AI Models Compared for Vulnerability Detection: DeepSeek 4.1 Flash Offers Better Cost-Effectiveness Than GPT-6 Astra

A cybersecurity research team conducted performance evaluations of AI models for vulnerability detection, comparing GPT-6 Astra and DeepSeek 4.1 Flash on a code base with known vulnerabilities. On the first run, Astra found approximately 75% of vulnerabilities while DeepSeek 4.1 Flash found 65%. When each model was run three times, the performance gap narrowed significantly, with Astra finding 90% of vulnerabilities and DeepSeek 4.1 Flash finding 85%. The cost difference is substantial: Astra costs around $1,000 per run compared to approximately $20 for DeepSeek 4.1 Flash. The research reveals that running DeepSeek three times for $60 total identifies more vulnerabilities than running Astra once at $1,000. The presenter notes this has important implications for continuous vulnerability scanning strategies and mentions a full interview with lead researcher Philipe is available.