Calibration
Vals LiveCodeBench → SWE-bench Verified
SWE-bench Verified is estimated from Vals LiveCodeBench with a offset logistic curve fitted on 21 models measured on both: y = 0.7626 + (0.9646 − 0.7626) / (1 + exp(−200.00·(x − 0.8777))), R² = 0.72, cross-validated error 4.4 pp. It is used for 8 estimates.
| Estimated model | Vals LiveCodeBench | SWE-bench Verified | Source |
|---|---|---|---|
| Gemini 3.8 Flash | 89.5% | 95.8% | estimated ± 4.4 pp, high confidence |
| Grok 4.6 | 88.2% | 90.5% | estimated ± 4.4 pp, high confidence |
| Gemini 3.6 Flash | 88.1% | 89.6% | estimated ± 4.4 pp, high confidence |
| Kimi K3 | 87.2% | 81.2% | estimated ± 4.4 pp, high confidence |
| Grok 4.3 | 84.5% | 76.3% | estimated ± 4.4 pp, high confidence |
| Qwen3.5 Flash | 83.3% | 76.3% | estimated ± 4.4 pp, high confidence |
| Kimi K2.7 Code | 82.1% | 76.3% | estimated ± 4.4 pp, high confidence |
| GLM-4.5 | 67.4% | 76.3% | estimated ± 4.4 pp, high confidence |
Fitted on
| Model | Vals LiveCodeBench | SWE-bench Verified |
|---|---|---|
| Claude Fable 5 | 89.8% | 95.0% |
| Claude Haiku 4.5 | 41.2% | 73.3% |
| Claude Opus 4.8 | 87.8% | 88.6% |
| Claude Opus 5 | 89.0% | 96.0% |
| Claude Sonnet 4.6 | 82.1% | 79.6% |
| Claude Sonnet 5 | 82.4% | 85.2% |
| DeepSeek V4 Flash 0731 | 87.3% | 79.0% |
| DeepSeek V4 Pro 0813 | 87.5% | 80.6% |
| GLM-4.7 | 82.2% | 73.8% |
| GPT-5.3 Codex | 87.3% | 85.0% |
| Grok 4.20 | 84.3% | 76.7% |
| Grok Code Fast 1 | 62.0% | 70.8% |
| Inkling | 85.5% | 77.6% |
| Inkling-Small | 85.9% | 80.2% |
| Kimi K2.6 | 86.8% | 80.2% |
| Laguna M.1 | 68.1% | 74.6% |
| Laguna XS.2 | 67.8% | 69.9% |
| MiniMax M3 | 82.2% | 80.5% |
| Nemotron 3 Ultra | 86.0% | 71.9% |
| Qwen3.6 Plus | 86.0% | 78.8% |
| Qwen3.7 Max | 87.1% | 80.4% |