Calibration
Vals SWE-bench → SWE-bench Verified
SWE-bench Verified is estimated from Vals SWE-bench with a linear curve fitted on 24 models measured on both: y = 0.5039·x + 0.4130, R² = 0.69, cross-validated error 4.5 pp. It is used for 2 estimates.
| Estimated model | Vals SWE-bench | SWE-bench Verified | Source |
|---|---|---|---|
| Muse Spark 1.2 | 86.6% | 84.9% | estimated ± 4.5 pp, high confidence |
| Composer 2.5 | 79.6% | 81.4% | estimated ± 4.5 pp, high confidence |
Fitted on
| Model | Vals SWE-bench | SWE-bench Verified |
|---|---|---|
| Claude Fable 5 | 95.0% | 95.0% |
| Claude Haiku 4.5 | 66.6% | 73.3% |
| Claude Opus 4.8 | 88.6% | 88.6% |
| Claude Opus 5 | 97.0% | 96.0% |
| Claude Sonnet 4.6 | 77.4% | 79.6% |
| Claude Sonnet 5 | 79.6% | 85.2% |
| DeepSeek V4 Flash 0731 | 88.8% | 79.0% |
| DeepSeek V4 Pro 0813 | 96.4% | 80.6% |
| Gemini 2.5 Pro | 54.4% | 63.8% |
| GLM-4.7 | 69.4% | 73.8% |
| GPT-5.3 Codex | 78.0% | 85.0% |
| Grok 4.20 | 72.2% | 76.7% |
| Inkling | 77.6% | 77.6% |
| Inkling-Small | 82.2% | 80.2% |
| Kimi K2.6 | 76.2% | 80.2% |
| Laguna M.1 | 57.6% | 74.6% |
| Laguna XS.2 | 55.2% | 69.9% |
| MiniMax M3 | 75.0% | 80.5% |
| Mistral Medium 3.5 128B | 66.4% | 77.6% |
| Muse Spark | 74.4% | 77.4% |
| Nemotron 3 Ultra | 69.0% | 71.9% |
| Qwen3.6-27B | 70.0% | 77.2% |
| Qwen3.6 Plus | 73.4% | 78.8% |
| Qwen3.7 Max | 68.8% | 80.4% |