benchgap
Calibration

BrowseComp → AA Cyber Index

AA Cyber Index is estimated from BrowseComp with a Michaelis–Menten + offset curve fitted on 8 models measured on both: y = 0.0000 + 2.0000·x / (3.35358 + x), R² = 0.40, cross-validated error 13.5 pp. It is used for 21 estimates.

Estimated modelBrowseCompAA Cyber IndexSource
Inkling 2 Small90.2%42.4%estimated ± 13.5 pp, low confidence
GPT-5.5 Pro90.1%42.4%estimated ± 13.5 pp, low confidence
GPT-5.4 Pro89.3%42.1%estimated ± 13.5 pp, low confidence
Claude Mythos 588.0%41.6%estimated ± 13.5 pp, low confidence
Claude Opus 4.683.7%39.9%estimated ± 13.5 pp, low confidence
GPT-5.482.7%39.6%estimated ± 13.5 pp, low confidence
Nemotron 3.5 Super VL 120B A12B78.0%37.7%estimated ± 13.5 pp, low confidence
Nemotron 3.5 Super VL 120B A12B NVFP477.5%37.5%estimated ± 13.5 pp, low confidence
Nemotron 3.5 Bolt 74B-A9B BF1670.0%34.5%estimated ± 13.5 pp, low confidence
Qwen3.5 397B69.0%34.1%estimated ± 13.5 pp, low confidence
Agents-A1-4B66.8%33.2%estimated ± 13.5 pp, low confidence
Nemotron 3.5 Bolt 74B-A9B66.8%33.2%estimated ± 13.5 pp, low confidence
GPT-5.265.8%32.8%estimated ± 13.5 pp, low confidence
Qwen3.5-27B61.0%30.8%estimated ± 13.5 pp, low confidence
Qwen3.5-35B-A3B61.0%30.8%estimated ± 13.5 pp, low confidence
Kimi K2.560.6%30.6%estimated ± 13.5 pp, low confidence
GLM-4.752.0%26.8%estimated ± 13.5 pp, low confidence
Sarvam 105B49.5%25.7%estimated ± 13.5 pp, low confidence
Solar Pro 449.2%25.6%estimated ± 13.5 pp, low confidence
LongCat-Flash-Lite-Sparse48.6%25.3%estimated ± 13.5 pp, low confidence
Sarvam 30B35.5%19.1%estimated ± 13.5 pp, low confidence

Fitted on

ModelBrowseCompAA Cyber Index
GPT-6 Astra91.5%33.4%
Kimi K3.1 Pro92.1%64.1%
Kimi K391.2%41.3%
MiniMax M383.5%22.4%
Mistral Large 484.5%50.9%
Nemotron 3 Ultra44.4%13.1%
Beam 1.0 501B-A23B76.8%33.7%
Step 5 Preview88.7%47.7%