Calibration
SWE-bench Pro → LiveCodeBench v6
LiveCodeBench v6 is estimated from SWE-bench Pro with a Michaelis–Menten + offset curve fitted on 22 models measured on both: y = 0.6337 + 2.0000·x / (3.98152 + x), R² = 0.85, cross-validated error 2.7 pp. It is used for 24 estimates.
| Estimated model | SWE-bench Pro | LiveCodeBench v6 | Source |
|---|---|---|---|
| Claude Opus 5.5 | 89.9% | 100.0% | estimated ± 2.7 pp, medium confidence |
| Claude Sonnet 5.5 | 81.3% | 97.3% | estimated ± 2.7 pp, medium confidence |
| Claude Mythos 5 | 80.3% | 96.9% | estimated ± 2.7 pp, medium confidence |
| Ling 3.1 Flash | 68.7% | 92.8% | estimated ± 2.7 pp, high confidence |
| Beam 1.0 501B-A23B | 66.1% | 91.8% | estimated ± 2.7 pp, high confidence |
| Beam 1.0 501B-A23B FP8 | 65.9% | 91.8% | estimated ± 2.7 pp, high confidence |
| Beam 1.0 501B-A23B NVFP4 | 65.6% | 91.7% | estimated ± 2.7 pp, high confidence |
| Claude Opus 4.7 (Adaptive) | 64.3% | 91.2% | estimated ± 2.7 pp, high confidence |
| GPT-5.6 Luna | 62.7% | 90.6% | estimated ± 2.7 pp, high confidence |
| Sage-A1.5 | 61.2% | 90.0% | estimated ± 2.7 pp, high confidence |
| Atria Dawn Preview | 59.6% | 89.4% | estimated ± 2.7 pp, high confidence |
| Laguna S 2.1 | 59.4% | 89.3% | estimated ± 2.7 pp, high confidence |
| GPT-5.4 | 57.7% | 88.7% | estimated ± 2.7 pp, high confidence |
| Step 3.7 Flash | 56.3% | 88.1% | estimated ± 2.7 pp, high confidence |
| GPT-5.2 | 55.6% | 87.9% | estimated ± 2.7 pp, high confidence |
| Spark-X2.6-4B | 54.3% | 87.4% | estimated ± 2.7 pp, high confidence |
| Claude Opus 4.6 | 53.4% | 87.0% | estimated ± 2.7 pp, high confidence |
| Muse Spark | 52.4% | 86.6% | estimated ± 2.7 pp, high confidence |
| Muse Glimmer 30B | 51.2% | 86.2% | estimated ± 2.7 pp, high confidence |
| Laguna M.1 | 49.2% | 85.4% | estimated ± 2.7 pp, high confidence |
| Laguna XS 2.1 | 47.6% | 84.7% | estimated ± 2.7 pp, high confidence |
| Laguna XS.2 | 46.3% | 84.2% | estimated ± 2.7 pp, high confidence |
| LongCat-Flash-Lite-Sparse | 40.6% | 81.9% | estimated ± 2.7 pp, high confidence |
| LLaDA2.2-flash | 30.1% | 77.4% | estimated ± 2.7 pp, high confidence |
Fitted on
| Model | SWE-bench Pro | LiveCodeBench v6 |
|---|---|---|
| Claude Opus 4.5 | 57.1% | 84.8% |
| dots3-note Preview | 61.0% | 91.5% |
| Granite 4.2 30B | 33.3% | 75.8% |
| Granite 4.2 8B | 19.1% | 73.2% |
| Kimi K2.6 | 58.6% | 89.6% |
| Kimi K2.5 | 50.7% | 85.0% |
| MAI-Thinking-1 | 52.8% | 87.7% |
| Mellum2.1-12B-A2.5B-Thinking | 28.0% | 82.0% |
| MiniCPM5-2B | 14.4% | 69.1% |
| Nemotron 3.5 Super VL 120B A12B | 51.8% | 88.2% |
| Nemotron 3.5 Super VL 120B A12B NVFP4 | 51.3% | 87.4% |
| Qwen3.5 397B | 50.9% | 83.6% |
| Qwen3.6-27B | 53.5% | 83.9% |
| Qwen3.6-35B-A3B | 49.5% | 80.4% |
| Qwen3.6 Plus | 56.6% | 87.1% |
| Qwen3.7 Max | 60.6% | 91.6% |
| Qwen3.7 Plus | 57.6% | 89.6% |
| Qwen3.8-27B | 61.7% | 90.3% |
| Qwen3.8-Flash-Next | 62.5% | 91.9% |
| Qwen3.8-Omni-Flash | 63.3% | 92.6% |
| Sakana Fugu | 59.0% | 92.9% |
| Sakana Fugu-Ultra | 73.7% | 93.2% |