benchgap
Calibration

WideResearch → BrowseComp

BrowseComp is estimated from WideResearch with a offset logistic curve fitted on 11 models measured on both: y = 0.5805 + (1.0732 − 0.5805) / (1 + exp(−22.30·(x − 0.7940))), R² = 0.93, cross-validated error 5.4 pp. It is used for 0 estimates.

Fitted on

ModelWideResearchBrowseComp
Atria Dawn Preview81.9%92.5%
dots3-note Preview78.9%83.3%
GLM-569.8%62.0%
Kimi K2.680.8%83.2%
Kimi K2.572.7%60.6%
Ling 3.0 Flash73.6%72.2%
Ling 3.1 Flash83.3%91.7%
Ornith-1.5-35B-A3B67.8%67.6%
Ornith-1.5-397B80.8%86.6%
Ornith-1.5-9B59.5%56.4%
Qwen3.5 397B74.0%69.0%