Singapore

Agentic coding costs: what Gemini Hydrogen changes

As of November 2026, a business running agent workflows should read the launch of a fast coding model as a change in the cost of running routine steps: Gemini Hydrogen, released on Nov 12, 2026, is priced at $1.50 per million input tokens and $7.50 per million output tokens on the paid tier, and it takes four thinking levels from low to xhigh. That price and those settings are where an ai automation for business budget gets its shape. Which model writes the best code is the question a developer asks; which model runs the routine steps of a billed workflow is the one a business lead has to answer.

Key takeaways:

Input tokens cost $1.50 per million and output tokens $7.50 per million, a five-fold spread that lets a reading-heavy step stay cheap while a writing-heavy step sets your bill.
Hydrogen takes four thinking levels, from low to xhigh, and defaults to high. The setting decides how much reasoning the model spends before it answers, step by step.
Pairing Hydrogen with Gemini Carbon inside Google Antigravity, under the Orchestrator option, saves 24% of quota and finishes 19% faster than running Carbon alone, on the test Google published.
The delivery constraint sits in permissions, acceptance and handover. A cheaper model shortens some build steps; it does not answer who signs off on what the agent may do.

What the token price buys, and where the cost lands

Google's Gemini API documentation lists one price for the model and one context window: $1.50 per million input tokens, $7.50 per million output tokens, with a 1,048,576-token window shared by input, thinking and output tokens. Synchronous single responses cap at 65,536 tokens; longer decodes run to the full window.

For a marketing operations team, the practical reading is a ratio of five to one. A step that mainly reads (a long brief, a CRM export, a page of specifications) costs a fifth of what a step that mainly writes does at the same token count. Agent flows mix both. Retrieval and classification drag in input; drafting copy, code or reports pushes output.

Run the arithmetic on one flow before agreeing to a scope. Take the ratio alone: for a fixed token count, the input side costs a fifth of the output side. So a flow that spends its tokens on reading stays cheap per document, and the same flow rebuilt to spend them on writing can climb several times over without a single new feature. Ask a vendor which side of the ledger your workflow sits on, and how they measured it.

Output limits shape web application development more than they shape copy. A single synchronous call of 65,536 tokens is roomy for a page component or a migration script and tight for a whole-repository change, which is why long decode up to the full window exists. If a build step has to return a component plus its tests in one response, check that the pipeline handles the long path rather than quietly truncating.

A desk with a tall pile of read documents beside one freshly drafted page.

Four thinking levels, and what each one is for

Hydrogen runs at low, medium, high or xhigh, with high as the default. The setting is a spend control before it is a quality control. A classification step that picks a category from a short list rarely needs the highest setting, and paying for it on every ticket in a queue compounds into real money by quarter end. That is the setting where ai automation for business budgets usually go wrong first.

A split that holds for most internal tools:

Step type	Level to test first	Why
Tagging, routing, extraction	low	The answer space is small and the check is cheap
Drafting reports, mapping fields	medium	Needs reasoning over context, not a full plan
Debugging a failing test, refactoring	high	The default, and the one to compare against
Multi-file change with an unclear cause	xhigh	The step where a wrong plan costs more than the tokens

Treat this as a pilot starting point rather than a benchmark. The only way to know whether medium does the job is to run your own inputs at both settings and read the outputs side by side. Keep the sample, keep the score, and move a step up only when the failure is a reasoning failure and not a missing instruction in the prompt. Teams that skip the comparison usually default to the highest level for everything and pay for it on every run.

The benchmark numbers, read for what they are

The standalone model scores 51 on the Artificial Analysis Intelligence Index, reaches 54 at its highest setting, and scores 57% on Terminal-Bench. The Antigravity blog compares it with Gemini Flash at 18.2% on the same test.

That gap describes a model that finishes work in a terminal-driven loop, which is the shape of most agentic build tooling. It says nothing about your codebase, your tests or your deployment rules. A benchmark run in Google's test setup answers a question about that setup. When a partner quotes the figure in a proposal, ask what they ran on your repository before the quote, and how many of the agent's changes passed your review without edits. The same caution applies to web application development timelines built on a vendor's benchmark slide.

Why the planner and executor split matters more than the model

The Antigravity post, Gemini Hydrogen and Orchestrator, describes a pairing in which Hydrogen works beside Gemini Carbon under one model selector. On the benchmark named above, the pair reached near-Carbon task success rates while consuming 24% less quota on average and finishing 19% faster.

The mechanism is familiar to anyone who has priced an operations process: put the expensive reasoning where the decision is made, and a cheaper model where the work repeats. In an agent flow that looks like a frontier model planning the steps and a faster model executing each one, with a check between them.

