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

#ModelScore
1GPT-6 (xHigh)37.79
2GPT-5.6 Sol (xHigh)35.44
3Claude Opus 5 (Max)32.44
4GPT-6 (Medium)29.33
5Claude Opus 5 (Medium)28.25
6GPT-5.6 Luna (xHigh)26.68
7GPT-5.6 Sol (Medium)25
8GLM-5.3 (High)21.91
9GPT-5.5 (xHigh)21.11
10GLM-5.3 (Max)20.39
11GPT-5.6 Terra (xHigh)19.14
12Claude Sonnet 5 (Max)16.99
13Qwen 3.8 Max (xHigh)16.47
14Grok 4.6 (xHigh)16.16
15GLM-5.3 Flash (High)12.98

Interactive version: theaggregate.ai/benchmark?slug=deepsecbench · How It Works · Data refreshed daily, snapshot 2026-09-19.