Cisco’s tiny Antares AI catches 150x more bugs per dollar than GPT-5.5

Cisco just threw a rock at the giant AI hornet’s nest. The networking giant released two small, open-source cybersecurity models—Antares-350M and Antares-1B—and the numbers they’re claiming are genuinely disruptive. Developer Aman Priyanshu posted on X that the smaller model identifies roughly 150 times more code vulnerabilities per dollar than sprawling AI agents like Cognition’s Devin Security Swarm. That’s not a marginal improvement. That’s a fundamentally different economic equation for security scanning.

Let’s talk about what that looks like in practice. According to a report from Axios, Cisco’s internal tests had Antares chewing through 500 code repositories in about 15 minutes. Total cost? Under a dollar. The same job using GPT-5.5 took five hours and racked up a bill north of $100. For a CISO staring down a budget spreadsheet, that comparison makes the argument all by itself. And because both Antares models run entirely on local hardware, the sensitive source code never touches a third-party cloud. That alone could be the deciding factor for heavily regulated industries.

The secret sauce isn’t just the size. The technical report reveals a training data mix that’s 72 percent security-specific concepts and 15 percent code search histories. It’s a focused diet, not a generalist buffet. Cisco is also playing a familiar dual strategy here: they’re keeping a larger 3-billion-parameter version for their own commercial products. That bigger model reportedly hangs with GPT-5.5 on accuracy and outperforms open models up to 200 times its size. Give away the razor, sell the blades.

I’m watching two things here. First, Cisco says they’re exploring an industry consortium for open AI security tools—if that materializes, it could standardize vulnerability detection in a way the market hasn’t seen. Second, the performance claims need third-party validation. Vendor benchmarks are marketing until someone independent replicates them. But if these numbers hold up, the era of throwing massive, expensive models at every security problem starts to look wasteful. Sometimes the right tool for the job is the small, cheap one you can run without asking permission.