Google released Gemini 4 Hydrogen on November 12, 2026. This guide was published November 15 and checks the current documentation rather than treating launch claims as independently tested results.
Gemini 4 Hydrogen is Google's new option for everyday coding and agent execution. The headline API price is $1.50 per million input tokens and $7.50 per million output tokens, including thinking. The more important deployment questions are whether it completes your tasks reliably, which API surface you use, and how retries, tools and human review affect the final bill.
What Google launched, and who can access it
Google's November 12 announcement positions Hydrogen for rapid execution, with Carbon and Argon remaining options for harder reasoning. Availability includes AI Studio, the Gemini API, Gemini for business, the Gemini app and Search; Antigravity access is offered to paid Google AI Pro and Ultra subscribers. Product access and API billing are separate questions.
Google also describes Antigravity Orchestrator switching between models. Its reported Terminal-Bench 4.0 result approached Carbon's success rate with 24% less quota and 19% faster completion. Those are Google's reported results for that setup, not a guarantee for Hydrogen alone or a Nerova benchmark. We have not independently tested Hydrogen.
Gemini 4 Hydrogen API pricing
The Gemini Developer API price list distinguishes several tiers. This table covers selected tiers, not every option. Prices are USD per million tokens; output includes thinking.
Hydrogen Developer API token rates
| Tier | Input | Output |
|---|---|---|
| Standard paid | $1.50 | $7.50 |
| Batch or Flex paid | $0.75 | $3.75 |
| Contributor | $0.75 | $3.75 |
Contributor is an invite-only private preview with data contribution, not simply another name for Batch or Flex. The pricing page marks free-tier and Contributor content as used for product improvement; the paid Standard, Batch and Flex columns say no. Review your chosen tier's terms before sending customer information. Account, billing and regional exceptions apply. Non-training use does not by itself mean zero data retention. These Developer API terms should not be assumed to describe an enterprise Cloud agreement.
A concrete token-cost example
For a hypothetical request with 100,000 billed input tokens and 10,000 billed output tokens, Standard arithmetic is (0.1 × $1.50) + (0.01 × $7.50) = $0.225. At the listed Batch or Flex rates, the same token quantities are $0.1125. This calculation excludes tool charges, retries, infrastructure, taxes and reviewer time. It does not predict how many tokens your workflow will consume.
Use this worksheet as a unit check, not a procurement decision. A low token bill is attractive only when the output is usable. For example, a workflow that requires an employee to rewrite every answer can cost more overall than a higher-priced route that produces acceptable work consistently.
The million-token window is a shared budget
The official model reference identifies the model as gemini-4-hydrogen and lists a unified 1,048,576-token input/output window. Supported inputs include text, images, video, audio and PDF, with text output. The default synchronous output limit is 65,536 tokens; longer decoding is supported through the documented continuation mechanisms. Explicit caching is unsupported, while implicit caching is available.
Do not plan around one million input tokens plus another million of output. Reserve room for the response and the work needed to produce it. In a document-review pilot, record the size of the assembled request, not merely the uploaded file. Retrieved material, earlier turns and tool results can change the budget substantially. A large context window is capacity, not evidence that every included detail will be used correctly.
Cloud migration: test more than the model name
The Gemini Enterprise Agent Platform guide specifies HIGH as Hydrogen's default thinking level, versus MEDIUM on Gemini 3.8 Flash. It rejects thinking_budget and MINIMAL. These are Cloud-specific integration checks: inventory the actual endpoint and SDK before applying them.
Cloud foreground segments are capped at 65,536 output tokens, including thinking and function calls. Long decoding requires supported continuation on Interactions or GenerateContent via v1beta1; the OpenAI compatibility layer cannot retrieve continuation-token completions. Background Interactions require server storage. Do not confuse that storage choice with the Developer API's product-improvement policy.
A migration test should cover interruption as well as success. Confirm what your application does when a stream ends early, a tool fails, the output budget is exhausted or a user cancels. Ensure a resumed request does not execute the same external action twice. A model change should never silently expand permissions.
How to decide whether Hydrogen belongs in your workflow
Our recommendation is a bounded comparison against the route you already use. Start with representative tasks and write acceptance criteria before seeing results. Include ordinary work, difficult edge cases and tasks that should be escalated to a person. Keep tool access, source material and review standards consistent.
- Measure accepted outcomes. Count tasks that meet the written criteria, not merely responses returned successfully.
- Include every attempt. Add failed calls, retries and repair prompts to the cost of the eventual result.
- Track reviewer minutes. Log correction effort separately from token spend; do not declare savings based on generation cost alone.
- Compare elapsed time. Measure request-to-accepted-result time, including tool waits and approvals.
- Choose a stopping rule. Define a minimum useful improvement and a sample large enough to cover your task mix before expanding traffic.
For a simple scorecard, divide total operating and review cost by accepted tasks. Also retain the failure rate: a route can look cheap per successful result while abandoning important cases. Examine the slowest and most expensive runs rather than relying only on averages.
Hydrogen may be a good execution candidate without becoming the only model in the system. Keep an escalation path for tasks it cannot complete, and evaluate the combined route rather than cherry-picking one successful stage. For a different frontier option, our Gemini 4 Carbon access and pricing guide separates public availability from budgeting decisions.
The business decision
Choose the workflow first, then the model. A short, repeatable process with a clear owner and measurable outcome gives you a better evaluation than an open-ended instruction to automate everything. Nerova's operational-improvement job is a place to explore that scope; it is not a claim that Nerova has integrated or benchmarked Hydrogen.