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. Saturation forecast: Around September 2028. 54 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Sol (xHigh) | 40.75 | #19 (GPT-6 Sol) |
| 2 | GPT-6.1 Sol (xHigh) | 39.49 | #7 (GPT-6.1 Sol) |
| 3 | GPT-6 Astra (xHigh) | 37.64 | #6 (GPT-6 Astra) |
| 4 | Claude Sonnet 5.5 (Max) | 37.5 | #14 (Claude Sonnet 5.5) |
| 5 | Kimi K3.1 Pro (Max) | 37.45 | #17 (Kimi K3.1 Pro) |
| 6 | GPT-5.6 Sol (xHigh) | 35.3 | #22 (GPT-5.6 Sol) |
| 7 | Claude Opus 5 (Max) | 32.31 | #16 (Claude Opus 5) |
| 8 | GPT-6.1 Sol (Medium) | 30.69 | #7 (GPT-6.1 Sol) |
| 9 | DeepSeek V4.1 Flash (1017) (Max) | 29.97 | #27 (DeepSeek V4.1 Flash (1017)) |
| 10 | GPT-6 Astra (Medium) | 29.21 | #6 (GPT-6 Astra) |
| 11 | GPT-6 Sol (Medium) | 29.19 | #19 (GPT-6 Sol) |
| 12 | Kimi K3.1 Pro (High) | 28.71 | #17 (Kimi K3.1 Pro) |
| 13 | Claude Opus 5 (Medium) | 28.14 | #16 (Claude Opus 5) |
| 14 | Claude Opus 5.5 (Max) | 26.57 | #3 (Claude Opus 5.5) |
| 15 | GPT-5.6 Luna (xHigh) | 26.57 | #79 (GPT-5.6 Luna) |
No result here: #9 Gemini 4 Argon, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=deepsecbench · How It Works · Data refreshed daily, snapshot 2026-11-05.