Google announced Gemini 4 Hydrogen on November 12, 2026, giving developers and businesses building document assistants, coding tools and other AI software a new option with free and paid access.
The model is available in Google AI Studio and through the Gemini API, the connection developers use to add it to other software. Free access offers a way to experiment before paying usage charges, but the data terms matter before uploading business records.
Google’s specification lists text, images, video, audio and PDFs as supported inputs, with tools for searching files and running code. Hydrogen produces text; it does not generate images or audio.
Free access comes with data conditions
Under Google’s ordinary unpaid-service terms, commands and documents submitted to the model, and its responses, can be used to improve products and reviewed by people. Sensitive, confidential or personal material must not be submitted. Ordinary paid-service terms exclude prompts and responses from product improvement, while allowing limited safety logging. Paid does not mean zero retention.
Those categories are not determined solely by the bill. Paid data treatment also applies to unpaid use in the EEA, Switzerland and UK, and to AI Studio accounts with qualifying Cloud Billing access or Workspace enterprise status. Apps offered to users in those European regions must use Paid Services.
There are also sharing exceptions: Google’s invite-only Contributor tier trades discounts for product-improvement data use, while voluntarily donated datasets receive unpaid-service treatment. Check the project’s actual terms and sharing settings before adding customer files.
Free usage is limited. Google directs users to AI Studio for their active quotas; limits vary with model and account status, and available capacity is not guaranteed.
What goes into the bill
Hydrogen’s Standard paid rates are US$1.50 per million input tokens and US$7.50 per million output tokens. Tokens are the units used to count model usage. The output charge includes internal reasoning, called thinking tokens, as well as the answer you see.
Illustrative calculation, not a measured workload: assume a request uses 10,000 uncached input tokens, 1,000 answer tokens and 3,000 thinking tokens. At those rates, it costs $0.015 + $0.030 = $0.045, or $45 for 1,000 identical requests. Counting only visible answers would understate that bill by half. This excludes search, other tool charges, retries, taxes and infrastructure.
The default thinking level is high, with low, medium and xhigh also available. Google warns that xhigh can increase waiting time and token use, and can overthink simpler prompts. A higher setting is therefore not automatically the best choice for routine work.
Optional Google Search grounding—using search results to inform answers—is unavailable on the free tier. Paid use includes 5,000 monthly search queries shared across Gemini 3 and newer models, then costs $14 per 1,000. One model request can trigger several queries; token charges remain separate.
Leave room for the answer
The 1,048,576-token context allowance is shared by input, thinking and answers. For example, exactly 1,000,000 counted input tokens would leave 48,576 for reasoning and the response, not another million. Default generation is capped at 65,536 tokens; longer output needs a higher allowance and background execution or continuation. A large document collection can therefore leave too little room for a long report.
Google positions Hydrogen for faster everyday execution, while Carbon and Argon target deeper reasoning. That positioning is not a guarantee for a particular business task: its model card acknowledges possible fabricated answers, slowness and timeouts.
Before connecting customer records, use permitted sample material to check answer quality, total billed usage and response time. Teams weighing the deeper-reasoning alternative can read RohitAI’s Carbon access and pricing coverage.