AI News
AI News AgentModel releaseGoogle2 min read

Gemini 4 Carbon reaches AI developers

Google is making Gemini 4 Carbon available to developers through the paid tier of its API and to Google AI Ultra subscribers. The model is designed for long programming, web development, scientific research, and cybersecurity tasks.

Google now lets developers build with Gemini 4 Carbon, a new model designed for long, complex tasks. The company presents it as its most powerful model, with advanced cybersecurity capabilities and the ability to handle multiple consecutive steps.

Carbon is designed for AI agent workflows. That means it can sustain a task for longer, break it into parts, and make steady progress instead of simply answering an isolated question.

In practice, it can help you:

  • Create and maintain software.
  • Develop complete web applications, from frontend to backend.
  • Support scientific research that requires multiple steps.
  • Detect and resolve computer security issues.

Google also highlights improved alignment, meaning a greater ability to follow intended instructions and objectives, along with more reliable task tracking. The goal is for Carbon not to abandon the work halfway through when a problem requires several linked decisions.

How developers can try it

Gemini 4 Carbon is already available through the paid tier of the Gemini API, the service that lets you integrate the model into your own applications. You can use it from Google AI Studio and Android Studio.

Google AI Ultra subscribers can also try it in AI Studio and Google Antigravity, Google's development platform focused on AI agents.

This is not being presented as a general, free release. To start building with Carbon, you need to use the paid API tier or have an AI Ultra subscription, depending on the tool you choose.

What changes for you

If you develop software, Carbon aims to reduce some of the repetitive coordination work. For example, it could analyze a request, suggest code changes, check for potential security issues, and continue through several iterations without requiring you to explain the entire context again at every step.

That does not remove the need to review its results. In programming, research, or cybersecurity tasks, an error can affect the final product, your data, or user security. Its real usefulness will depend on how well it maintains context and meets objectives during extended processes.

Google is positioning Gemini 4 Carbon for a stage in which models do more than answer questions: they work alongside developers on complete projects. What matters now is whether its ability to handle long-running tasks holds up in daily use once problems move beyond controlled demonstrations.