Cisco's Antares: 350M params beating GPT-5.5 at code security
cisco released antares-350m and antares-1b, open models that claim to find 150x more vulnerabilities per dollar than gpt-5.5. this challenges the notion that bigger models are better for enterprise security.
cisco open-sourced antares-350m and antares-1b, small language models specialized in detecting software vulnerabilities. developer aman priyanshu reported that the 350m model identifies 150 times more vulnerabilities per dollar than large agents like cognition's devin, while internal tests show it scanned 500 code repositories in 15 minutes for under a dollar. by comparison, gpt-5.5 took five hours and cost over $100 for the same task. because the models run locally, they address data privacy concerns by keeping sensitive code on-premise. cisco retains a 3b parameter version for internal products that reportedly rivals gpt-5.5 and explores an open-source security consortium, positioning specialized small models as a superior economic and architectural choice for enterprise defense.