Google Unveils Gemini 4 Carbon for Autonomous Security Work

Google has announced Gemini 4 Carbon, a flagship model tuned for cybersecurity, coding and multi-day autonomous agent work: it locates vulnerabilities, verifies they can be exploited, and patches the code. It takes first place on the Artificial Analysis Intelligence Index v5.0 with 59 points. Published scores are 88.2% on SEC-bench Pro for finding real vulnerabilities, 54.8% on ExploitGym for exploitability, 71% on CWE-bench for automated fixes, and 1769 Elo on WebDev Arena, fourth for interface work. Google reports Carbon runs autonomously for several days, tracing root causes and fixing them in large codebases, generates secure code twice as often as Gemini 4 Argon in internal tests, produces Chrome patches the browser team rates 10x more effective than another frontier model’s, cuts serious instruction violations on code tasks by 80%, and makes half as many unfounded claims as Argon. The model reportedly withstood every prompt-injection attack tested in one benchmark. Access is limited at first to security specialists, with broad release later and API pricing of $2 per million input tokens and $10 per million output tokens on an introductory tier.

Introducing Gemini 4 Carbon