated ± 2.6 pp, medium confidence
Gemini 3.5 Flash83.6%67.2%estimated ± 2.6 pp, medium confidence
GPT-5.6 Sol83.0%65.4%estimated ± 2.6 pp, medium confidence
Seed 2.1 Pro81.6%61.3%estimated ± 2.6 pp, medium confidence
GPT-5.481.2%60.1%estimated ± 2.6 pp, medium confidence
GPT-5.581.2%60.1%estimated ± 2.6 pp, medium confidence
Gemini 3 Pro81.0%59.5%estimated ± 2.6 pp, medium confidence
GPT-5.6 Terra80.7%58.6%estimated ± 2.6 pp, medium confidence
Muse Spark80.4%57.7%estimated ± 2.6 pp, medium confidence
Seed 2.1 Turbo80.1%56.8%estimated ± 2.6 pp, medium confidence
GPT-5.279.5%55.0%estimated ± 2.6 pp, medium confidence
Kimi K2.679.4%54.7%estimated ± 2.6 pp, medium confidence
Qwen3.5 397B79.0%53.5%estimated ± 2.6 pp, low confidence
Qwen3.7 Plus79.0%53.5%estimated ± 2.6 pp, low confidence
Qwen3.6 Plus78.8%52.9%estimated ± 2.6 pp, low confidence
Hy478.6%52.3%estimated ± 2.6 pp, low confidence
Kimi K2.578.5%52.0%estimated ± 2.6 pp, low confidence
GPT-5.6 Luna78.4%51.7%estimated ± 2.6 pp, low confidence
Grok 4.378.1%50.8%estimated ± 2.6 pp, low confidence
MiniMax M378.1%50.8%estimated ± 2.6 pp, low confidence
Pareto 26.978.0%50.5%estimated ± 2.6 pp, low confidence
MiMo-V2.577.9%50.2%estimated ± 2.6 pp, low confidence
Claude Opus 4.677.3%48.4%estimated ± 2.6 pp, low confidence
Gemma 4 31B76.9%47.2%estimated ± 2.6 pp, low confidence
Qwen3.5-122B-A10B76.9%47.2%estimated ± 2.6 pp, low confidence
GPT-5.4 mini76.6%46.3%estimated ± 2.6 pp, low confidence
Step 5 Preview76.0%44.5%estimated ± 2.6 pp, low confidence
Hy4 flash75.8%43.9%estimated ± 2.6 pp, low confidence
Qwen3.6-27B75.8%43.9%estimated ± 2.6 pp, low confidence
Qwen3.6-35B-A3B75.3%42.4%estimated ± 2.6 pp, low confidence
Grok 4.2075.2%42.1%estimated ± 2.6 pp, low confidence
Qwen3.5-35B-A3B75.1%41.8%estimated ± 2.6 pp, low confidence
Qwen3.5-27B75.0%41.5%estimated ± 2.6 pp, low confidence

Fitted on

ModelMMMU-ProPerceptionBench
dots3-note Preview79.1%53.4%
Inkling 2 Small81.6%64.7%
Kimi K3.1 Pro84.1%68.3%
Kimi K381.6%58.5%
Qwen3.8 Max82.3%63.5%