Cisco's tiny open AI models catch vulnerabilities 150× cheaper than GPT-5.5

Cisco has introduced two small open-source AI models designed specifically for cybersecurity: Antares-350M and Antares-1B. Both are optimized to spot vulnerabilities in software code—a task traditionally handled by either manual code review or larger, more expensive AI systems.

The headline claim is cost and speed. According to Cisco's own testing cited by Axios, Antares scanned 500 code repositories in about 15 minutes for under a dollar. By contrast, GPT-5.5 completed the same work in five hours and cost over $100. Developer Aman Priyanshu stated on X that the smallest model catches about 150 times more vulnerabilities per dollar than large AI agents like Cognition's Devin Security Swarm—a striking efficiency metric if accurate. Both models run locally rather than sending code to external servers, a feature that appeals to organizations concerned with keeping proprietary or regulated code on-premises.

The technical foundation involves training on domain-specific data: the models learned from roughly 72 percent security-concept data and 15 percent code search histories. Cisco is not releasing its largest model, a three-billion-parameter version, which it is keeping for its own products. According to the article, this larger version performs close to GPT-5.5 and reportedly beats open models up to 200 times its size. Beyond the two open releases, Cisco is also exploring an industry consortium for open AI security tools, suggesting a longer-term commitment to collaborative security AI development.

Cisco's release of Antares reflects a broader industry trend: small, specialized AI models trained on domain-specific data can outperform massive general-purpose models on narrow tasks. The models were trained on roughly 72 percent security-concept data and 15 percent code search histories—a focused diet that likely accounts for their efficiency at vulnerability detection. The cost and speed advantage over GPT-5.5 is striking: a task that costs over $100 and takes five hours with the larger model can be completed for under a dollar in 15 minutes with Antares, a difference that matters for companies running continuous security scanning on large codebases.

By releasing Antares-350M and Antares-1B as open models while retaining a larger three-billion-parameter version internally, Cisco is hedging its bets: it gains community adoption and reputation through the open release while preserving the most capable variant for its own commercial products. The parallel exploration of an industry consortium for open AI security tools suggests Cisco may be betting that security AI will become a collaborative infrastructure layer, similar to how open-source libraries operate today.

No comments yet. Be the first to share your thoughts!