Google releases Gemini 4 Carbon for security firms

Google released Gemini 4 Carbon on Wednesday, positioning the new model as a dedicated tool for cybersecurity defense. The company announced that Carbon currently leads the Artificial Analysis Intelligence Index v5.0, outperforming Claude Opus 5.5 and GPT-6 Astra. The model also holds the top spot on SEC-bench Pro, a test for discovering real-world bugs, and ExploitGym, which measures a model's ability to turn vulnerabilities into exploits.

Carbon improves upon Argon, a model Google launched earlier this month. The company stated that Carbon demonstrates a significant leap in cyberdefense capabilities, specifically in finding vulnerabilities and generating secure code. Google noted that its internal teams have observed the model running for days to handle complex, multi-step engineering tasks. This design allows the system to persist through dead ends rather than taking shortcuts.

The new model also shows better resistance to prompt injection attacks and improved collaboration with users. Google claims Carbon offers greater transparency by cutting unsupported claims by two times compared to Argon. The company described the release as reflecting the pace at which it is pushing the frontier forward.

Gemini 4 Carbon is initially available only to a select group of cybersecurity defense companies. Google plans to expand access to other partners through its Fairwind Program. The pricing structure mirrors the introductory rates set for Argon. Customers will pay $2 per million input tokens and $10 per million output tokens.

The release marks the latest addition to Google's frontier model lineup. The company emphasized that Carbon is built for completing longer tasks with higher accuracy. The new model aims to reduce errors in code generation while maintaining high performance in security-focused applications.