Gemini 4 Hydrogen explained: price, thinking levels, free tier and how it compares with Carbon and Argon

Gemini 4 Hydrogen is a large language model from Google, the third member of the Gemini 4 family after Gemini 4 Carbon and Gemini 4 Argon. Google describes Carbon and Argon as its frontier models for long, deep reasoning, and Hydrogen as the fast model for everyday coding and agentic work. Agentic work means the model takes several steps on its own, such as reading files, calling tools, running tests and fixing what failed.

It reads text, images, video, audio and PDF files and writes text. It can call functions, run code, use Google Search and Maps for grounding, follow a JSON schema for structured output, and operate a computer screen through computer use. It cannot generate images or audio and does not work with the Live API, according to its model page.

Google is putting it everywhere at once: the Gemini API in AI Studio and Android Studio, Gemini for business, the Gemini app, Google Search for all users, Gemini in Google Sheets, and Google Antigravity for paid Google AI Pro and Ultra subscribers. It is also now the default model behind the Antigravity agent in Gemini Managed Agents.

A thinking level controls how much the model reasons before it answers. More thinking usually means better answers on hard problems, but also more waiting and more output tokens to pay for. Hydrogen offers low, medium, high and a new xhigh, and uses high when you do not choose.

Google's advice in the developer guide is practical: start with high, keep xhigh for hard multi-step problems where high falls short, and note that xhigh can overthink simple prompts. Run your evaluation set at all four levels with a generous max_output_tokens so answers are not cut off, then pick the lowest level that meets your quality bar.

On Google's numbers, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at high, one point above Argon's 50, and 54 at xhigh. Those are Google's reported results, not ours.

Read next: Thinking step by step: what reasoning buys you and what it costs

Orchestrator is a new choice in the model menu of Google Antigravity, Google's coding agent app. Instead of using one model for a whole task, it switches between Carbon and Hydrogen inside the same session. In Google's example, Carbon explores the code and writes a plan, Hydrogen makes the edits and runs the tests, and if Hydrogen gets stuck it hands back to Carbon to repair the plan.

Google says this reached nearly Carbon's success rate on Terminal-Bench while using 24 percent less quota on average and finishing 19 percent faster. The Antigravity post names Terminal-Bench 4.1 and the main launch post names Terminal-Bench 4.0 for the same result, so treat the version as unconfirmed. Orchestrator (Carbon) and Orchestrator (Argon) are rolling out to Google AI Ultra subscribers, and Orchestrator (Argon) to paid Google AI Pro subscribers.

This is the same idea as model routing, which many teams already build by hand: send planning and hard decisions to a strong, expensive model, and routine execution to a fast, cheap one. The difference Google stresses is that the switch happens many times within one task, not once at the start.

Read next: Model routing and cascades: sending each request to the cheapest model that can do it

Hydrogen is Google's answer to a question every team building agents now faces: which model should do the routine steps? A frontier model on every step is slow and expensive, and most steps in an agent run are routine, such as reading a file, running a test or making a small edit. Hydrogen is priced and tuned for those steps, and Orchestrator shows Google expects it to sit beside a bigger model rather than replace it.

If you are on Gemini 3.8 Flash today, there is no rush on price alone, because Flash stays cheaper until 1 January 2027. Test Hydrogen on your own tasks at medium and high, count how many turns and tokens it really uses, and switch when the numbers say so. If you use explicit caching, plan that change first, because Hydrogen does not support it.

Read next: The AI Engineering course: models, tokens, agents and the cost of running them