Google releases Gemini 4 Hydrogen with new migration controls Google releases Gemini 4 Hydrogen with new migration controls

A wider release with different access routes

Hydrogen is rolling out through the Gemini app, Search, AI Studio, the Gemini API and Google's business products. Google positions it as an execution model for routine coding and multi-step work, alongside Carbon and Argon for deeper reasoning. Access differs by product: Antigravity requires a paid Google AI Pro or Ultra subscription. Free Gemini app users receive Hydrogen through automatic selection when Google considers it suitable for their prompt. That distinction matters when a team wants a specific model for repeatable testing instead of whatever the consumer app selects.

The request budget changes shape

Google's documentation replaces Flash's separate input and output allowances with one shared pool for input, thinking and the final answer. A long document therefore leaves less room for the model's reasoning and response. Synchronous output remains capped at 65,536 tokens. Longer responses require background execution with a larger output allowance, or client-side continuation. Merely turning on background mode leaves the default limit unchanged. The default thinking setting also moves from medium to high. The four supported levels are low, medium, high and xhigh; the old minimal setting returns an error.

The model reads more formats than it produces

The developer specification lists text, images, video, audio and PDFs as inputs, with text as the output. Function calling, structured responses, file search and code execution are supported. Image generation, audio generation and the Live API are separate capabilities that this model does not provide. This makes Hydrogen a reader and operator inside a broader application: it can interpret a recording or document and produce instructions or a structured result. Generating a spoken reply or a new image still needs another component, with its own integration and billing.

Thinking is part of the output bill

Standard API pricing is $1.50 per million input tokens and $7.50 per million output tokens, including thinking. For example, 100,000 uncached input tokens and 20,000 billed output tokens cost $0.30 before grounding or tool charges. Google's table lists implicit cached input at $0.075 per million tokens. It also distinguishes data treatment: free-tier content can improve Google's products, while paid Standard content does not. That makes the tier selection a data-handling decision as well as a spending decision for teams evaluating the same model.

Existing clients need more than a renamed model

The migration checklist removes explicit cached-content references and ignores custom temperature, top-p and top-k values. It also requires declared native tools or a structured-output schema instead of instructions that imitate tool calls in raw text. Conversation state and thought signatures must survive subsequent turns. These changes affect the surrounding client, even when the application asks the same question. A successful first response does not exercise the cache path, a later tool result or a long-running job.