benchgap
Calibration

BrowseComp → CyberGym

CyberGym is estimated from BrowseComp with a inverse Michaelis–Menten curve fitted on 15 models measured on both: y = 1.07522·(x − 0.0000) / (0.0000 + 2.0000 − x), R² = 0.60, cross-validated error 7.3 pp. It is used for 19 estimates.

Estimated modelBrowseCompCyberGymSource
GPT-5.5 Pro90.1%88.2%estimated ± 7.3 pp, medium confidence
GPT-5.4 Pro89.3%86.7%estimated ± 7.3 pp, medium confidence
dots3-note Preview83.3%76.7%estimated ± 7.3 pp, medium confidence
Sage-A1.581.7%74.3%estimated ± 7.3 pp, medium confidence
Ling 3.0 Flash Sante80.2%72.0%estimated ± 7.3 pp, medium confidence
Beam 1.0 501B-A23B FP876.7%66.9%estimated ± 7.3 pp, medium confidence
Beam 1.0 501B-A23B NVFP475.6%65.3%estimated ± 7.3 pp, medium confidence
Agents-A175.5%65.2%estimated ± 7.3 pp, medium confidence
Nemotron 3.5 Bolt 74B-A9B BF1670.0%57.9%estimated ± 7.3 pp, medium confidence
Qwen3.5 397B69.0%56.6%estimated ± 7.3 pp, medium confidence
Ornith-1.5-35B-A3B67.6%54.9%estimated ± 7.3 pp, medium confidence
Agents-A1-4B66.8%53.9%estimated ± 7.3 pp, medium confidence
GPT-5.265.8%52.7%estimated ± 7.3 pp, medium confidence
Qwen3.5-35B-A3B61.0%47.2%estimated ± 7.3 pp, low confidence
Ornith-1.5-9B56.4%42.2%estimated ± 7.3 pp, low confidence
Spark-X2.6-4B53.1%38.9%estimated ± 7.3 pp, low confidence
Sarvam 105B49.5%35.4%estimated ± 7.3 pp, low confidence
LongCat-Flash-Lite-Sparse48.6%34.5%estimated ± 7.3 pp, low confidence
Sarvam 30B35.5%23.2%estimated ± 7.3 pp, low confidence

Fitted on

ModelBrowseCompCyberGym
Atria Dawn Preview92.5%86.5%
Claude Mythos 588.0%83.8%
Claude Opus 4.683.7%66.6%
Claude Opus 4.779.3%73.1%
DeepSeek V4 Flash 073173.2%76.7%
DeepSeek V4 Pro83.4%83.3%
GLM-562.0%43.2%
GLM-5.168.0%68.7%
GPT-5.482.7%79.0%
GPT-5.584.4%81.8%
GPT-5.6 Luna83.3%77.9%
GPT-5.6 Sol92.2%84.5%
GPT-5.6 Terra87.5%81.8%
Ling 3.1 Flash91.7%87.9%
Step 5 Preview88.7%84.7%