Calibration
Toolathlon → BrowseComp
BrowseComp is estimated from Toolathlon with a linear curve fitted on 13 models measured on both: y = 0.9612·x + 0.3158, R² = 0.85, cross-validated error 4.3 pp. It is used for 2 estimates.
| Estimated model | Toolathlon | BrowseComp | Source |
|---|---|---|---|
| Qwen3.6 Plus | 39.8% | 69.8% | estimated ± 4.3 pp, high confidence |
| Qwen3.6-35B-A3B | 26.9% | 57.4% | estimated ± 4.3 pp, medium confidence |
Fitted on
| Model | Toolathlon | BrowseComp |
|---|---|---|
| Claude Opus 4.8 | 59.9% | 84.3% |
| DeepSeek V4 Flash 0731 | 47.8% | 73.2% |
| DeepSeek V4 Pro 0813 | 51.8% | 83.4% |
| GLM-5 | 38.0% | 62.0% |
| GPT-5.4 | 54.6% | 82.7% |
| GPT-5.5 | 55.6% | 84.4% |
| GPT-5.6 Luna | 53.4% | 83.3% |
| GPT-5.6 Sol | 58.0% | 92.2% |
| GPT-5.6 Terra | 53.1% | 87.5% |
| Kimi K2.6 | 50.0% | 83.2% |
| Kimi K2.5 | 27.8% | 60.6% |
| Qwen3.5 397B | 36.3% | 69.0% |
| Step 3.7 Flash | 49.5% | 75.8% |