CyberGym-E2E S1 (615 tasks)
Agents · 2026-06-03
CyberGym-E2E S1 success rate on the earlier 615-task set. A task passes S1 when the agent's generated proof of concept crashes the unpatched build.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.2 | 30.2 |
| Gemini 3 Pro Preview | 29.6 |
| Opus 4.5 | 24.9 |
| Sonnet 4.5 | 18.1 |