Google introduced Gemini 4 Carbon with reported gains in cybersecurity and coding, initially making the model available to selected cyberdefense companies.
Google announced Gemini 4 Carbon on October 28, 2026, presenting it as a frontier model with improvements in coding, agentic work and cybersecurity. The company says Carbon ranks first on the Artificial Analysis Intelligence Index v5.0, ahead of Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra.
Google also points to Carbon’s results on two security-focused evaluations. The model currently leads SEC-bench Pro, which tests the discovery of real-world software bugs, and ExploitGym, which measures whether a model can turn vulnerabilities into exploits. The Verge reports that Google says Carbon is better at finding vulnerabilities and producing secure code; those performance claims are the company’s, rather than independent assessments in the article.
Google describes the model as suited to extended, multi-step work. The company says internal teams have run it for days on engineering tasks, and reports improvements in resisting prompt-injection attacks and avoiding unsupported claims. Google says the latter were reduced by a factor of two compared with its earlier Gemini 4 Argon model.
Access is currently restricted to a select group of cybersecurity defense companies. Google says it plans to expand availability to other partners through its Fairwind Program, but the announcement does not give a date for a broader rollout. The introductory price matches Argon’s: $2 per million input tokens and $10 per million output tokens.
The launch positions Carbon as both a general-purpose frontier model and a security tool, but its early access remains narrow. The reported benchmark placements and internal task results offer a snapshot of Google’s claims at launch; wider partner access will provide a chance for more organizations to assess the model in their own workflows.