Calibration
BrowseComp → DeepSearchQA
DeepSearchQA is estimated from BrowseComp with a Hill curve fitted on 11 models measured on both: y = 0.7463 + (1.2000 − 0.7463)·x^6.00 / (0.97379^6.00 + x^6.00), R² = 0.31, cross-validated error 8.1 pp. It is used for 17 estimates.
| Estimated model | BrowseComp | DeepSearchQA | Source |
|---|---|---|---|
| Agents-A1 | 75.5% | 82.7% | estimated ± 8.1 pp, low confidence |
| Agents-A1-4B | 66.8% | 78.9% | estimated ± 8.1 pp, low confidence |
| Claude Mythos 5 | 88.0% | 90.6% | estimated ± 8.1 pp, low confidence |
| GLM-4.7 | 52.0% | 75.7% | estimated ± 8.1 pp, low confidence |
| GPT-5.2 | 65.8% | 78.6% | estimated ± 8.1 pp, low confidence |
| GPT-5.4 Pro | 89.3% | 91.6% | estimated ± 8.1 pp, low confidence |
| GPT-5.5 Pro | 90.1% | 92.1% | estimated ± 8.1 pp, low confidence |
| Kimi K2.5 (Reasoning) | 60.6% | 77.1% | estimated ± 8.1 pp, low confidence |
| LongCat-Flash-Lite-Sparse | 48.6% | 75.3% | estimated ± 8.1 pp, low confidence |
| Ornith-1.5-35B-A3B | 67.6% | 79.2% | estimated ± 8.1 pp, low confidence |
| Ornith-1.5-397B | 86.6% | 89.6% | estimated ± 8.1 pp, low confidence |
| Ornith-1.5-9B | 56.4% | 76.3% | estimated ± 8.1 pp, low confidence |
| Qwen3.5-27B | 61.0% | 77.2% | estimated ± 8.1 pp, low confidence |
| Qwen3.5-35B-A3B | 61.0% | 77.2% | estimated ± 8.1 pp, low confidence |
| Qwen3.5 397B | 62.0% | 77.5% | estimated ± 8.1 pp, low confidence |
| Solar Pro 4 | 49.2% | 75.4% | estimated ± 8.1 pp, low confidence |
| Step 5 Preview | 88.7% | 91.1% | estimated ± 8.1 pp, low confidence |