AI DevelopmentNew Release8 min readPublished November 12, 2026

A practical model choice starts with the completed task

Gemini 4 Hydrogen: Price, Benchmarks and When to Use It

Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 output. Compare thinking levels, benchmark evidence and the costs of real agent work.

DA
Digital Applied Team
Senior strategists
PublishedNovember 12, 2026
Read time8 min
SourcesGoogle API documentation and Artificial Analysis

Gemini 4 Hydrogen gives developers a new candidate for everyday coding and tool-using agents at $1.50 per million input tokens and $7.50 per million output tokens. Google made gemini-4-hydrogen generally available on November 12, with free and paid API access, according to its Gemini API release notes. The paid API defaults to high thinking and adds an xhigh setting for harder work.

The important decision is whether Hydrogen can finish your recurring tasks reliably enough to justify its total cost. A model that answers quickly but needs several repairs can cost more than a slower one that succeeds immediately. Start with the work: a coding change, a document review or a sequence of tool calls. Then choose the model and thinking level around a definition of success that somebody can actually check.

Key takeaways
  1. 01
    Price the full responseOutput billing includes thinking, so the visible answer alone does not describe the bill.
  2. 02
    Choose thinking deliberatelyThe API offers low, medium, high and xhigh; high is the default rather than a universal recommendation.
  3. 03
    Keep benchmark claims in contextGoogle reports its own task results. Artificial Analysis supplies a separate evaluation with different tasks and settings.
  4. 04
    Measure completed workUse successful tasks, retries, elapsed time and review effort to decide whether to expand a pilot.

01 — PositioningStart with the everyday workload

Google positions Hydrogen below the most demanding long-horizon reasoning work served by Carbon and Argon. Its Hydrogen introduction emphasises coding, tool use and multimodal work. That positioning is useful as a shortlist, but it is not a promise that every everyday request is easy. A small change in a payment calculation can demand more care than a much larger change to decorative copy.

For a development team, a sensible first workload might be a bounded maintenance issue with a reproducible failure and an existing test. Hydrogen can be asked to inspect the relevant files, propose a change and explain the evidence. The team still decides what permissions it receives and what makes the change acceptable. A passing test is useful evidence; a confident description of a test that never ran is not.

For a content team, consider extracting facts from approved source documents before drafting a product comparison. Keep the source references with each extracted claim and have an editor check consequential details. The model choice should reflect the cost of an incorrect answer. Calling both tasks everyday does not make their acceptance criteria interchangeable.

Coding
Bounded changes
Start with a known failure

Use a reproducible issue, a small change surface and a clear check. Review what changed as well as whether the command succeeded.

Good pilot
Documents
Source-based extraction
Keep evidence beside claims

Check that figures, dates and exclusions survive extraction. A fluent summary can still omit the condition that matters.

Good pilot
Architecture
Difficult judgments
Allow an escalation path

A lower-priced default can hand unresolved decisions to a stronger model or a person. Escalation needs an explicit trigger.

Compare separately

02 — Token pricesSeparate the rate from the final bill

Google’s standard API pricing lists Hydrogen at $1.50 per million input tokens and $7.50 per million output tokens, with $0.075 per million for implicit cache hits. Thinking tokens count as output. Prices are flat across the model’s full context window. These are token rates, not a subscription price or a fixed charge for an agent completing a job.

The distinction matters because agents may make several model calls, repeat context, use tools and retry. A proposal that multiplies the final answer’s word count by the output price leaves out part of the work. Build a budget from recorded usage fields and tool charges, then add the cost of the infrastructure that executes the workflow. Keep human review visible rather than treating it as free.

The comparison with Carbon and Argon also has a calendar attached. Their introductory standard rates end after November 30. If a pilot starts in November but the production forecast covers December, using only the introductory price creates an avoidable surprise. The earlier Carbon access and pricing analysis explains that separate buying decision.

