Calibration
APEX-Agents-AA → JobBench
JobBench is estimated from APEX-Agents-AA with a linear curve fitted on 5 models measured on both: y = 1.5412·x + -0.0943, R² = 0.90, cross-validated error 8.2 pp. It is used for 11 estimates.
| Estimated model | APEX-Agents-AA | JobBench | Source |
|---|---|---|---|
| Apodex 1.1 | 31.2% | 38.7% | estimated ± 8.2 pp, low confidence |
| Apodex 1.1 Mini | 31.2% | 38.7% | estimated ± 8.2 pp, low confidence |
| Claude Opus 4.6 (Adaptive) | 33.0% | 41.4% | estimated ± 8.2 pp, low confidence |
| Gemini 3.1 Pro | 32.0% | 39.9% | estimated ± 8.2 pp, low confidence |
| GPT-OSS 120B | 3.1% | 0.0% | estimated ± 8.2 pp, low confidence |
| GPT-OSS 20B | 0.7% | 0.0% | estimated ± 8.2 pp, low confidence |
| Grok 4.3 | 17.0% | 16.8% | estimated ± 8.2 pp, low confidence |
| Kimi K2.5 (Reasoning) | 11.5% | 8.3% | estimated ± 8.2 pp, low confidence |
| MiMo-V2.5-Pro | 2.4% | 0.0% | estimated ± 8.2 pp, low confidence |
| Nemotron 3 Super 100B | 1.8% | 0.0% | estimated ± 8.2 pp, low confidence |
| Qwen3.8 Max Preview | 42.4% | 55.9% | estimated ± 8.2 pp, low confidence |