Cisco Foundation AI released two open-weight language models today: Antares-350M and Antares-1B. These aren't toys. They're built to do one job and do it better than anything else at the price point, finding security vulnerabilities in codebases by linking CVE entries to specific files.
The benchmark tells the story. On Cisco's 500-entry vulnerability dataset, Antares scanned repos in about an hour for under $1. GPT-5.5 took 4.5 hours and cost $141. Both run locally, which matters. Sensitive source code stays on your hardware, not some lab's servers. Antares isn't leading on code generation or reasoning benchmarks because it isn't built for those. It's built to catch the thing that actually matters to security teams: knowing which files have known bugs.
There's a pattern here that keeps repeating and hasn't quite broken through in the headlines. The frontier model labs are in a speed race they can't win at scale. You need a model that runs fast, costs nothing, and solves a specific problem? Build it. Train it on the thing you care about. Ship it. The labs are still treating specialization as a downgrade from generality, but teams with actual security problems don't care that a model can write essays about Renaissance art if it's too slow and expensive to scan their codebase on Monday morning.
Cisco also announced Antares-3B is coming soon, running locally on modest hardware. That's the whole architecture: small models, narrow purpose, works offline, costs pennies. The existing releases are on Hugging Face. This is what happens when you build for practitioners instead of benchmarks.
The counterargument is worth saying out loud: frontier models are improving so fast that a specialized model trained months ago will be leapfrogged by the next GPT or Claude release. Maybe. But Antares' efficiency advantage compounds. Cheaper to run means more frequent scanning. More frequent scanning catches real bugs before they matter. The frontier models will get better at security tasks too, but they'll always be heavier and more expensive. The question isn't whether GPT-5.6 will outperform Antares on some test set. The question is whether anyone will pay $141 per scan when Antares costs $1 and runs on a local box.
There's also a subtext here about the infrastructure game. As frontier models get bigger and more expensive to train and deploy, the real economic value shifts to people building the tools that make those expensive systems unnecessary for specific jobs. Cisco isn't competing with OpenAI here. Cisco is showing that the thing you actually want, fast, cheap vulnerability scanning, doesn't need to compete with them.
SOCIAL SHARE CARD GENERATOR