Calibration
SWE-bench Pro → AA Coding Index
AA Coding Index is estimated from SWE-bench Pro with a Hill curve fitted on 38 models measured on both: y = 0.2114 + (0.8676 − 0.2114)·x^6.00 / (0.53173^6.00 + x^6.00), R² = 0.72, cross-validated error 7.5 pp. It is used for 2 estimates.
| Estimated model | SWE-bench Pro | AA Coding Index | Source |
|---|---|---|---|
| Sage-A1.5 | 61.2% | 67.0% | estimated ± 7.5 pp, medium confidence |
| Atria Dawn Preview | 59.6% | 64.8% | estimated ± 7.5 pp, medium confidence |
Fitted on
| Model | SWE-bench Pro | AA Coding Index |
|---|---|---|
| Claude Fable 5 | 80.0% | 76.5% |
| Claude Fable 5.1 | 81.2% | 81.6% |
| Claude Opus 4.7 (Adaptive) | 64.3% | 73.6% |
| Claude Opus 4.8 | 69.2% | 74.3% |
| Claude Opus 5 | 79.2% | 78.0% |
| Claude Sonnet 5 | 63.2% | 71.6% |
| DeepSeek V4 Flash 0731 | 52.6% | 69.1% |
| DeepSeek V4 Pro 0813 | 55.4% | 68.8% |
| Gemini 3.5 Flash | 55.1% | 70.1% |
| Gemini 3.5 Flash-Lite | 54.2% | 49.3% |
| GLM-5.1 | 58.4% | 55.8% |
| GLM-5.2 | 62.1% | 68.8% |
| GPT-5.4 | 57.7% | 71.1% |
| GPT-5.5 | 58.6% | 74.9% |
| GPT-5.6 Luna | 62.7% | 71.5% |
| GPT-5.6 Sol | 64.6% | 77.4% |
| GPT-5.6 Terra | 63.4% | 76.7% |
| Granite 4.2 8B | 19.1% | 22.4% |
| Grok 4.5 | 64.7% | 72.5% |
| Inkling | 54.3% | 52.1% |
| Inkling-Small | 55.9% | 52.9% |
| Kimi K2.6 | 58.6% | 61.8% |
| Kimi K2.5 | 50.7% | 46.8% |
| Ling 3.0 Flash | 56.6% | 50.6% |
| MiMo-V2.5-Pro | 57.2% | 60.2% |
| MiniMax M2.7 | 56.2% | 52.6% |
| MiniMax M3 | 59.0% | 58.6% |
| Muse Glimmer 30B | 51.2% | 49.0% |
| Muse Spark | 52.4% | 58.6% |
| Muse Spark 1.1 | 61.5% | 71.3% |
| Qwen3.6-27B | 53.5% | 53.7% |
| Qwen3.6-35B-A3B | 49.5% | 41.9% |
| Qwen3.6 Plus | 56.6% | 54.5% |
| Qwen3.7 Max | 60.6% | 66.0% |
| Qwen3.7 Plus | 57.6% | 55.9% |
| Qwen3.8-27B | 61.7% | 68.1% |
| Qwen3.8-Flash-Next | 62.5% | 73.1% |
| Step 3.7 Flash | 56.3% | 39.6% |