Google releases Gemini 4 Carbon to lead AI cybersecurity benchmarks

Google released its newest frontier model, Gemini 4 Carbon, on Wednesday. The company stated the system offers major improvements in coding, agentic tasks, and cybersecurity capabilities.

Gemini 4 Carbon currently ranks first on the Artificial Analysis Intelligence Index version 5.0. Claude Opus 5.5 and GPT-6 Astra follow in the rankings. The model also holds the top position on SEC-bench Pro and ExploitGym, two benchmarks that measure the ability to find real-world bugs and convert vulnerabilities into exploits.

Availability is currently limited to select cybersecurity defense firms. Google plans to expand access to other partners within its Fairwind Program. This phased rollout follows the release of the Argon model earlier this month. Carbon improves upon Argon by running for days to resolve complex, multi-step engineering problems. Google noted that internal teams observed the model persisting through dead ends rather than taking shortcuts.

The company reported that Carbon reduces unsupported claims by two times compared to its predecessor. It also shows better resistance to prompt injection attacks and improved collaboration features. Google described the release as reflecting the pace of its frontier advancements.

Pricing for Gemini 4 Carbon matches the introductory rates set for Argon. Input tokens cost two dollars per million, while output tokens cost ten dollars per million. This pricing structure allows enterprises to test the model's extended task handling without a cost increase.

The release positions Google ahead of OpenAI and Anthropic on specific security metrics. The model's focus on long-horizon tasks and secure code generation targets complex infrastructure challenges. No other major details on feature expansions were provided in the initial announcement.