Google has released Gemini 4 Carbon to developers, calling it its most capable model so far. Rather than adding general-purpose chatbot polish, Google is focusing Carbon on two areas: cybersecurity, and long-running agentic tasks where an AI works through complex, multi-step jobs with little supervision.
Google says Carbon delivers frontier-level performance in both areas. It is aimed at sustained workflows in software engineering, full-stack web development and scientific research, the kind of work where a model has to plan, run, check its own output and keep going for an extended period. Google also says Carbon has stronger alignment and more reliable follow-through. If that holds up in practice, it would address a common complaint about AI agents: they start a task confidently and then drift off track, stall, or quietly cut corners partway through.
For now, those remain Google’s claims. The announcement includes no benchmark scores, pricing details or context window figures, so developers will need to test Carbon themselves before treating it as a clear upgrade over earlier Gemini models or the competition.
Access is limited to developers and paying users, with no word yet on availability in the consumer Gemini app. Carbon is available through the paid tier of the Gemini API in Google AI Studio and Android Studio. Subscribers to Google AI Ultra, Google’s top consumer AI plan, can try it in AI Studio and in Antigravity, Google’s agent-focused development environment. Releasing to developers first is a familiar pattern for Google’s top models, and it puts Carbon in front of the people most likely to stress-test its coding and agent claims.
The timing is notable. Anthropic released its Claude Fable 5.2 and Mythos 5.2 models the same day, and competition among the major AI labs now centres on autonomous coding agents and security work rather than conversational quality. Each lab is trying to show that its model can be trusted with a codebase for hours without needing a human to step in.
The cybersecurity focus deserves particular attention. A model that is very good at finding and reasoning about vulnerabilities can protect software, and the same capability could help attackers. Other AI labs have restricted access to their most capable security models or added specific safeguards around them. Google’s brief announcement says nothing about what protections apply to Carbon’s security capabilities, who can use them, or whether its paid-tier availability brings extra vetting.
That is the main open question. If Carbon is as strong at security work as Google says, developers will want to know what limits apply to it.