CyberGym-E2E: leaderboard
CyberGym-E2E (2026) evaluates coding agents across the vulnerability lifecycle on real OSS-Fuzz vulnerabilities in C and C++ projects, with a 10-dollar cost cap and a 90-minute limit per task. The headline is the end-to-end success rate (%) on the 920-task benchmark: from the codebase alone the agent must write an input that crashes the vulnerable build, patch the crash and keep the project tests passing; higher is better.
Metric: End-to-end success rate (%, stage S3) on the 920 OSS-Fuzz vulnerabilities (139 C/C++ projects) of the expanded benchmark: given only the codebase and build environment, the agent must write a PoC input that crashes the vulnerable build under a sanitizer, a patch that removes that crash, and keep the project functionality tests passing; with a $10 cost cap and a 90-minute limit per task; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 65.9 | #123 |
| 2 | Gemini 3.1 Pro (Preview) | 43.8 | #94 |
| 3 | Claude Opus 4.6 | 37.9 | #102 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cybergym-e2e · How It Works · Data refreshed daily, snapshot 2026-11-13.