DeepsecBench: leaderboard
Vercel benchmark for finding vulnerabilities in application code. Models inspect 50 entry-point files against a hidden golden set of 231 human-judged findings; the primary score is recall-weighted F2, with runs repeated three times and reported by median.
Metric: Recall-weighted F2 score (%). Source: vercel.com. Status: years away from saturation. 38 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 (xHigh) | 37.79 |
| 2 | GPT-5.6 Sol (xHigh) | 35.44 |
| 3 | Claude Opus 5 (Max) | 32.44 |
| 4 | GPT-6 (Medium) | 29.33 |
| 5 | Claude Opus 5 (Medium) | 28.25 |
| 6 | GPT-5.6 Luna (xHigh) | 26.68 |
| 7 | GPT-5.6 Sol (Medium) | 25 |
| 8 | GLM-5.3 (High) | 21.91 |
| 9 | GPT-5.5 (xHigh) | 21.11 |
| 10 | GLM-5.3 (Max) | 20.39 |
| 11 | GPT-5.6 Terra (xHigh) | 19.14 |
| 12 | Claude Sonnet 5 (Max) | 16.99 |
| 13 | Qwen 3.8 Max (xHigh) | 16.47 |
| 14 | Grok 4.6 (xHigh) | 16.16 |
| 15 | GLM-5.3 Flash (High) | 12.98 |
Interactive version: theaggregate.ai/benchmark?slug=deepsecbench · How It Works · Data refreshed daily, snapshot 2026-09-19.