CyberGym-E2E S4 (615 tasks)
Agents · 2026-06-03
CyberGym-E2E cumulative S4 success rate on the earlier 615-task set. A task must pass S1 through S3 and the patch must also fix the specific held-out ground-truth vulnerability.
Top models (higher is better)
| Model | Score |
|---|---|
| Opus 4.5 | 7.6 |
| GPT-5.2 | 6.5 |
| Gemini 3 Pro Preview | 5.0 |
| Sonnet 4.5 | 3.4 |