Google launches Gemini 4 Hydrogen with global availability and per-token pricing
Image generated by the Inferama Editorial Team
01

A model built to handle everyday tasks

Google introduced Gemini 4 Hydrogen on November 12, 2026, as a model for everyday coding and agentic workflows. The company distinguishes it from Gemini 4 Carbon and Gemini 4 Argon, which it describes as models for deep reasoning and long-horizon tasks. Hydrogen is designed for fast execution and for working alongside those models in workflows that combine planning with operational tasks.

Google’s model card says Hydrogen is adapted from Gemini 4 Carbon. Its unified context window supports up to 1,048,576 tokens shared between input and output; that is not one million input tokens plus a separate million output tokens. The default synchronous output limit is up to 65,536 tokens. The card also describes an option for up to one million tokens of background long decode, within the shared context window and through the corresponding workflow.

The model accepts text, images, audio, and video, and generates text. The model card identifies coding, debugging, repository navigation, multimodal tasks, and tool use as intended scenarios. That orientation does not mean the model acts without oversight or that its answers are correct by default: Google itself lists limitations such as hallucinations, delays, and timeouts.

Google announced a rollout of Hydrogen across its developer, enterprise, and consumer products. Actual availability depends on the product, plan, and access method. In Gemini, for example, Google says free accounts may receive it through automatic model selection when appropriate for a prompt; that does not guarantee a manual selection option for every free account.

02

Access channels and product differences

The November 12 announcement describes a global rollout across Google AI Studio, the Gemini API, Gemini for business, the Gemini app, and Google Search, as well as Google Antigravity for Google AI Pro and Ultra subscribers. The model card also lists Google Sheets and other distribution channels. A channel’s appearance in the announcement does not mean every feature is available with every subscription, in every region, or through every usage mode.

For developers, Google announced access through the Gemini API and AI Studio. For businesses, access is offered in Gemini for business through the API or model selector, in Google Antigravity, and in Gemini in Sheets for eligible business subscriptions. In the Gemini app, the model selector offers Hydrogen depending on the plan; free accounts may access it through automatic selection when the system deems it appropriate. Google Search is included in the rollout announced for all users, but the announcement does not specify a direct option to choose the model.

The API is the route for integrating the model into custom applications and automated workflows. That flexibility involves configuring authentication, managing quotas, and checking the applicable terms of service. The model card says no special hardware or software is required to use the model, but it points to different terms for the API and enterprise products. The model’s name alone does not confirm that a particular feature is enabled for a specific account or product.

It is useful to distinguish access to the model from the capabilities of products that incorporate it. The app may add tools, connectors, and its own limits; Search presents answers within its own experience; and a Gemini API call depends on the implementation and selected parameters. The announced global rollout does not establish that every feature, integration, or plan will become available at exactly the same time in every market.

What to check for each access channel

ChannelAnnounced accessWhat to check
Gemini API and AI StudioDeveloper accessQuotas, API terms, parameters, and tool charges.
Gemini appModel selector depending on plan; automatic selection for free accounts when applicablePlan, account availability, and whether Hydrogen appears as a manual or automatic selection.
Google SearchRollout announced for all usersThe experience available in your account; the announcement does not specify a manual model selector.
Gemini for business and SheetsAccess subject to business eligibility or subscriptionSubscription, workspace permissions, and service terms.
Google AntigravityAccess announced for certain Google AI plans and business usesSubscription, usage mode, and availability of the specific feature.
03

API pricing and cost limits

Google announced standard paid rates for Gemini 4 Hydrogen of $1.50 per million input tokens and $7.50 per million output tokens. The rates apply to the announced context window, and reasoning is included in the stated price. The published rates do not mean every request costs the same: spending depends on the number of input and output tokens, the usage mode, and any additional services involved.

Google’s model card presents these prices as flat rates across the context window. For Batch and Flex, the pricing catalog lists $0.75 per million input tokens and $3.75 per million output tokens; Priority has different rates. Implicit caching costs $0.075 per million input tokens. These alternative rates apply to different modes and should not be confused with the standard rate or assumed to apply automatically to every Google product.

A simple estimate can separate input and output: multiply each token count by its per-million price, then add the amounts. For example, a request with 100,000 input tokens and 20,000 output tokens, at the announced standard rates, would cost approximately $0.30: $0.15 for input and $0.15 for output. The example illustrates the calculation only; it excludes potential tool charges and other billing conditions.

Google says the free tier has limits and conditions on data use. Free access should therefore not be treated as unlimited capacity or as an alternative with the same terms as a paid plan. Before deploying an integration, it is practical to test representative requests, measure input and output usage, and set spending limits.

Published API rates

Rates depend on the usage mode. Confirm the applicable terms before preparing a budget.

