November 12, 2026. Google’s Gemini API release notes date the general availability of Gemini 4 Hydrogen (gemini-4-hydrogen) to today. This is a released API model, not merely a preview rumor. For teams running assistants and automated workflows, the practical question is which settings deliver a useful answer without unnecessary expense. Release notes.
The answer is not the whole bill
Google lists Standard paid pricing of US$1.50 per million input tokens and US$7.50 per million output tokens, including thinking tokens. Those are processing charges, not a flat fee per task. Free-tier access exists, but the pricing table marks free-tier content as used to improve Google products and paid-tier content as not used for that purpose. Do not treat a free trial as permission to submit confidential company material; check applicable terms and your organization’s policy. Pricing.
More thinking is a choice, not an automatic improvement
Hydrogen defaults to high thinking, whereas Gemini 3.8 Flash defaults to medium. Its additional xhigh setting can spend more tokens and time. Google warns that simple extraction prompts can overthink or exhaust max_output_tokens at xhigh. For an IT team, a short ticket-classification task and a complicated incident investigation need not use the same setting. Compare correctness, completion rate, elapsed time and billed usage on representative non-sensitive tasks before changing production defaults. Thinking guide.
Check the agent configuration too
For antigravity-preview-09-2026, omitting agent_config now selects Hydrogen. That does not mean every saved agent automatically changes: Google says agents created with agents.create retain the model set at creation. Inventory both direct calls and saved agents so that a model change is intentional. Agent model selection.
What we have not established
This report checks Google’s published release and configuration documentation. We have not benchmarked Hydrogen, verified a universal speed or quality advantage, or established drop-in compatibility for existing integrations. Treat a staged evaluation and a rollback plan as part of the upgrade, not proof that the new model will suit every workflow.