Google has launched Gemini 4 Carbon, a model built for sustained engineering and cybersecurity work, and is limiting access to selected cyber defenders through its Fairwind program. Ordinary developers and consumers do not have access yet.

According to Google’s announcement, the model scored 88.2% on SEC-bench Pro, a test of finding real software vulnerabilities, and 71% on CWE-bench, which tests producing fixes for categories of security weakness. It also scored 59 on the Artificial Analysis Intelligence Index v5. Those scores are Google’s own. Reuters reported that its secure-code success rate is more than twice that of Gemini 4 Argon on Google’s internal benchmarks. Introductory pricing is $2 per million input tokens and $10 per million output tokens.

Carbon follows Gemini 4 Argon, which launched on October 14 with a similar limited rollout, by about two weeks. More trusted partners will be added to Fairwind gradually, and the model is in voluntary testing with the US government.

Why it matters: Google is now releasing its strongest security-capable models to defenders first. The approach assumes the same skills that find vulnerabilities can be used to exploit them. Whether the access limits hold, and how partners are chosen, are not described in the reports I could find.