CWE-bench v1 Judge Panel Pass@4
Code · 2026-09-25
Share of the 120 CWE-bench v1 tasks passed by the majority of three agentic judges in at least one of four one-hour high-reasoning rollouts.
Top models (higher is better)
| Model | Score |
|---|---|
| Grok 4.7 | 82.0 |
| Claude Opus 5.5 | 81.0 |
| Claude Fable 5.1 | 77.0 |
| GPT-6 Astra | 77.0 |
| DeepSeek V4.1 Flash | 75.0 |
| Gemini 4 Argon | 73.0 |
| Grok 4.6 | 73.0 |
| Hy4 Preview | 73.0 |
| GPT-6 Sol | 72.0 |
| Muse Spark 1.3 | 68.0 |