ModeInput per million tokensOutput per million tokens
Standard$1.50$7.50
Batch and Flex$0.75$3.75
Implicit caching$0.075 for inputNo output rate stated in the published figure
PriorityDifferent ratesDifferent rates
04

What the published tests show—and what they do not

Google’s model card reports results from several evaluations, including 57.0% on Terminal-Bench 4.1, adjusted for cheating; 58.8% on HLE-Diamond; 53.4% Climb WSR accuracy on MKA; and 71.6 points on Agents’ Last Exam v1.1. These are provider-reported results tied to specific benchmarks and metrics; they are not an independent Inferama test and do not guarantee the same performance on a real-world task.

The figures can only be interpreted correctly alongside the benchmark version, metric, and configuration. The announcement also mentions Terminal-Bench 4.0 for Orchestrator, a system that switches between models within a session. That result should not be attributed to Hydrogen operating alone or combined with Hydrogen’s 57.0% on Terminal-Bench 4.1: the benchmark version and configuration differ.

The model card also documents an internal automated safety evaluation against Gemini 3.8 Flash. Google reports improvements on some indicators and a 5.57 percentage-point increase in unjustified refusals, where a higher value represents a regression. The card says the evaluations are automated, are not equivalent to human evaluation or red teaming, and are not directly comparable with results in earlier cards because the evaluations have been improved.

For frontier safety, Google says Hydrogen showed no material increase in general capability over Gemini 4 Argon in its material capability change assessment. The company reports targeted tests in the biological and cybersecurity domains, stating that the model did not reach the critical capability thresholds but did reach the alert thresholds. The full safety report was listed as forthcoming in the published model card. These are evaluation results and decisions reported by Google, not an independent certification that the model is risk-free.

Results reported by Google

Scores correspond to specific benchmarks and are not guarantees of general performance.

EvaluationReported resultStated scope
Terminal-Bench 4.157.0%Cheating-adjusted success.
HLE-Diamond58.8%Benchmark score.
MKA53.4%Climb WSR accuracy.
Agents’ Last Exam v1.171.6%Benchmark score.
05

Documented limitations and usage criteria

The official model card acknowledges hallucinations, possible delays, and timeouts. It also warns that the model may use more tokens to maximize performance, especially at higher reasoning levels. Its knowledge cutoff is March 2026, although Google notes that knowledge in some domains may be less current. Answers about recent events therefore need verification, even when a task is completed quickly.

Google describes four reasoning levels: low, medium, high—the default—and xhigh. The company presents them as options for adjusting quality, latency, and cost. Changing the level changes the request configuration; it does not guarantee accuracy, valid code, safety, or reproducible results. A low temperature or a higher reasoning level also does not replace testing, human review, or deterministic controls.

For coding tasks, run tests, inspect changes, and check that the model has not modified files outside the intended scope. For agents that use tools, limit permissions, validate data before taking action, and request confirmation for sensitive operations. The model card says Google continues to work on resistance to jailbreaks and indirect prompt injection; safeguards therefore reduce risks but do not justify giving unrestricted access to repositories, accounts, or production systems.

The model is worth evaluating when speed and iterative execution matter, especially for everyday coding or multimodal workflows. To assess it carefully, compare its results on your own workload, track input and output tokens, check which channel and plan are available, and validate any output that could affect users, data, or systems with appropriate tests.

Practical evaluation before integration

  1. 01Define a representative task and a verifiable outcome, such as a test that must pass or a document that must meet explicit requirements.
  2. 02Test the channel and reasoning level you would use in production; record latency, input and output tokens, errors, and cost.
  3. 03Review every tool action and limit permissions, data, and operations to what is necessary.
  4. 04Compare results with an alternative and repeat the test with edge cases; do not treat a single run as a guarantee.
  5. 05Before deployment, add deterministic checks and human oversight for high-impact decisions or changes.
06

What to consider when deciding

Gemini 4 Hydrogen arrives as Google’s fast-execution model, with access announced across consumer, enterprise, and developer products. Its key published specifications include a shared context window of up to 1,048,576 tokens, default synchronous output of up to 65,536 tokens, and standard API rates. Each figure has conditions: input and output share the context capacity, long output requires a specific workflow, and pricing varies by mode and usage.

Benchmarks and safety evaluations describe bounded results, not a universal measure of quality or safety. The model card acknowledges limitations, and Google still listed the full safety report as forthcoming. For real-world applications, decisions should rest on tests using representative tasks, cost assessment, and controls appropriate to the risk.

Before using it, confirm that the model is enabled for your product and plan, check the API terms if you are integrating the service, and verify current facts or important actions outside the model. For code and agents, test changes and restrict permissions. For costs, measure input and output tokens separately and account for the rates that apply to each mode.

07

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction