Calibration
JobBench → GDPval-AA
GDPval-AA is estimated from JobBench with a linear curve fitted on 13 models measured on both: y = 0.6690·x + 0.1431, R² = 0.83, cross-validated error 6.2 pp. It is used for 11 estimates.
| Estimated model | JobBench | GDPval-AA | Source |
|---|---|---|---|
| Claude 4.1 Opus | 21.9% | 29.0% | estimated ± 6.2 pp, medium confidence |
| Claude 4 Sonnet | 18.4% | 26.6% | estimated ± 6.2 pp, medium confidence |
| Claude Haiku 4.5 | 16.0% | 25.0% | estimated ± 6.2 pp, medium confidence |
| Claude Sonnet 4.5 | 27.7% | 32.8% | estimated ± 6.2 pp, medium confidence |
| Gemini 3 Flash | 11.4% | 21.9% | estimated ± 6.2 pp, medium confidence |
| Gemini 3 Pro | 11.4% | 21.9% | estimated ± 6.2 pp, medium confidence |
| GPT-5.1-Codex | 26.2% | 31.8% | estimated ± 6.2 pp, medium confidence |
| GPT-5.2 | 34.3% | 37.3% | estimated ± 6.2 pp, medium confidence |
| GPT-5.2-Codex | 26.0% | 31.7% | estimated ± 6.2 pp, medium confidence |
| GPT-5.3 Codex | 33.7% | 36.8% | estimated ± 6.2 pp, medium confidence |
| Qwen3.5 Plus | 18.5% | 26.7% | estimated ± 6.2 pp, medium confidence |