Google launches Gemini 4 Hydrogen for developers
Google now lets developers build with Gemini 4 Hydrogen, a model aimed at programming, debugging, web development, and AI agent workflows. It is available in Google AI Studio, Android Studio, Gemini for businesses, and Google Antigravity for Pro and Ultra subscribers.

Google now lets developers build with Gemini 4 Hydrogen, a model designed for everyday programming tasks and AI agent workflows. These agents can chain together multiple steps, such as reviewing a repository, identifying an error, and suggesting a fix.
Hydrogen is designed to work with code in repositories, debug applications, and create websites interactively. It also supports multimodal workflows, meaning tasks that combine text, images, or other types of information.
A model for faster iteration
The model can run on its own with four configurable reasoning levels. In practice, you can choose whether to prioritize speed or devote more resources to problems that need deeper analysis.
It can also work alongside higher-capability models from the same family, such as Gemini 4 Carbon and Gemini 4 Argon. Google envisions this combination as a way to divide the work: Hydrogen can handle frequent tasks, while the larger models take on problems that require more reasoning.
The company says Hydrogen follows instructions more closely and generates fewer made-up claims than previous versions. That should lead to more reliable code, although you will still need to review its responses before adding them to a real project.
Where you can try it
Starting today, developers can access Hydrogen through:
- The Gemini API in Google AI Studio.
- Android Studio.
- Gemini for businesses.
- Google Antigravity, if you have a paid Google AI Pro or Ultra subscription.
Google has also published a guide covering the technical details, setup instructions, and controls for choosing the model's reasoning level.
In Google Antigravity, a feature called Orchestrator can coordinate Hydrogen with Gemini 4 Carbon. The goal is for several models to collaborate on the same task instead of forcing a single model to solve everything.
For you, the most concrete change is the kind of help you can expect while programming: Hydrogen does not just complete a line of code. It can review parts of a project, work on errors, and take part in chained tasks within a development workflow. The final quality will depend on the instructions it receives, the context provided, and human review.
The next thing to watch is how these models perform outside demonstrations: whether they actually save time in large repositories, maintain accuracy during long tasks, and how much it costs to run workflows with several coordinated agents.

