CWE-Bench-AA: leaderboard
Artificial Analysis's run of 120 held-out CWE-Bench tasks private to Collinear AI and AA, covering the ten OWASP Top 10 (2025) categories: each agent audits and patches a real open-source repository in one 2-hour attempt on AA's Stirrup harness, and a deterministic verifier requires the exploit blocked with legitimate behaviour preserved. Cells for models that declined more than 5% of tasks are withheld. Pass@1 (%).
Metric: Pass@1 (%). Source: artificialanalysis.ai. Saturation forecast: Around October 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4.1 Flash (1017) (Max) | 69.17 | #27 (DeepSeek V4.1 Flash (1017)) |
| 2 | DeepSeek V4.1 Flash (Max) | 68.33 | #37 (DeepSeek V4.1 Flash) |
| 3 | Grok 4.7 (xHigh) | 68.33 | #26 (Grok 4.7) |
| 4 | MiMo-V2.6-Pro | 63.33 | #48 |
| 5 | Kimi K3.1 Pro | 60.83 | #17 |
| 6 | Muse Spark 1.3 (xHigh) | 60.83 | #31 (Muse Spark 1.3) |
| 7 | mistral-large-4-oss (High) | 58.33 | #60 (mistral-large-4-oss) |
| 8 | GPT-6 Luna (Max) | 56.67 | #61 (GPT-6 Luna) |
| 9 | GLM-5.3 Flash | 55.83 | #77 |
| 10 | GLM-5.3 (Max) | 55.83 | #44 (GLM-5.3) |
| 11 | Kimi K3 (Max) | 53.33 | #35 (Kimi K3) |
| 12 | Step 5 Preview | 52.5 | #53 |
| 13 | Mistral Large 4 | 50.83 | #106 |
| 14 | Beam 1.0 | 49.17 | #102 |
| 15 | MiniMax-M3 | 41.67 | #151 |
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-aa · How It Works · Data refreshed daily, snapshot 2026-11-05.