Calibration
SWE-bench Pro → Vals SWE-bench
Vals SWE-bench is estimated from SWE-bench Pro with a Hill curve fitted on 35 models measured on both: y = 0.0000 + (1.0285 − 0.0000)·x^4.13 / (0.43114^4.13 + x^4.13), R² = 0.54, cross-validated error 7.5 pp. It is used for 3 estimates.
| Estimated model | SWE-bench Pro | Vals SWE-bench | Source |
|---|---|---|---|
| Ling 3.1 Flash | 68.7% | 89.8% | estimated ± 7.5 pp, medium confidence |
| Atria Dawn Preview | 59.6% | 81.4% | estimated ± 7.5 pp, medium confidence |
| Laguna S 2.1 | 59.4% | 81.2% | estimated ± 7.5 pp, medium confidence |
Fitted on
| Model | SWE-bench Pro | Vals SWE-bench |
|---|---|---|
| Claude Fable 5 | 80.0% | 95.0% |
| Claude Opus 4.8 | 69.2% | 88.6% |
| Claude Opus 5 | 79.2% | 97.0% |
| Claude Sonnet 5 | 63.2% | 79.6% |
| DeepSeek V4 Flash 0731 | 52.6% | 88.8% |
| DeepSeek V4 Pro 0813 | 55.4% | 96.4% |
| Gemini 3.5 Flash | 55.1% | 78.8% |
| Gemini 3.5 Flash-Lite | 54.2% | 75.0% |
| GLM-5.1 | 58.4% | 76.4% |
| GLM-5.2 | 62.1% | 82.8% |
| GPT-5.3 Codex | 56.8% | 78.0% |
| GPT-5.5 | 58.6% | 82.6% |
| GPT-5.6 Luna | 62.7% | 93.0% |
| GPT-5.6 Sol | 64.6% | 96.2% |
| GPT-5.6 Terra | 63.4% | 95.4% |
| Grok 4.20 | 51.8% | 72.2% |
| Grok 4.5 | 64.7% | 86.6% |
| Inkling | 54.3% | 77.6% |
| Inkling-Small | 55.9% | 82.2% |
| Kimi K2.6 | 58.6% | 76.2% |
| Laguna M.1 | 49.2% | 57.6% |
| Laguna XS.2 | 46.3% | 55.2% |
| Ling 3.0 Flash | 56.6% | 65.2% |
| MiMo-V2.5 | 56.1% | 71.0% |
| MiMo-V2.5-Pro | 57.2% | 74.0% |
| MiniMax M2.7 | 56.2% | 73.8% |
| MiniMax M3 | 59.0% | 75.0% |
| Muse Spark | 52.4% | 74.4% |
| Muse Spark 1.1 | 61.5% | 82.0% |
| Qwen3.6-27B | 53.5% | 70.0% |
| Qwen 3.6 Max (preview) | 57.3% | 72.8% |
| Qwen3.6 Plus | 56.6% | 73.4% |
| Qwen3.7 Max | 60.6% | 68.8% |
| Qwen3.8-27B | 61.7% | 86.0% |
| Qwen3.8 Max | 67.7% | 85.6% |