Gemini 4 Hydrogen now has an official route into everyday coding and document work. Google announced the model on November 12, with distribution across AI Studio, the Gemini API, the Gemini app, and other products. For developers moving an existing integration, the immediate issue is compatibility: some familiar settings cause errors, while others are silently ignored.
Choose the entry point before estimating the bill
Google's launch announcement directs developers to AI Studio and the Gemini API. Access through Antigravity is offered to paying Google AI Pro and Ultra subscribers. In the Gemini app, free users receive Hydrogen through automatic model selection when appropriate; that is not a promise of unrestricted manual selection.
The announced model rates are $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across the million-token context window. Those are separate usage rates, not a subscription price or a fixed charge for finishing a task. Free access in the consumer app should not be read as free API usage.
A successful request does not mean every setting worked
The Google Cloud model documentation specifies four thinking levels: LOW, MEDIUM, HIGH, and XHIGH. HIGH is the default. Sending thinking_budget or selecting MINIMAL produces an HTTP 400 validation error. Custom temperature, top_p, and top_k values are ignored, so an accepted request does not establish that your sampling configuration took effect.
There is also one shared input-and-output allowance of 1,048,576 tokens. Output includes reasoning, function calls, and the final answer. A foreground segment produces at most 65,536 output tokens; longer generation needs continuation handling. The headline context capacity is not a promise that one ordinary response will contain that much text. The long-decode documentation supports continuation through the Interactions API or GenerateContent API (v1beta1): the Google Gen AI SDK handles it automatically, while a custom loop must pass continuation_token, which is supported only in v1beta1, the SDK's default API version.
Our recommendation is to migrate a small, checkable set of tasks before replacing a production workflow. Keep inputs, thinking settings, and acceptance criteria consistent. Record elapsed time, actual usage, and rework rather than deciding from a benchmark headline alone.
Keep a route back to another model
Hydrogen is documented as a short-term generally available release, with at least 45 days' notice before retirement. General availability should not be mistaken for a long-term commitment to this particular model identifier. Keep model selection replaceable in the integration.
There is no need to move every difficult task at once. Google still positions Carbon and Argon for deeper, long-horizon reasoning. Start Hydrogen on bounded execution tasks whose results can be checked, while retaining alternatives for architectural decisions or difficult debugging. Before deployment, verify both parameter cleanup and the fallback path.