Antares-3B approaches GPT-5.5 at $0.002 a task

Antares-3B outperforms open-weight models 200 times its size at finding software vulnerabilities, per a paper posted to arXiv.

The economics are the point. A full 500-task evaluation sweep finishes in about 15 minutes on a single H100, working out to under 2 seconds and under $0.002 per task. That puts continuous vulnerability scanning inside a CI budget instead of a frontier-API bill.

The model starts from IBM's Granite base. Supervised fine-tuning on cybersecurity reasoning and repository-exploration data comes first, then reinforcement learning from verifiable rewards, run against real vulnerable repositories. The paper says Antares approaches GPT-5.5's accuracy. It doesn't say Antares beats it, or how much of the gap is left.

If a 3B-parameter model can get this close to frontier accuracy at $0.002 a task, the constraint on security scanning stops being model capability and starts being how often you're willing to run it. Teams that gated vulnerability scans to nightly or pre-release runs because of API cost lose that excuse. Worth watching whether Antares' recipe, RL over verifiable rewards on real vulnerable code, shows up in other narrow, checkable domains where a small model can beat a much larger one on price.