Calibration
JobBench → BrowseComp
BrowseComp is estimated from JobBench with a linear curve fitted on 9 models measured on both: y = 0.6572·x + 0.5401, R² = 0.75, cross-validated error 6.2 pp. It is used for 4 estimates.
| Estimated model | JobBench | BrowseComp | Source |
|---|---|---|---|
| Claude 4.1 Opus | 21.9% | 68.4% | estimated ± 6.2 pp, medium confidence |
| Claude Haiku 4.5 | 16.0% | 64.5% | estimated ± 6.2 pp, medium confidence |
| GPT-5 (high) | 8.5% | 59.6% | estimated ± 6.2 pp, low confidence |
| Qwen3.5 Plus | 18.5% | 66.2% | estimated ± 6.2 pp, medium confidence |