benchgap
Calibration

Toolathlon-Verified → BrowseComp

BrowseComp is estimated from Toolathlon-Verified with a Hill curve fitted on 10 models measured on both: y = 0.0000 + (0.8972 − 0.0000)·x^4.63 / (0.36879^4.63 + x^4.63), R² = 0.79, cross-validated error 6.4 pp. It is used for 1 estimate.

Estimated modelToolathlon-VerifiedBrowseCompSource
Laguna S 2.149.7%71.7%estimated ± 6.4 pp, medium confidence

Fitted on

ModelToolathlon-VerifiedBrowseComp
Claude Opus 580.6%90.8%
DeepSeek V4 Flash 073170.3%73.2%
DeepSeek V4 Pro 081374.1%83.4%
dots3-note Preview55.6%83.3%
Inkling-Small54.4%77.4%
Kimi K373.2%91.2%
Ornith-1.5-35B-A3B48.7%67.6%
Ornith-1.5-397B71.2%86.6%
Ornith-1.5-9B41.2%56.4%
Step 5 Preview74.1%88.7%