Calibration
OSWorld-Verified → BrowseComp
BrowseComp is estimated from OSWorld-Verified with a offset logistic curve fitted on 13 models measured on both: y = 0.6282 + (0.8382 − 0.6282) / (1 + exp(−70.17·(x − 0.6443))), R² = 0.95, cross-validated error 3.1 pp. It is used for 14 estimates.
| Estimated model | OSWorld-Verified | BrowseComp | Source |
|---|---|---|---|
| Claude Opus 4.5 | 66.3% | 79.4% | estimated ± 3.1 pp, high confidence |
| Claude Sonnet 4.5 | 61.4% | 65.1% | estimated ± 3.1 pp, high confidence |
| Gemini 3.5 Flash | 78.4% | 83.8% | estimated ± 3.1 pp, high confidence |
| Gemini 3.5 Flash-Lite | 74.0% | 83.8% | estimated ± 3.1 pp, high confidence |
| Gemini 3.6 Flash | 83.0% | 83.8% | estimated ± 3.1 pp, high confidence |
| GPT-5.3 Codex | 64.7% | 74.3% | estimated ± 3.1 pp, high confidence |
| GPT-5.4 mini | 72.1% | 83.7% | estimated ± 3.1 pp, high confidence |
| GPT-5.4 nano | 39.0% | 62.8% | estimated ± 3.1 pp, medium confidence |
| Holo3-122B-A10B | 78.9% | 83.8% | estimated ± 3.1 pp, high confidence |
| Holo3-35B-A3B | 82.6% | 83.8% | estimated ± 3.1 pp, high confidence |
| Muse Glimmer 30B | 65.9% | 78.3% | estimated ± 3.1 pp, high confidence |
| Qwen3.8-27B | 84.3% | 83.8% | estimated ± 3.1 pp, high confidence |
| UI-Mate-27B | 77.0% | 83.8% | estimated ± 3.1 pp, high confidence |
| UI-Mate-9B | 66.2% | 79.1% | estimated ± 3.1 pp, high confidence |