CyberBench v1.1 Patch
Agents
The 56-task patch track of CyberBench v1.1. A patch must compile, remove the original sanitizer failure and match maintainer-fixed reference behavior on hidden holdouts. This revised roster and clean-room grading protocol are distinct from the earlier 59-scored-item track.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Fable 5.1 | 87.5 |
| Gemini 3.8 Flash | 87.5 |
| Claude Opus 5 | 85.7 |
| DeepSeek V4.1 Flash | 85.7 |
| MiMo-V2.6-Flash | 85.7 |
| MiMo-V2.6-Pro | 85.7 |
| GPT-6 Astra | 82.1 |
| Muse Spark 1.3 | 82.1 |
| Grok 4.6 | 80.4 |