Calibration
Vals SWE-bench → AA-SciCode
AA-SciCode is estimated from Vals SWE-bench with a linear curve fitted on 41 models measured on both: y = 0.3844·x + 0.2041, R² = 0.48, cross-validated error 4.2 pp. It is used for 1 estimate.
| Estimated model | Vals SWE-bench | AA-SciCode | Source |
|---|---|---|---|
| Qwen3.5 Flash | 64.4% | 45.2% | estimated ± 4.2 pp, medium confidence |
Fitted on
| Model | Vals SWE-bench | AA-SciCode |
|---|---|---|
| Claude Fable 5 | 95.0% | 61.0% |
| Claude Opus 4.8 | 88.6% | 54.4% |
| Claude Opus 5 | 97.0% | 56.4% |
| Claude Sonnet 5 | 79.6% | 54.3% |
| DeepSeek V4 Flash 0731 | 88.8% | 50.3% |
| DeepSeek V4 Pro 0813 | 96.4% | 51.0% |
| Gemini 2.5 Pro | 54.4% | 46.3% |
| Gemini 3.1 Pro | 78.8% | 58.7% |
| Gemini 3.5 Flash | 78.8% | 53.9% |
| Gemini 3.5 Flash-Lite | 75.0% | 41.3% |
| Gemini 3.6 Flash | 79.6% | 53.4% |
| Gemini 3.7 Flash | 80.8% | 57.2% |
| Gemini 3.8 Flash | 80.0% | 56.6% |
| GLM-5.1 | 76.4% | 44.8% |
| GLM-5.2 | 82.8% | 51.2% |
| GLM-5.3 | 95.4% | 59.0% |
| GPT-5.4 mini | 73.0% | 52.1% |
| GPT-5.4 nano | 69.8% | 47.2% |
| GPT-5.5 | 82.6% | 55.8% |
| GPT-5.6 Luna | 93.0% | 53.6% |
| GPT-5.6 Sol | 96.2% | 57.1% |
| GPT-5.6 Terra | 95.4% | 55.0% |
| Grok 4.3 | 71.4% | 48.3% |
| Grok 4.5 | 86.6% | 55.0% |
| Grok 4.6 | 95.6% | 56.5% |
| Inkling | 77.6% | 47.0% |
| Inkling-Small | 82.2% | 49.7% |
| Kimi K2.6 | 76.2% | 51.5% |
| Kimi K2.7 Code | 78.2% | 47.8% |
| Kimi K3 | 93.4% | 59.5% |
| Ling 3.0 Flash | 65.2% | 42.0% |
| MiMo-V2.5-Pro | 74.0% | 50.6% |
| MiniMax M2.7 | 73.8% | 50.1% |
| MiniMax M3 | 75.0% | 47.1% |
| Mistral Medium 3.5 128B | 66.4% | 40.2% |
| Muse Spark 1.1 | 82.0% | 58.8% |
| Muse Spark 1.2 | 86.6% | 57.4% |
| Nemotron 3 Ultra | 69.0% | 40.3% |
| Qwen3.6-27B | 70.0% | 42.8% |
| Qwen3.7 Max | 68.8% | 49.5% |
| Qwen3.8-27B | 86.0% | 46.6% |