benchgap
Calibration

BrowseComp → OSWorld-Verified

OSWorld-Verified is estimated from BrowseComp with a Hill curve fitted on 13 models measured on both: y = 0.4576 + (1.2000 − 0.4576)·x^6.00 / (0.88159^6.00 + x^6.00), R² = 0.85, cross-validated error 5.4 pp. It is used for 21 estimates.

Estimated modelBrowseCompOSWorld-VerifiedSource
Agents-A175.5%66.8%estimated ± 5.4 pp, medium confidence
Agents-A1-4B66.8%57.6%estimated ± 5.4 pp, medium confidence
dots3-note Preview83.3%76.6%estimated ± 5.4 pp, medium confidence
GLM-4.752.0%48.8%estimated ± 5.4 pp, low confidence
GPT-5.4 Pro89.3%84.3%estimated ± 5.4 pp, low confidence
GPT-5.5 Pro90.1%85.3%estimated ± 5.4 pp, low confidence
Inkling77.1%68.7%estimated ± 5.4 pp, medium confidence
Inkling-Small77.4%69.1%estimated ± 5.4 pp, medium confidence
Kimi K2.560.6%52.8%estimated ± 5.4 pp, low confidence
Kimi K2.5 (Reasoning)60.6%52.8%estimated ± 5.4 pp, low confidence
Ling 3.0 Flash72.2%63.0%estimated ± 5.4 pp, medium confidence
LongCat-Flash-Lite-Sparse48.6%47.8%estimated ± 5.4 pp, low confidence
Nemotron 3.5 Lightning 30B A3B NVFP436.8%46.2%estimated ± 5.4 pp, low confidence
Nemotron 3 Ultra44.4%47.0%estimated ± 5.4 pp, low confidence
Ornith-1.5-35B-A3B67.6%58.3%estimated ± 5.4 pp, medium confidence
Ornith-1.5-397B86.6%80.9%estimated ± 5.4 pp, medium confidence
Ornith-1.5-9B56.4%50.5%estimated ± 5.4 pp, low confidence
Qwen3.5 397B62.0%53.8%estimated ± 5.4 pp, medium confidence
Beam77.4%69.1%estimated ± 5.4 pp, medium confidence
Solar Pro 449.2%47.9%estimated ± 5.4 pp, low confidence
Step 3.7 Flash75.8%67.1%estimated ± 5.4 pp, medium confidence