CWE-bench v1 - Judge Panel Pass@4: leaderboard
Collinear CWE-bench archived v1, Judge Panel Pass@4, reported as a percentage. 120 private audit-and-patch tasks; four rollouts per task at high reasoning, one hour each; timeouts graded on code at timeout. Native model-specific harnesses. Programmatic success requires exploit blocked and existing tests passing. Separate majority panel of three rubric-based agentic judges checks correctness, happy-path, generalization, integrity and soundness; any failed axis rejects a fix. Pass@1 and Pass@4 remain separate. This is archived Collinear v1, not the Artificial Analysis Stirrup two-hour implementation or CWE-bench v2. Source: https://cwe-bench.com/v1
Metric: Judge Panel Pass@4 (%). Source: cwe-bench.com. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 4 Carbon | 83 | #2 |
| 2 | Grok 4.7 | 82 | #43 |
| 3 | Claude Opus 5.5 | 81 | #5 |
| 4 | GPT-6 Astra | 77 | #6 |
| 5 | Claude Fable 5.1 | 77 | #14 |
| 6 | DeepSeek V4.1 Flash | 75 | #58 |
| 7 | Gemini 4 Argon | 73 | #11 |
| 8 | Grok 4.6 | 73 | #44 |
| 9 | Hy4 preview | 73 | #66 |
| 10 | GPT-6 Sol | 72 | #24 |
| 11 | Muse Spark 1.3 | 68 | #42 |
No result here: #1 Claude Fable 5.2, #10 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cwe-bench-v1-judge-panel-pass-4 · How It Works · Data refreshed daily, snapshot 2026-11-12.