Calibration

STEMBench → AA-Omniscience Accuracy

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap
License: https://creativecommons.org/licenses/by-nc/4.0/

AA-Omniscience Accuracy is estimated from STEMBench with a inverse Michaelis–Menten curve fitted on 8 models measured on both: y = 0.26672·(x − 0.0000) / (0.0000 + 0.0404 − x), R² = 0.38, cross-validated error 13.3 pp. It is used for 0 estimates.

Fitted on

Model	STEMBench	AA-Omniscience Accuracy
Claude Opus 5.5	2.6%	65.5%
DeepSeek V4.1 Flash 1017	2.5%	46.4%
Gemini 4 Argon	2.4%	48.3%
GPT-6.1 Sol	2.8%	61.8%
GPT-6 Astra	2.8%	61.6%
MiMo-V2.6-Pro	2.2%	34.8%
Mistral Large 4	2.5%	23.9%
Muse Wonder	2.7%	38.0%