Calibration
SWE-bench Verified → DeepSWE
DeepSWE is estimated from SWE-bench Verified with a Hill curve fitted on 9 models measured on both: y = 0.0000 + (1.1142 − 0.0000)·x^6.00 / (0.85573^6.00 + x^6.00), R² = 0.54, cross-validated error 12.7 pp. It is used for 33 estimates.
| Estimated model | SWE-bench Verified | DeepSWE | Source |
|---|---|---|---|
| Claude Mythos 5 | 95.5% | 73.4% | estimated ± 12.7 pp, low confidence |
| Claude Opus 4.6 | 80.8% | 46.3% | estimated ± 12.7 pp, low confidence |
| GPT-5.2 | 80.0% | 44.6% | estimated ± 12.7 pp, low confidence |
| BTL-4 | 78.4% | 41.4% | estimated ± 12.7 pp, low confidence |
| MiMo-V2-Pro | 78.0% | 40.6% | estimated ± 12.7 pp, low confidence |
| Claude Sonnet 4.5 | 77.2% | 39.0% | estimated ± 12.7 pp, low confidence |
| Kimi K2.5 (Reasoning) | 76.8% | 38.2% | estimated ± 12.7 pp, low confidence |
| MiMo-V2-Omni | 74.8% | 34.4% | estimated ± 12.7 pp, low confidence |
| Laguna M.1 | 74.6% | 34.0% | estimated ± 12.7 pp, low confidence |
| Claude 4.1 Opus | 74.5% | 33.8% | estimated ± 12.7 pp, low confidence |
| Step 3.5 Flash | 74.4% | 33.6% | estimated ± 12.7 pp, low confidence |
| Flint | 74.2% | 33.2% | estimated ± 12.7 pp, low confidence |
| MAI-Thinking-1 | 73.5% | 31.9% | estimated ± 12.7 pp, low confidence |
| MiMo-V2-Flash | 73.4% | 31.7% | estimated ± 12.7 pp, low confidence |
| Jais-3-MoVA-36B-A4B | 72.8% | 30.6% | estimated ± 12.7 pp, low confidence |
| Claude 4 Sonnet | 72.7% | 30.4% | estimated ± 12.7 pp, low confidence |
| MAI-Code-1.1-Flash | 72.6% | 30.3% | estimated ± 12.7 pp, low confidence |
| Claude 4 Opus | 72.5% | 30.1% | estimated ± 12.7 pp, low confidence |
| Jais-3-32B | 72.4% | 29.9% | estimated ± 12.7 pp, low confidence |
| Qwen3.5-27B | 72.4% | 29.9% | estimated ± 12.7 pp, low confidence |
| Laguna XS 2.1 | 70.9% | 27.2% | estimated ± 12.7 pp, low confidence |
| Grok Code Fast 1 | 70.8% | 27.1% | estimated ± 12.7 pp, low confidence |
| Solar Open 2 | 70.4% | 26.4% | estimated ± 12.7 pp, low confidence |
| Jais-3-375B-A23B | 70.2% | 26.0% | estimated ± 12.7 pp, low confidence |
| Jais-3-7B | 70.2% | 26.0% | estimated ± 12.7 pp, low confidence |
| Laguna XS.2 | 69.9% | 25.5% | estimated ± 12.7 pp, low confidence |
| Qwen3-Coder 480B-A35B | 69.6% | 25.0% | estimated ± 12.7 pp, low confidence |
| Qwen3.5-35B-A3B | 69.2% | 24.3% | estimated ± 12.7 pp, low confidence |
| LongCat-Flash-Lite-Sparse | 68.2% | 22.7% | estimated ± 12.7 pp, low confidence |
| DeepSeek V3.1 | 66.0% | 19.4% | estimated ± 12.7 pp, low confidence |
| Kimi K2 | 65.8% | 19.1% | estimated ± 12.7 pp, low confidence |
| Spark-X2.6-4B | 62.8% | 15.1% | estimated ± 12.7 pp, low confidence |
| Claude 3.7 Sonnet | 62.3% | 14.4% | estimated ± 12.7 pp, low confidence |
Fitted on
| Model | SWE-bench Verified | DeepSWE |
|---|---|---|
| Claude Opus 5 | 96.0% | 68.8% |
| DeepSeek V4 Flash 0731 | 79.0% | 54.4% |
| DeepSeek V4 Pro 0813 | 80.6% | 62.7% |
| Ember-1 | 92.2% | 75.2% |
| Ornith-1.5-35B-A3B | 79.0% | 22.0% |
| Ornith-1.5-397B | 86.0% | 56.0% |
| Beam 1.0 501B-A23B | 81.6% | 43.3% |
| Beam 1.0 501B-A23B FP8 | 80.2% | 43.0% |
| Beam 1.0 501B-A23B NVFP4 | 80.7% | 42.7% |