Google launches Gemini 4 Carbon with a ‘significant leap’ in cybersecurity

Google has unveiled Gemini 4 Carbon, its most recent frontier model, boasting "significant enhancements" in coding, agentic tasks, and cybersecurity. In a statement on Wednesday, Google emphasized Carbon's top position on the Artificial Analysis Intelligence Index (v5.0) benchmark, followed closely by Claude Opus 5.5 and GPT-6 Astra.

The new Gemini 4 Carbon model represents an improvement over the recently released Argon model. Currently, it is accessible to a limited set of cybersecurity defense firms, but it is set to become available to other Fairwind Program partners in the near future.

Google describes Carbon as embodying the company's "pace at which we're pushing the frontier forward," with a "significant leap" in cyberdefense that enables it to better identify vulnerabilities and produce secure code. Carbon currently holds the top spot on SEC-bench Pro, which gauges an AI model's capacity to uncover real-world bugs, and ExploitGym, which assesses a model's ability to convert vulnerabilities into exploits.

The model is designed to handle longer tasks, with Google's internal teams observing it running for days to navigate through obstacles and address complex, multi-step engineering tasks. Additionally, Carbon demonstrates a stronger resistance to prompt injection attacks and collaborative tendencies with users, as reported by Google.

Furthermore, Carbon aims to provide more "transparency and honesty" by reducing "unsupported claims" by half when compared to its predecessor, Argon. The model will be introduced at the same competitive introductory price as Argon, priced at $2 per million input tokens and $10 per million output tokens.