benchgap
Calibration

AA-HLE → LABBench2

LABBench2 is estimated from AA-HLE with a linear curve fitted on 7 models measured on both: y = 0.4257·x + 0.6276, R² = 0.64, cross-validated error 2.4 pp. It is used for 25 estimates.

Estimated modelAA-HLELABBench2Source
Claude Sonnet 5.555.0%86.2%estimated ± 2.4 pp, medium confidence
Muse Wonder Full53.6%85.6%estimated ± 2.4 pp, medium confidence
Kimi K3.1 Pro52.5%85.1%estimated ± 2.4 pp, medium confidence
Muse Wonder50.1%84.1%estimated ± 2.4 pp, medium confidence
MiMo-V2.6-Pro49.4%83.8%estimated ± 2.4 pp, medium confidence
MiniMax M3.1 Flash48.0%83.2%estimated ± 2.4 pp, medium confidence
Step 5 Preview46.5%82.6%estimated ± 2.4 pp, medium confidence
GPT-6.1 Terra46.2%82.4%estimated ± 2.4 pp, medium confidence
Claude Haiku 5.544.4%81.7%estimated ± 2.4 pp, medium confidence
Grok 4.743.1%81.1%estimated ± 2.4 pp, medium confidence
Ling 3.1 Flash39.4%79.5%estimated ± 2.4 pp, low confidence
DeepSeek V4.1 Flash 101739.2%79.5%estimated ± 2.4 pp, low confidence
GPT-6 Luna38.5%79.2%estimated ± 2.4 pp, low confidence
MiMo-V2.6-Flash35.1%77.7%estimated ± 2.4 pp, low confidence
Beam 1.0 501B-A23B34.7%77.5%estimated ± 2.4 pp, low confidence
Apodex 1.134.1%77.3%estimated ± 2.4 pp, low confidence
Apodex 1.1 Mini34.1%77.3%estimated ± 2.4 pp, low confidence
Nemotron 3.5 Super VL 120B A12B30.7%75.8%estimated ± 2.4 pp, low confidence
Nemotron 3.5 Super VL 120B A12B NVFP430.7%75.8%estimated ± 2.4 pp, low confidence
A.X K229.6%75.4%estimated ± 2.4 pp, low confidence
GLM-529.3%75.2%estimated ± 2.4 pp, low confidence
GPT-5.4 nano28.3%74.8%estimated ± 2.4 pp, low confidence
GLM-5-Turbo27.8%74.6%estimated ± 2.4 pp, low confidence
GPT-5.1-Codex25.7%73.7%estimated ± 2.4 pp, low confidence
GPT-5.1-Codex-Max25.7%73.7%estimated ± 2.4 pp, low confidence

Fitted on

ModelAA-HLELABBench2
Claude Opus 554.9%84.2%
Claude Sonnet 541.3%80.1%
Gemini 3.7 Flash47.9%82.1%
Gemini 3.8 Flash47.8%86.2%
Gemini 4 Argon57.1%88.8%
GPT-5.6 Sol49.5%82.1%
GPT-5.6 Terra42.9%81.2%