CWE-Bench: leaderboard
CWE-Bench evaluates code agents on audit-and-patch tasks, where the input modality is text/code. The pinned metric is Pass@1 (%), measuring whether the model’s first attempt finds and fixes the hidden vulnerability without breaking existing tests. This is a source-attributed result table mirrored by BenchmarkList. Source: BenchmarkList’s attributed result table.
Metric: Pass@1 (%). Source: benchmarklist.com. Saturation forecast: Around October 2028. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (High) | 47.8 | #20 (Claude Fable 5) |
| 2 | GPT-5.6 Sol (High) | 44.2 | #22 (GPT-5.6 Sol) |
| 3 | Gemini 3.7 Flash (High) | 44 | #24 (Gemini 3.7 Flash) |
| 4 | Claude Opus 4.8 (High) | 42 | #39 (Claude Opus 4.8) |
| 5 | Grok 4.6 (High) | 38.2 | #29 (Grok 4.6) |
| 6 | Qwen 3.8 Max (High) | 37.5 | #34 (Qwen 3.8 Max) |
| 7 | Hy4 preview (High) | 33.8 | #45 (Hy4 preview) |
| 8 | GLM-5.3 (High) | 31.1 | #44 (GLM-5.3) |
| 9 | DeepSeek V4 Flash (High) | 30.4 | #155 (DeepSeek V4 Flash) |
| 10 | MiniMax-M3 (High) | 22.7 | #151 (MiniMax-M3) |
| 11 | Kimi K3 (High) | 22.5 | #35 (Kimi K3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cwe-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.