Google standard Gemini API rates, retrieved November 17, 2026. Output includes thinking; tool charges and other consumption options are separate. Source: Google API pricing linked above.
ModelInput / output per 1MPeriod
Hydrogen$1.50 / $7.50Standard current rate
Carbon and Argon$2 / $10Through November 30, 2026
Carbon and Argon$4 / $20From December 1, 2026

03 — Worked exampleCalculate a task before forecasting a month

Consider an illustrative text-only call with 20,000 uncached input tokens and 4,000 billed output tokens, including thinking. At Hydrogen’s standard rates, input costs $0.03 and output costs $0.03: $0.06 in total. One thousand identical calls would cost $60 before tools, infrastructure, taxes or retries. This is arithmetic using an assumed workload, not a measured Hydrogen benchmark.

Now suppose an accepted task requires two such calls. The model charge becomes $0.12 per completed task, before any failed attempts. If a cheaper configuration produces more retries, the apparently lower rate can lose its advantage. Conversely, a higher-thinking setting may justify itself if it reduces costly rework. Neither outcome follows from the rate card; both require observations from the task you intend to run.

Use a small ledger with task identifier, model, thinking setting, billed input, billed output, tool costs and acceptance result. Record failed runs too. Excluding them makes a pilot look economical precisely when reliability is costing money. Keep the examples repeatable and remove private customer material unless the approved data-handling arrangement covers its use.

Illustrative arithmetic

20,000 ÷ 1,000,000 × $1.50 + 4,000 ÷ 1,000,000 × $7.50 = $0.06. The output assumption includes thinking. It does not predict how many tokens Hydrogen will actually use.

04 — Thinking settingsTreat deeper thinking as a choice

The Hydrogen API guide lists low, medium, high and xhigh thinking, with high as the default. These settings give you a practical control to test. They should not become a hierarchy in which the largest label is automatically the professional choice. The useful setting is the one that meets the task’s quality and time requirements at an acceptable total cost.

A deterministic extraction task might not benefit from the same deliberation as an ambiguous debugging problem. Build paired examples: a straightforward case, a missing-data case and a case where the correct response is to ask for clarification. Compare whether the setting improves the difficult cases without turning simple requests into an unnecessarily slow interaction. Keep tool availability and instructions constant during that comparison.

Choose the setting explicitly in production configuration so a default does not silently become a product decision. Also keep application defaults separate from API defaults: a setting shown inside the Gemini app is not evidence of how your own API request runs. If you later change the setting, treat it as a behavior change and rerun the small acceptance set that justified the original choice.

A useful experiment

Compare settings on the same representative tasks and record correctness, time to useful completion and billed tokens. Do not declare a winner from one impressive answer.

05 — Benchmark evidenceRead the score beside its test conditions

Google reports a Terminal-Bench 4.1 result of 57% for Hydrogen against 18.2% for Gemini 3.8 Flash, alongside a claim of 2.5-times-faster task completion at Hydrogen’s high setting. These are Google-reported results from its launch article, retrieved November 17. They support investigating the model, but they do not establish a success rate for an unrelated repository or customer workflow.

The Artificial Analysis Hydrogen page, also retrieved November 17, reports an Intelligence Index score of 54 for xhigh and an average $3.16 per index task. That is a separate evaluation. An index task is not your support ticket, website edit or document review. The score is not a percentage accuracy figure, and the cost is not an advertised fixed fee for arbitrary work.

The practical interpretation is to use public evaluations to select candidates, then decide using your own acceptance set. A strong aggregate score can coexist with a specific weakness that matters to you. Keep the evaluator, thinking setting and retrieval date beside any number copied into a procurement document. Comparing a current score with an older version of the index can create a ranking that the evaluator never published.

