Google introduces Gemini 4 Carbon model for cybersecurity and long-running software work

Google announced Gemini 4 Carbon on 28 October 2026. According to a post on Google's blog by Koray Kavukcuoglu, senior vice president at Google DeepMind, Carbon builds on Gemini 4 Argon, released two weeks earlier, and is designed for long, multi-step tasks such as tracing a bug to its root cause in a large codebase and patching it.

• A small group of cyber defense teams gets the model first, starting in the week of the announcement; Google will then widen access step by step to more trusted partners through its Fairwind Program.
• In results published by Google, Carbon scored 88.2 percent on SEC-bench Pro for finding vulnerabilities and 71.0 percent on CWE-bench for patching; these tests were run without guardrails.
• Attack success on an indirect prompt injection benchmark was reported at 0.0 percent, compared with 0.7 percent for Argon.
• Introductory pricing matches Argon at 2 dollars per million input tokens and 10 dollars per million output tokens, rising to 4 and 20 dollars after the introductory period.

Why it matters

Google describes the staged rollout as an intentional choice: defenders get access first, and developers, businesses and consumers will follow once more feedback on guardrails has been gathered. The post also says Google is taking part in the US government's voluntary pre-release testing process. Companies in Türkiye cannot access Carbon directly yet, but for IT teams connecting AI agents to business processes, resistance to prompt injection could be one of the measures to check when choosing a model. It may make sense to watch for the general availability date and to keep the permissions given to agents narrow.