dfbench Validation Precision
Cybersecurity · 2026-08-19
Precision of an agent's findings on dfbench's curated validation subset of 27 examples, measured against reported findings in a known vulnerability's file. Its denominator is each agent's own findings.
Benchmark source: Depthfirst
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| Claude Opus 5.5 | 48.5 | dfbench: Opus 5.5 high · depthfirst |
| Gemini 3.6 Flash | 43.5 | dfbench: Gemini 3.6 Flash high · depthfirst |
| Claude Opus 5 | 40.5 | dfbench: Opus 5 med · depthfirst |
| DeepSeek V4.1 Flash | 36.4 | dfbench: DeepSeek v4.1 Flash high · depthfirst |
| Muse Spark 1.3 | 34.1 | dfbench: Muse Spark 1.3 xhigh · depthfirst |
| GPT-6 Luna | 33.3 | dfbench: GPT 6 Luna med · depthfirst |
| Kimi K3 | 33.3 | dfbench: Kimi K3 high · depthfirst |
| GLM-5.2 | 30.8 | dfbench: GLM 5.2 high · depthfirst |
| Grok 4.7 | 30.5 | dfbench: Grok 4.7 high · depthfirst |
| Grok 4.5 | 29.5 | dfbench: Grok 4.5 high · depthfirst |
| GPT-5.6 Luna | 27.3 | dfbench: GPT 5.6 Luna med · depthfirst |
| GPT-5.6 Sol | 24.6 | dfbench: GPT 5.6 Sol med · depthfirst |
| MiMo-V2.6-Flash | 24.4 | dfbench: MiMo-v2.6 Flash · depthfirst |
| GPT-6 Sol | 22.6 | dfbench: GPT 6 Sol med · depthfirst |
| MiMo-V2.6-Pro | 18.1 | dfbench: MiMo-v2.6 Pro · depthfirst |