Gemini 4 Hydrogen arrives: faster agents, pricing and the output limit to check

What Google launched

Google introduced Gemini 4 Hydrogen on November 12, 2026. Its developer announcement positions the model for repository changes, debugging, interactive web development and work across different media. Developers can use the Gemini API through AI Studio and Android Studio, with additional routes through Gemini for business and Google Antigravity.

For a team building an app, that suggests a concrete starting point: a small change with a clear completion check. A faster first response is useful only if the resulting change survives review. That is our interpretation of the release, rather than a result from a Promptwire hands-on test.

Public record: Google developer announcement — November 12, 2026

Pricing and thinking levels

Google lists $1.50 per million input tokens and $7.50 per million output tokens, with the same rates across the one-million-token context range. Hydrogen offers low, medium, high and xhigh thinking settings; high is the default.

An illustrative request using 100,000 billed input tokens and 10,000 billed output tokens would cost $0.225 at those rates: $0.15 for input plus $0.075 for output. This is arithmetic, not a measured workload or a complete integration quote. Budget for repeated attempts and check the billing rules for the service you use.

Google also introduced Antigravity Orchestrator, which can switch models during a session. The company reports that it approached Carbon’s task success on Terminal-Bench 4.0 while using 24% less quota and finishing 19% faster. Those figures describe Google’s evaluation of the orchestration setup, not a guaranteed saving for every Hydrogen request.

Public record: Google launch announcement — November 12, 2026

A million-token window is not the default output length

DeepMind’s model card describes a combined input-and-output context of up to one million tokens. It lists up to 64,000 default synchronous output tokens, with up to one million through background long decode. A large document budget and a large ordinary response are therefore different capabilities.

The card also lists a March 2026 knowledge cutoff, possible hallucinations, occasional slow responses or timeouts, and higher token use at stronger thinking settings. The full Safety & Security Report is marked as forthcoming on the card checked for this article.

Before designing a workflow around a very long answer, confirm which mode your endpoint exposes. Break a large deliverable into reviewable parts when that makes omissions easier to catch. This is an editorial implementation suggestion.

Public record: Google DeepMind model card — checked November 13, 2026

Free app access follows automatic selection

The Gemini Apps help page distinguishes direct model choice from automatic routing. With no paid Google AI plan, users have Auto or Flash-Lite; when automatic selection is enabled, prompts needing more reasoning may use Hydrogen. That does not mean every free prompt runs on Hydrogen.

The help page lists Hydrogen among paid-plan chat options, with task features and model availability depending on the plan and remaining usage. Work and school accounts depend on their Workspace licence. It also says the response-details menu can show which model answered.

For a useful comparison, record that model label alongside the output. Otherwise two apparently similar app sessions may be measuring different systems or different usage conditions.

Public record: Gemini Apps Help — model access, checked November 13, 2026

Promptwire’s take: time the accepted change

Our suggested trial is a bounded website fix: repair a broken mobile menu, preserve its keyboard behaviour and run the existing checks. Give the model the same starting files and acceptance criteria used for your current workflow.

Record elapsed time through review, billed tokens, failed attempts and any manual repair. Try a second task involving a document or screenshot if that reflects your actual work. Keep the examples separate from material used to tune the instructions.

The release makes a practical choice more interesting: which work can a fast execution model complete reliably, and which work still benefits from deeper planning? An answer supported by accepted changes and total cost will be more useful than a single impressive demo. Promptwire has not independently benchmarked Hydrogen.