Calibration

BrowseComp → JobBench

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

JobBench is estimated from BrowseComp with a inverse Michaelis–Menten curve fitted on 9 models measured on both: y = 1.48037·(x − 0.3823) / (0.3823 + 2.0000 − x), R² = 0.74, cross-validated error 10.1 pp. It is used for 4 estimates.

Estimated model	BrowseComp	JobBench	Source
Agents-A1	75.5%	33.9%	estimated ± 10.1 pp, low confidence
Agents-A1-4B	66.8%	24.7%	estimated ± 10.1 pp, low confidence
GPT-5.4 Pro	89.3%	50.8%	estimated ± 10.1 pp, low confidence
GPT-5.5 Pro	90.1%	51.8%	estimated ± 10.1 pp, low confidence