Calibration
Vals SWE-bench → SWE-bench Pro
SWE-bench Pro is estimated from Vals SWE-bench with a Michaelis–Menten + offset curve fitted on 35 models measured on both: y = 0.0982 + 2.0000·x / (2.41398 + x), R² = 0.49, cross-validated error 5.4 pp. It is used for 0 estimates.
Fitted on
| Model | Vals SWE-bench | SWE-bench Pro |
|---|---|---|
| Claude Fable 5 | 95.0% | 80.0% |
| Claude Opus 4.8 | 88.6% | 69.2% |
| Claude Opus 5 | 97.0% | 79.2% |
| Claude Sonnet 5 | 79.6% | 63.2% |
| DeepSeek V4 Flash 0731 | 88.8% | 52.6% |
| DeepSeek V4 Pro 0813 | 96.4% | 55.4% |
| Gemini 3.5 Flash | 78.8% | 55.1% |
| Gemini 3.5 Flash-Lite | 75.0% | 54.2% |
| GLM-5.1 | 76.4% | 58.4% |
| GLM-5.2 | 82.8% | 62.1% |
| GPT-5.3 Codex | 78.0% | 56.8% |
| GPT-5.5 | 82.6% | 58.6% |
| GPT-5.6 Luna | 93.0% | 62.7% |
| GPT-5.6 Sol | 96.2% | 64.6% |
| GPT-5.6 Terra | 95.4% | 63.4% |
| Grok 4.20 | 72.2% | 51.8% |
| Grok 4.5 | 86.6% | 64.7% |
| Inkling | 77.6% | 54.3% |
| Inkling-Small | 82.2% | 55.9% |
| Kimi K2.6 | 76.2% | 58.6% |
| Laguna M.1 | 57.6% | 49.2% |
| Laguna XS.2 | 55.2% | 46.3% |
| Ling 3.0 Flash | 65.2% | 56.6% |
| MiMo-V2.5 | 71.0% | 56.1% |
| MiMo-V2.5-Pro | 74.0% | 57.2% |
| MiniMax M2.7 | 73.8% | 56.2% |
| MiniMax M3 | 75.0% | 59.0% |
| Muse Spark | 74.4% | 52.4% |
| Muse Spark 1.1 | 82.0% | 61.5% |
| Qwen3.6-27B | 70.0% | 53.5% |
| Qwen 3.6 Max (preview) | 72.8% | 57.3% |
| Qwen3.6 Plus | 73.4% | 56.6% |
| Qwen3.7 Max | 68.8% | 60.6% |
| Qwen3.8-27B | 86.0% | 61.7% |
| Qwen3.8 Max | 85.6% | 67.7% |