SWE-bench Verified Extended (Mercor held-out)
Code
Mercor's held-out Extended twin of SWE-bench Verified: 50 Mercor real-world issue-resolution tasks in which an agent edits a codebase, scored by the share of tasks resolved. The public twin is SWE-bench Verified (swe-bench-verified). Mercor does not publish the held-out tasks or say how they were drawn or written, so this is its own benchmark, not a component of the public one.
Top models (higher is better)
| Model | Score |
|---|---|
| DeepSeek V4 Pro 0813 | 83.3 |
| DeepSeek V4.1 Flash | 82.7 |
| Claude Opus 5 | 82.0 |
| Claude Fable 5 | 78.7 |
| GPT-5.6 Terra | 78.7 |
| Kimi K3 | 78.7 |
| Grok 4.6 | 78.0 |
| GPT-5.6 Luna | 74.7 |
| Grok 4.5 | 70.7 |
| GLM-5.3 | 68.0 |
| Sonnet 5 | 67.3 |
| Claude Opus 5.5 | 66.0 |
| GLM-5.3 Flash | 66.0 |
| Opus 4.8 | 63.3 |
| GPT-5.6 Sol Pro | 63.3 |
| Claude Fable 5.1 | 61.3 |
| MiMo-V2.6-Pro | 61.3 |
| Opus 4.7 | 58.0 |
| Claude Sonnet 5.5 | 57.3 |
| Grok 4.7 | 55.3 |
| Gemini 3.8 Flash | 54.0 |
| Muse Spark 1.3 | 54.0 |
| Claude Haiku 5.5 | 53.3 |
| GLM-5.2 | 47.3 |
| GPT-6.1 Sol | 45.3 |
| GPT-5.4 | 40.0 |
| Beam 1.0 501B A23B | 39.3 |
| GPT-5.5 | 39.3 |
| GPT-6 Luna | 39.3 |
| GPT-6 Sol | 39.3 |
| Gemini 3.6 Flash | 33.3 |
| Gemini 3.7 Flash | 30.0 |
| MiniMax M3 | 28.7 |
| Sonnet 4.6 | 27.3 |
| Inkling | 26.7 |
| Kimi K2.7 Code | 26.7 |
| Gemini 3.5 Flash | 22.7 |
| Gemini 3.1 Pro Preview | 21.3 |
| DeepSeek-V3.2 | 10.0 |
| Nemotron 3 Ultra 550B A55B (NVFP4) | 8.7 |