Calibration
SWE-bench Verified → AA-SciCode
AA-SciCode is estimated from SWE-bench Verified with a Michaelis–Menten curve fitted on 28 models measured on both: y = 2.0000·x / (2.51760 + x), R² = 0.77, cross-validated error 3.9 pp. It is used for 14 estimates.
| Estimated model | SWE-bench Verified | AA-SciCode | Source |
|---|---|---|---|
| Claude 3.5 Sonnet | 49.0% | 32.6% | estimated ± 3.9 pp, high confidence |
| Claude 4.1 Opus | 74.5% | 45.7% | estimated ± 3.9 pp, high confidence |
| Claude 4 Sonnet | 72.7% | 44.8% | estimated ± 3.9 pp, high confidence |
| Claude Haiku 4.5 | 73.3% | 45.1% | estimated ± 3.9 pp, high confidence |
| Claude Sonnet 4.5 | 77.2% | 46.9% | estimated ± 3.9 pp, high confidence |
| Ember-1 | 92.2% | 53.6% | estimated ± 3.9 pp, high confidence |
| GPT-4.1 | 54.6% | 35.6% | estimated ± 3.9 pp, high confidence |
| Grok Code Fast 1 | 70.8% | 43.9% | estimated ± 3.9 pp, high confidence |
| MAI-Code-1.1-Flash | 72.6% | 44.8% | estimated ± 3.9 pp, high confidence |
| MiMo-V2-Omni | 74.8% | 45.8% | estimated ± 3.9 pp, high confidence |
| MiMo-V2-Pro | 78.0% | 47.3% | estimated ± 3.9 pp, high confidence |
| o3-mini | 49.3% | 32.8% | estimated ± 3.9 pp, high confidence |
| Qwen3.5-27B | 72.4% | 44.7% | estimated ± 3.9 pp, high confidence |
| Qwen3.5-35B-A3B | 69.2% | 43.1% | estimated ± 3.9 pp, high confidence |