Interpretation guide. Google and Artificial Analysis measurements are attributed in the text; the final row describes a proposed evaluation, not Digital Applied test results.
EvidenceWhat it supportsWhat it does not establish
Google Terminal-Bench resultA vendor-reported improvement on that testYour repository’s success rate
Artificial Analysis indexA separate broad capability comparisonPercentage correctness on your tasks
Your acceptance setA decision about a defined workloadA universal model ranking

06 — Context limitsA large window still needs a useful brief

The model specification gives Hydrogen a unified 1,048,576-token context window shared by input and output, including thinking. It lists 65,536 tokens for default synchronous output, with longer generation through long decode. Inputs include text, images, audio, video and PDFs; output is text. Capacity and output modality are separate properties.

A large window can let an agent inspect more material, but it does not decide which material is relevant. A hypothetical policy-review task might include several versions of a document. If the prompt never identifies the effective version, more context can make the ambiguity worse. Label the source, date and authority of each document before asking the model to resolve a business question.

Likewise, accepting image input does not mean the model generates product images. Choose the model according to the result the application needs, not the broad word multimodal. For long jobs, define intermediate deliverables that can be checked and resumed. A huge requested output is difficult to review and expensive to repeat when a small late-stage error invalidates the result.

Capacity is not a brief

Supply the right source version, a concrete question and a clear output contract. A million-token limit does not remove the need to resolve contradictory instructions or missing facts.

07 — IntegrationMake tool behavior part of the evaluation

An agent is more than a model call. Its instructions, tools, permissions and recovery behavior determine whether it can turn a useful answer into a valid result. Test an empty search result, a failed command and an ambiguous document. The desired behavior may be to stop and explain what is missing. A model that continues confidently through a failed prerequisite is unsuitable for that workflow even if its prose reads well.

Start with read-only access or a disposable environment for the first evaluation. For a coding pilot, preserve the original working state and inspect the diff. For a document pilot, retain source references and distinguish an extracted fact from a recommendation. These are proposed safeguards for a useful trial, not claims that Digital Applied has run a Hydrogen deployment study.

The agent evaluation guide provides broader context for measuring behavior across complete runs. Keep the pilot small enough that a person can review every failure. Adding more tasks before understanding the first failures can produce a large spreadsheet without a clear decision.

Failure worth inspecting

A tool reports that a file could not be opened. Does the agent identify the missing evidence, seek a permitted alternative or invent the contents? That behavior matters more than the confidence of the final paragraph.

08 — AdoptionPromote the workload that earns its place

Use Hydrogen as a candidate default, then make exceptions explicit. A team might accept it for routine extraction and bounded code changes while retaining another model or a human for difficult architecture decisions. The Argon pricing and capability guide helps frame that comparison without requiring one model to win every category.

Before expanding, write down what the pilot established: which tasks passed, which failed, what the accepted tasks cost and how much review remained. Include a rollback route and a named owner for configuration changes. If nobody can explain why the model was chosen, a later price change or quality regression becomes much harder to assess.

For businesses building this into a production process, AI transformation work should connect the model choice to a measurable operational outcome. A useful first milestone is a repeatable task completed with acceptable accuracy, cost and oversight. Buying access to a new model is only one input to that outcome.

Decision rule

Expand only the task categories that meet your written acceptance criteria. Keep unresolved categories on the existing route until there is evidence to change them.

Your next step

Choose a setting, then prove one useful task

Hydrogen’s pricing and capability make it worth evaluating for everyday agent work. Start with a bounded workload, set thinking deliberately and count all billed output, failed runs and review effort. Public benchmarks can justify the shortlist; a repeatable acceptance test should justify the deployment.

Digital Applied

Turn a model trial into useful work

Define a practical AI workflow, its acceptance criteria and the evidence needed to put it into production.

Clear task scopeCost and quality checksPractical rollout
Start with one decision

Start with one recurring task

  • →Identify the current effort
  • →Define an acceptable result
  • →Compare complete runs
Common questions

Questions, answered.

Google’s standard API rate is $1.50 per million input tokens and $7.50 per million output tokens, including thinking. That is a token price, so the cost of an agent task also depends on repeated calls, retries, tools and any infrastructure used to execute the work.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

Google Search

See more Digital Applied analysis in your Google results by adding us as a preferred source.

Add as a preferred source