Google launches Gemini 4 Carbon for cybersecurity, with access initially restricted

Google launched Gemini 4 Carbon, a model aimed at advanced programming and cybersecurity tasks such as identifying vulnerabilities in controlled environments, exploiting flaws and creating code fixes. Google says the model scored 59 points on the Artificial Analysis index, ahead of Claude Opus 5.5 at 58, GPT-6 Astra at 57 and Gemini 4 Argon at 50. Tecnoblog describes Artificial Analysis as an independent consultancy that measures performance close to real-world AI use.

The model is not yet available to the general public. It will initially be used by digital defense specialists during security evaluations with U.S. authorities, and Google expects to provide API access in one to two weeks. There is no timetable for the Gemini app or browser interface. In tests presented by Google, Carbon scored 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71% on CWE-bench, but trailed competitors on some programming and autonomous computer-operation tests.