Gemini 4 Hydrogen and the Cost of Running Coding Agents

Google’s Gemini 4 Hydrogen gives coding agents four reasoning levels. See how to split work between Hydrogen and Carbon, and measure what each accepted patch costs.

Gemini 4 Hydrogen cover with a blue Gemini star, a code icon and a test checkmark.

A coding agent can spend a single session reading a repository, changing a few lines, running tests and investigating an unexpected failure. Each of those steps can require a different amount of reasoning. With Gemini 4 Hydrogen, announced on November 12, 2026, Google offers developers a model for everyday coding work with four thinking levels and a price that stays flat across its context window. The practical question is how much it costs to get a patch you can actually merge. Google’s launch announcement

Where Hydrogen fits in a coding workflow

Imagine a PHP application where two workers sometimes process the same queue job. A team might ask Gemini 4 Carbon to trace the failure across the job handler, database transaction and retry logic, then use Hydrogen to write a regression test and implement the agreed fix. A second Hydrogen session could review the diff for unrelated edits and run the relevant test suite. If the patch passes the tests but still leaves the same race condition open, the team can return to Carbon with the failing case and the attempted fix.

Hydrogen can also handle a well-defined issue from start to finish. A bug report with a reproducible failure and a clear acceptance test gives the agent a way to verify its work. Argon and Carbon remain useful options when the cause is unclear, several implementations look plausible, or the fix changes a contract used across the application. For more on that choice, see our comparison of Gemini 4 Carbon and Argon.

Google has built this kind of model switching into Antigravity’s Orchestrator. In Google’s Terminal-Bench 4.0 test, Orchestrator approached Carbon’s success rate while using 24% less quota and completing tasks 19% faster on average. Those figures describe Google’s benchmark setup and its quota measure. For a team paying through the API, the dollar cost depends on the calls each model makes and the tokens those calls consume. Google’s Orchestrator benchmark

Four thinking levels create four different trade-offs

Hydrogen offers Low, Medium, High and XHigh. High is the default. In Artificial Analysis’s November 12 evaluation, the model scored 27, 41, 51 and 54 on its Intelligence Index, respectively. The jump from Medium to High was larger than the jump from High to XHigh on that collection of tests. The index combines several capabilities, so a three-point gain gives no direct estimate of how many more patches will pass your tests. Artificial Analysis evaluation results · Google Cloud model documentation

Individual tests show why the choice matters. In Google DeepMind’s model card, Hydrogen scored 57.0% on the adjusted Terminal-Bench 4.1 evaluation, compared with 63.9% for Claude Sonnet 5.5. On SWE-sweep, which evaluates an agent’s ability to discover and fix bugs, Hydrogen scored 13.5% compared with 3.1% for Sonnet 5.5. The two results give different rankings because they test different kinds of work. Google DeepMind’s Gemini 4 Hydrogen model card

What the API price means for an actual patch

Google announced standard API rates of $1.50 per million input tokens and $7.50 per million output tokens. Its Cloud documentation specifies a shared context limit of 1,048,576 tokens. Input, reasoning, function calls and the final response all draw on that allowance. At the announced rates, an illustrative run with 80,000 billed input tokens and 12,000 billed output tokens costs about $0.21 in model charges before any cache discounts, separate tool fees or fallback calls. Google’s pricing announcement · Google Cloud token limits

A flat token price still leaves room for a wide variation in cost per issue. Two attempts at Medium can cost more than one at High when the agent repeatedly reads the same files and runs the same commands. Conversely, a narrowly scoped change with a fast test can finish at Medium with little reason to spend more. Count retries and fallback calls in both cases.

How to test Hydrogen on your own repository

Take a set of previously resolved issues and give each model the repository snapshot from before the fix. Include routine changes, ambiguous bug reports and at least a few tasks where a patch can pass existing tests while still getting the behavior wrong. Compare Hydrogen at Medium and High with your current setup under the same tool permissions, time limit and acceptance checks.

Track the share of patches accepted, the elapsed time per issue, and the total model spend divided by the number of accepted patches. Keep the cost of failed runs in the numerator. Also record human review time, especially if quick patches often need author follow-up or cleanup. Those measurements show whether Hydrogen saves work on a particular class of issues and which failures are worth sending to a stronger model.

Google has begun rolling Hydrogen out through AI Studio, the Gemini API and the Gemini app. Free Gemini app users can receive it through automatic model selection, while Google announced Antigravity access for paid Google AI Pro and Ultra subscribers. Google Cloud lists the API model ID as gemini-4-hydrogen. Google’s rollout details · Google Cloud model reference

Sources