APEX-SWE
Code · 2026-01-13
Pass@1 on Mercor's 200-task APEX-SWE software engineering benchmark. The paper and leaderboard report means across repeated evaluations; observations declare that sampling protocol separately from the canonical distinct-task count. Published model-level nSamples values do not establish per-harness repeat counts or verified success denominators.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 67.6 |
| Claude Sonnet 5.5 | 66.4 |
| Claude Opus 5 | 63.7 |
| Claude Fable 5.1 | 63.6 |
| Claude Fable 5 | 58.9 |
| Grok 4.6 | 57.7 |
| Claude Haiku 5.5 | 55.1 |
| Muse Wonder | 54.6 |
| MiMo-V2.6-Pro | 54.2 |
| Grok 4.5 | 53.6 |
| Grok 4.7 | 53.4 |
| DeepSeek V4.1 Flash | 51.9 |
| Gemini 4 Argon | 49.2 |
| DeepSeek V4.1 Flash 1017 | 49.1 |
| Muse Spark Open | 49.1 |
| GPT-6 Astra | 48.6 |
| GPT-5.6 Sol | 48.0 |
| Kimi K3 | 48.0 |
| Opus 4.7 | 47.6 |
| Opus 4.8 | 47.3 |
| GPT-6.1 Sol | 46.9 |
| Sonnet 5 | 46.4 |
| GLM-5.3 Flash | 46.1 |
| DeepSeek V4 Pro 0813 | 45.9 |
| GLM-5.3 | 45.6 |
| GPT-6 Sol | 45.0 |
| Mercury 3 | 43.1 |
| Gemini 3.7 Flash | 42.0 |
| Opus 4.6 | 41.8 |
| Beam 1.0 501B A23B | 41.6 |
| GPT-5.3-Codex | 41.5 |
| GPT-5.5 | 40.8 |
| Muse Spark 1.2 | 40.5 |
| Kimi K2.7 Code | 40.3 |
| Gemini 3.6 Flash | 39.7 |
| Muse Spark 1.3 | 39.1 |
| Muse Spark 1.1 | 38.9 |
| GPT-6 Luna | 38.8 |
| GLM-5.2 | 37.4 |
| Mistral Large 4 Preview | 37.0 |