Calibration
APEX-Agents-AA → BrowseComp
BrowseComp is estimated from APEX-Agents-AA with a Hill curve fitted on 11 models measured on both: y = 0.0000 + (0.8645 − 0.0000)·x^3.71 / (0.09097^3.71 + x^3.71), R² = 0.94, cross-validated error 3.4 pp. It is used for 13 estimates.
| Estimated model | APEX-Agents-AA | BrowseComp | Source |
|---|---|---|---|
| Apodex 1.1 | 31.2% | 85.6% | estimated ± 3.4 pp, high confidence |
| Apodex 1.1 Mini | 31.2% | 85.6% | estimated ± 3.4 pp, high confidence |
| Claude Opus 4.6 (Adaptive) | 33.0% | 85.7% | estimated ± 3.4 pp, high confidence |
| Gemini 3.1 Pro | 32.0% | 85.6% | estimated ± 3.4 pp, high confidence |
| GLM-5 | 14.5% | 73.4% | estimated ± 3.4 pp, high confidence |
| GLM-5.2 | 33.7% | 85.8% | estimated ± 3.4 pp, high confidence |
| GPT-OSS 120B | 3.1% | 1.6% | estimated ± 3.4 pp, medium confidence |
| GPT-OSS 20B | 0.7% | 0.0% | estimated ± 3.4 pp, medium confidence |
| Grok 4.3 | 17.0% | 78.7% | estimated ± 3.4 pp, high confidence |
| MiMo-V2.5-Pro | 2.4% | 0.6% | estimated ± 3.4 pp, medium confidence |
| MiniMax M2.7 | 10.6% | 55.2% | estimated ± 3.4 pp, medium confidence |
| Nemotron 3 Super 100B | 1.8% | 0.2% | estimated ± 3.4 pp, medium confidence |
| Qwen3.8 Max Preview | 42.4% | 86.2% | estimated ± 3.4 pp, medium confidence |