Calibration
ExploitGym → BrowseComp
BrowseComp is estimated from ExploitGym with a linear curve fitted on 6 models measured on both: y = 0.2866·x + 0.8067, R² = 0.93, cross-validated error 1.8 pp. It is used for 9 estimates.
| Estimated model | ExploitGym | BrowseComp | Source |
|---|---|---|---|
| Claude Mythos Preview | 17.5% | 85.7% | estimated ± 1.8 pp, medium confidence |
| DeepSeek V4.1 Flash | 15.3% | 85.1% | estimated ± 1.8 pp, medium confidence |
| GLM-5.3 | 15.0% | 85.0% | estimated ± 1.8 pp, medium confidence |
| GPT-6.1 Sol | 35.1% | 90.7% | estimated ± 1.8 pp, medium confidence |
| GPT-6 Luna | 11.6% | 84.0% | estimated ± 1.8 pp, medium confidence |
| GPT-6 Sol | 22.1% | 87.0% | estimated ± 1.8 pp, medium confidence |
| MiMo-V2.6-Flash | 6.0% | 82.4% | estimated ± 1.8 pp, medium confidence |
| MiMo-V2.6-Pro | 17.8% | 85.8% | estimated ± 1.8 pp, medium confidence |
| Muse Spark 1.1 | 0.8% | 80.9% | estimated ± 1.8 pp, low confidence |