CWE-Bench (AA)
Finding and fixing the vulnerability classes CWE catalogues, in real repositories. Run by Artificial Analysis.
As of 18 Oct 2026, DeepSeek V4.1 Flash leads CWE-Bench (AA) on BenchLeader with 68.3%, ahead of Grok 4.7 at 68.3%, across 18 model configurations with a published result.
- Published by
- Artificial Analysis
- Category
- Agents & tools
- Index weight
- 0.5
- Models
- 18
- Data as of
- 18 Oct 2026
Source: Artificial Analysis (artificialanalysis.ai). Data taken from the public leaderboard.
- 1DeepSeek V4.1 Flash (max)68.3%
- 2Grok 4.7 (xhigh)68.3%
- 3GPT-6 Sol (max)64.2%
- 4GPT-6 Astra (max)63.3%
- 5MiMo-V2.6-Pro63.3%
- 6Claude Opus 5.5 (max)58.3%
- 7GPT-6 Luna (max)56.7%
- 8GLM 5.3 (max)55.8%
- 9GLM 5.3 Flash55.8%
- 10Kimi K3 (max)53.3%
- 11Gemini 3.8 Flash (high)52.5%
- 12Step 5 Preview52.5%
- 13Claude Fable 5.1 (max)50.8%
- 14Mistral Large 450.8%
- 15Beam 1.049.2%
18 of 18
| # | ||||
|---|---|---|---|---|
| 1 |  DeepSeek V4.1 Flashmaxopen ↗ | 68.3% | 61.5 | 2026-09-10 |
| 2 |  Grok 4.7xhigh | 68.3% | 60.6 | 2026-09-21 |
| 3 |  GPT-6 Solmax | 64.2% | 65.6 | 2026-09-22 |
| 4 | GPT-6 Astramax | 63.3% | 69.7 | 2026-09-03 |
| 5 |  MiMo-V2.6-Proopen ↗ | 63.3% | 65.3 | 2026-09-21 |
| 6 |  Claude Opus 5.5max | 58.3% | 71.6 | 2026-09-22 |
| 7 | GPT-6 Lunamax | 56.7% | 58.6 | 2026-09-22 |
| 8 |  GLM 5.3maxopen ↗ | 55.8% | 63.6 | 2026-08-18 |
| 9 | GLM 5.3 Flashopen ↗ | 55.8% | 62.2 | 2026-08-26 |
| 10 |  Kimi K3maxopen ↗ | 53.3% | 64.8 | 2026-07-16 |
| 11 |  Gemini 3.8 Flashhigh | 52.5% | 63.3 | 2026-09-02 |
| 12 |  Step 5 Previewopen | 52.5% | 62.2 | 2026-09-18 |
| 13 | Claude Fable 5.1max | 50.8% | 69.9 | 2026-09-01 |
| 14 |  Mistral Large 4open | 50.8% | 58.7 | 2026-10-06 |
| 15 | New Beam 1.0open | 49.2% | 56.4 | 2026-10-16 |
| 16 |  MiniMax M3open ↗ | 41.7% | 56.2 | 2026-06-01 |
| 17 |  Nemotron 3 Ultra 550B A55Bthinkingopen ↗ | 27.5% | 55.6 | 2026-06-04 |
| 18 |  Qwen3.8 27Bxhighopen ↗ | 26.7% | 56.3 | 2026-08-14 |
More agents & tools benchmarks · All LLM benchmarks
Cite as: BenchLeader, “CWE-Bench (AA) leaderboard”, https://www.benchleader.com/benchmarks/aa_cwe_bench, data as of 18 Oct 2026.
CWE-Bench (AA): questions
- What does CWE-Bench (AA) measure?
- Finding and fixing the vulnerability classes CWE catalogues, in real repositories. Run by Artificial Analysis. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads CWE-Bench (AA)?
- DeepSeek V4.1 Flash leads CWE-Bench (AA) with 68.3% as of 18 Oct 2026, ahead of Grok 4.7 at 68.3%.
- How many models have CWE-Bench (AA) results?
- 18 model configurations have a CWE-Bench (AA) result on BenchLeader, all taken from Artificial Analysis.
- Who runs CWE-Bench (AA) and how often is it updated?
- CWE-Bench (AA) is published by Artificial Analysis. BenchLeader re-reads the published results once a day and records the date each result was published.
- Does CWE-Bench (AA) count toward the BenchLeader Index?
- Yes. CWE-Bench (AA) contributes to the agents & tools category of the BenchLeader Index, normalised so that 50 is the average of the evaluated models.