benchgap
Calibration

AA-Omniscience Accuracy → HLE-Verified

HLE-Verified is estimated from AA-Omniscience Accuracy with a inverse Michaelis–Menten curve fitted on 6 models measured on both: y = 1.36858·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.65, cross-validated error 6.0 pp. It is used for 9 estimates.

Estimated modelAA-Omniscience AccuracyHLE-VerifiedSource
Claude Sonnet 5.554.0%50.6%estimated ± 6.0 pp, low confidence
DeepSeek V4.1 Flash46.4%41.3%estimated ± 6.0 pp, low confidence
Grok 4.747.4%42.5%estimated ± 6.0 pp, low confidence
Ling 3.1 Flash29.1%23.3%estimated ± 6.0 pp, low confidence
Mercury 2.522.0%16.9%estimated ± 6.0 pp, low confidence
MiMo-V2.6-Flash27.0%21.4%estimated ± 6.0 pp, low confidence
MiMo-V2.6-Pro34.8%28.8%estimated ± 6.0 pp, low confidence
Mistral Large 425.8%20.3%estimated ± 6.0 pp, low confidence
Step 5 Preview41.5%35.8%estimated ± 6.0 pp, low confidence