benchgap
Calibration

CharXiv → RealWorldQA

RealWorldQA is estimated from CharXiv with a linear curve fitted on 7 models measured on both: y = 0.2136·x + 0.6808, R² = 0.72, cross-validated error 1.1 pp. It is used for 16 estimates.

Estimated modelCharXivRealWorldQASource
Claude Mythos 593.5%88.0%estimated ± 1.1 pp, medium confidence
Claude Opus 4.7 (Adaptive)91.0%87.5%estimated ± 1.1 pp, medium confidence
Claude Opus 4.889.9%87.3%estimated ± 1.1 pp, medium confidence
Claude Sonnet 4.677.4%84.6%estimated ± 1.1 pp, low confidence
Claude Sonnet 588.3%86.9%estimated ± 1.1 pp, medium confidence
Gemini 3.1 Flash-Lite73.2%83.7%estimated ± 1.1 pp, low confidence
Gemini 3.5 Flash84.2%86.1%estimated ± 1.1 pp, medium confidence
Gemini 3.7 Flash88.7%87.0%estimated ± 1.1 pp, medium confidence
GLM-5.3-Flash89.4%87.2%estimated ± 1.1 pp, medium confidence
Inkling82.0%85.6%estimated ± 1.1 pp, medium confidence
Inkling-Small81.3%85.4%estimated ± 1.1 pp, medium confidence
MiMo-V2.581.0%85.4%estimated ± 1.1 pp, medium confidence
Muse Glimmer 30B78.8%84.9%estimated ± 1.1 pp, medium confidence
Muse Spark 1.188.4%87.0%estimated ± 1.1 pp, medium confidence
Sakana Fugu85.1%86.3%estimated ± 1.1 pp, medium confidence
Sakana Fugu-Ultra86.6%86.6%estimated ± 1.1 pp, medium confidence