benchgap
Calibration

Vals MMLU-Pro → AA-Omniscience Accuracy

AA-Omniscience Accuracy is estimated from Vals MMLU-Pro with a offset logistic curve fitted on 45 models measured on both: y = 0.2215 + (0.9024 − 0.2215) / (1 + exp(−34.58·(x − 0.9052))), R² = 0.65, cross-validated error 7.7 pp. It is used for 9 estimates.

Estimated modelVals MMLU-ProAA-Omniscience AccuracySource
Qwen3.8 Max88.6%45.3%estimated ± 7.7 pp, medium confidence
Gemini 3.1 Flash-Lite86.2%34.6%estimated ± 7.7 pp, medium confidence
GLM-5.3-Flash86.1%34.3%estimated ± 7.7 pp, medium confidence
Qwen3.5 Flash84.1%28.8%estimated ± 7.7 pp, medium confidence
MiMo-V2.582.9%26.7%estimated ± 7.7 pp, medium confidence
GLM-4.581.2%24.8%estimated ± 7.7 pp, medium confidence
Claude Haiku 4.578.7%23.3%estimated ± 7.7 pp, medium confidence
Laguna XS.269.1%22.2%estimated ± 7.7 pp, low confidence
Laguna M.168.8%22.2%estimated ± 7.7 pp, low confidence

Fitted on

ModelVals MMLU-ProAA-Omniscience Accuracy
Claude Fable 591.5%65.4%
Claude Fable 5.192.4%68.4%
Claude Opus 4.789.9%44.7%
Claude Opus 4.889.6%48.8%
Claude Opus 591.6%60.9%
Claude Sonnet 4.687.3%38.6%
Claude Sonnet 587.5%40.1%
DeepSeek V4 Flash 073186.2%40.4%
DeepSeek V4 Pro 081387.0%49.1%
Gemini 3.5 Flash89.5%51.9%
Gemini 3.5 Flash-Lite85.8%26.5%
Gemini 3.6 Flash89.3%50.0%
Gemini 3.7 Flash90.1%55.3%
Gemini 3.8 Flash90.2%50.5%
Gemini 3 Flash88.6%45.8%
GLM-4.682.2%21.4%
GLM-4.782.7%29.3%
GLM-5.186.9%23.7%
GLM-5.286.7%24.3%
GLM-5.386.8%32.3%
GPT-5.4 mini84.6%37.5%
GPT-5.4 nano77.2%25.7%
GPT-5.588.1%58.0%
GPT-5.6 Luna86.0%42.7%
GPT-5.6 Sol89.1%59.4%
GPT-5.6 Terra86.7%46.8%
Grok 4.385.8%34.6%
Grok 4.589.2%51.6%
Grok 4.689.4%48.2%
Inkling86.3%39.3%
Inkling-Small85.6%31.8%
Kimi K2.687.6%32.6%
Kimi K388.0%47.6%
Ling 3.0 Flash82.0%18.2%
MiMo-V2.5-Pro84.6%22.4%
MiniMax M2.780.4%26.8%
MiniMax M384.2%44.9%
Mistral Medium 3.5 128B75.3%23.0%
Muse Spark87.3%49.6%
Muse Spark 1.188.7%52.1%
Muse Spark 1.288.3%45.4%
Nemotron 3 Ultra85.8%21.6%
Qwen3.6 Plus87.7%26.4%
Qwen3.7 Max89.3%31.1%
Qwen3.8-27B84.3%17.3%