SWE Atlas Codebase QnA Extended (Mercor held-out)
Code
Mercor's held-out Extended twin of SWE Atlas Codebase QnA: 50 Mercor codebase question-answering tasks for coding agents, scored by pass@1. The public twin is SWE Atlas - Codebase QnA (swe-atlas-codebase-qna-scale-ai-2). 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 |
|---|---|
| Claude Sonnet 5.5 | 44.7 |
| Claude Opus 5.5 | 43.3 |
| Claude Opus 5 | 40.0 |
| Claude Fable 5.1 | 38.0 |
| DeepSeek V4.1 Flash | 36.7 |
| Sonnet 5 | 32.7 |
| GPT-6.1 Sol | 31.3 |
| GPT-6 Sol | 31.3 |
| GPT-5.6 Sol Pro | 24.7 |
| Opus 4.8 | 24.0 |
| Kimi K3 | 24.0 |
| GPT-5.6 Luna | 19.3 |
| Grok 4.6 | 19.3 |
| GPT-5.5 | 17.3 |
| GLM-5.2 | 16.0 |
| GPT-5.6 Terra | 16.0 |
| GPT-6 Luna | 15.3 |
| Grok 4.7 | 15.3 |
| Grok 4.5 | 12.0 |
| Opus 4.7 | 11.3 |
| GPT-5.4 | 10.7 |
| MiniMax M3 | 10.0 |
| Sonnet 4.6 | 6.7 |
| Gemini 3.6 Flash | 4.7 |
| Kimi K2.7 Code | 4.7 |
| Gemini 3.1 Pro Preview | 2.7 |
| Gemini 3.5 Flash | 2.7 |
| Inkling | 1.3 |
| MiniMax M2.7 | 0.7 |
| Claude Fable 5 | 0.0 |
| DeepSeek-V3.2 | 0.0 |
| gpt-oss-120b | 0.0 |
| Kimi K2 Thinking | 0.0 |