Calibration
Vals LiveCodeBench → SWE-bench Pro
SWE-bench Pro is estimated from Vals LiveCodeBench with a offset logistic curve fitted on 32 models measured on both: y = 0.5697 + (0.8227 − 0.5697) / (1 + exp(−200.00·(x − 0.8833))), R² = 0.69, cross-validated error 4.8 pp. It is used for 4 estimates.
| Estimated model | Vals LiveCodeBench | SWE-bench Pro | Source |
|---|---|---|---|
| Gemini 3.7 Flash | 88.7% | 74.1% | estimated ± 4.8 pp, high confidence |
| Grok 4.3 | 84.5% | 57.0% | estimated ± 4.8 pp, high confidence |
| Qwen3.5 Flash | 83.3% | 57.0% | estimated ± 4.8 pp, high confidence |
| GLM-4.5 | 67.4% | 57.0% | estimated ± 4.8 pp, medium confidence |
Fitted on
| Model | Vals LiveCodeBench | SWE-bench Pro |
|---|---|---|
| Claude Fable 5 | 89.8% | 80.0% |
| Claude Fable 5.1 | 90.5% | 81.2% |
| Claude Opus 4.8 | 87.8% | 69.2% |
| Claude Opus 5 | 89.0% | 79.2% |
| Claude Sonnet 5 | 82.4% | 63.2% |
| DeepSeek V4 Flash 0731 | 87.3% | 52.6% |
| DeepSeek V4 Pro 0813 | 87.5% | 55.4% |
| Gemini 3.5 Flash | 87.6% | 55.1% |
| Gemini 3.5 Flash-Lite | 79.0% | 54.2% |
| GLM-5.1 | 81.4% | 58.4% |
| GLM-5.2 | 69.5% | 62.1% |
| GPT-5.3 Codex | 87.3% | 56.8% |
| GPT-5.5 | 85.3% | 58.6% |
| GPT-5.6 Sol | 82.6% | 64.6% |
| GPT-5.6 Terra | 85.9% | 63.4% |
| Grok 4.20 | 84.3% | 51.8% |
| Grok 4.5 | 87.4% | 64.7% |
| Inkling | 85.5% | 54.3% |
| Inkling-Small | 85.9% | 55.9% |
| Kimi K2.6 | 86.8% | 58.6% |
| Laguna M.1 | 68.1% | 49.2% |
| Laguna XS.2 | 67.8% | 46.3% |
| Ling 3.0 Flash | 84.0% | 56.6% |
| MiMo-V2.5 | 81.5% | 56.1% |
| MiMo-V2.5-Pro | 81.4% | 57.2% |
| MiniMax M2.7 | 79.9% | 56.2% |
| MiniMax M3 | 82.2% | 59.0% |
| Muse Spark 1.1 | 85.9% | 61.5% |
| Qwen3.6 Plus | 86.0% | 56.6% |
| Qwen3.7 Max | 87.1% | 60.6% |
| Qwen3.8-27B | 84.0% | 61.7% |
| Qwen3.8 Max | 87.9% | 67.7% |