Google launches Gemini 4 Carbon with focus on cybersecurity and long-running agentic workflows

Google DeepMind has released Gemini 4 Carbon to developers, positioning it as the company's strongest model for frontier performance in cybersecurity and long-horizon capabilities. The announcement came from Google's models and research team, which describes Carbon as an architecture built specifically for sustained agentic workflows.

What changes for those building agents

The central proposition of Gemini 4 Carbon is to solve one of the most persistent bottlenecks in applied AI: the ability to maintain coherence and reliable execution across tasks that require multiple steps. According to the release, the model was designed for "sustained agentic workflows" — that is, operations in which an agent must chain decisions, tool calls and verifications without losing the thread.

Carbon tackles complex, multi-step problems in software engineering, full-stack web development and scientific research. The emphasis on reinforced alignment and reliable follow-through indicates that Google is prioritizing predictability of model behavior, not just isolated reasoning benchmarks.

Cybersecurity as a frontier differentiator

Carbon's positioning in cybersecurity is one of the most distinctive points of the announcement. Google states that the model shows "strong cybersecurity capabilities," suggesting applications in vulnerability analysis, secure code review and detection of attack patterns in development environments.

The combination of extended agentic capabilities and cybersecurity responds to a growing market demand: agents that autonomously audit, fix and monitor software systems, but with guarantees that they will not introduce new risks during execution.

Where to test and how to access

Access to Gemini 4 Carbon begins with the paid API tier, available in Google AI Studio and Android Studio. Google AI Ultra subscribers have early access in both AI Studio and Google Antigravity, Google's agent-first development platform.

The mention of Antigravity is strategic: the platform was conceived to orchestrate AI agents in engineering tasks, and Carbon arrives as the model that sustains this type of end-to-end operation. The announcement does not detail per-token pricing or context limits, information that should be confirmed in the official API documentation.