CyberGym-E2E S3 (615 tasks)
Agents · 2026-06-03
CyberGym-E2E cumulative S3 success rate on the earlier 615-task set. In addition to passing S1 and S2, the patched project must retain its developer-tested functionality; S3 is the benchmark's discover-and-patch success criterion.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3 Pro Preview | 22.6 |
| GPT-5.2 | 20.7 |
| Opus 4.5 | 19.2 |
| Sonnet 4.5 | 10.6 |