Google Announces Gemini 4 Carbon for Cybersecurity
According to Google, Carbon topped the updated Artificial Analysis Intelligence Index ranking, surpassing other models.
The company has announced Gemini 4 Carbon—a model for cybersecurity, programming, and extended agent-based scenarios. Initially, it will be made available to a limited group of cybersecurity specialists, with access gradually expanded through the Fairwind Program. Google plans to release it for developers, businesses, and general users later, following additional verification of its protective mechanisms.
In external tests conducted without protective restrictions, Carbon scored 88.2% in SEC-bench Pro for vulnerability detection, 54.8% in ExploitGym for exploitability verification, and 71.0% in CWE-bench for patch preparation. The company also reports improved secure code quality and a reduction in serious behavioral deviations in internal scenarios.
The introductory price for the model will be $2 per million input tokens and $10 per million output tokens. Google states that Carbon is designed for multi-stage engineering tasks and can operate continuously for several days, including root cause analysis of errors and code remediation.