For a Singapore team commissioning an agent, this changes the scope question. Ask for the flow diagram with model choice marked at every step, and ask what happens when the executor fails twice in a row. A design that routes a failure back to the planner with the error attached is cheaper to run than one that escalates every step. The quota saving came from a benchmark run in Google's own setup; your own split needs measuring rather than assuming. A quote for ai automation for business work that names only one model for the whole flow has not answered the question.

One colleague sketches a plan on glass while another follows the same steps at a checkpoint.

What Gemini Hydrogen changes for companies automating business workflows

Search results on agentic AI for business still explain the concept in general terms: agents perceive, plan and act with limited human direction. None of the pages a business lead finds on the first page of results says what a new model generation changes about the economics of running those agents. The answer sits in the pricing table rather than the launch copy, and it differs from the previous generation in ways a finance team will notice faster than a developer will.

The price per output token sets the cost of every step that produces a document, a piece of code or a report, and that is the largest line in most content-heavy flows. The thinking levels let a team spend reasoning budget per step instead of per flow, so the same workflow can run cheap on routine work and expensive only where a wrong answer is costly. A second, faster model designed to execute what a larger model plans turns the plan-execute split from an architecture diagram into a priced option.

What stays put is the work around the model. Authorisation for outbound messages and budget changes still needs a person. Acceptance criteria still decide whether the agent shipped. Handover documentation still determines whether your team can run the thing after the vendor leaves. A model release can halve the infrastructure line of an agent budget and leave the governance line untouched. Budgets built on the assumption that the model is the expensive part tend to over-weight the wrong column.

The decision for a reader here is not which model to pick. It is to scope the first flow around cost per token and a planner-executor split, with the heavier model planning and the cheaper one executing, and to leave the model comparison inside that scope rather than in front of it. For the automation case, the practical test is a pilot with two steps at two thinking levels, the token count written down, and a decision at the end of the month. One reporting flow, one costing sheet, one review meeting.

Where Topkee sits in this

The model is the replaceable layer. Topkee builds custom AI agents for enterprises aimed at marketing, e-commerce and operations teams, and it treats model choice as one setting among several: the platform notes that model choice, subscriptions, node resources and visibility determine the usable scope of a job template.

That framing is also the honest answer to a launch week full of benchmark charts. A faster model raises the ceiling of what one flow can do; it does not change which flow the business needs, which data the flow may read, or who signs off on its output. Teams that buy a model usually get a cheaper run. Teams that buy a delivered flow with review points get something they can operate after the invoice is paid, and the published scopes above are built for the second case.

Delivery is where a faster coding model shows up in a quote. Topkee's integration work handles system development and integration as one project covering the agent, the app, cloud integration and maintenance, and it builds and bills shared backends, interfaces and integration work once. Its workflow design follows four principles it publishes: sourced data, bounded execution, verifiable output and transferable delivery. Each delivery stage carries a named artefact. A function and interface list with a fixed quote opens the work, a usable build with sample outputs follows, and test records and deployment notes arrive at acceptance, with a monthly service record after launch.

Topkee publishes its package scopes at a few levels. The smallest covers one agent flow and a single standard system connection. The middle one adds linked agent flows, web pages and two system connections. The largest goes to several flows, up to a dozen mobile pages plus four backend pages, shared iOS and Android code, and two two-way integrations. Acceptance comes with a month of defect warranty. The quoted amounts are in the notes behind this article and none of them moves with a model release, because they price the delivery work rather than the tokens.

The part of Topkee's agent platform that a model release touches least is the approval gate. Outbound sending or budget changes need separate authorisation, and the agent raises a confirmation item when data is missing instead of filling the gap with a guess. A cheaper executor makes it affordable to run that gate on every step; it does not remove the gate.

A team member pauses at a desk to approve a pending outbound message before sending.

What to put in the next vendor conversation

Bring a short question set to the next scoping call, and bring your own numbers to it.

Ask for the token estimate per flow, split into input and output, with the thinking level named for each step. Ask how the flow routes a failure, and what the agent may do without a person approving it. Ask for the acceptance artefacts by name, and whether handover includes the interface list and the deployment notes. The ai automation for business page lists the package scopes behind those numbers.

A team with those answers can compare two quotes on the same basis. A team without them compares a project price against a guess.

For context on how agents change the way marketing teams measure outcomes, How Agentic AI Drives Measurable SEM Growth for SMBs covers the measurement side, and Why Your AI Advertising Pilots Fail to Scale covers the readiness checks that usually come first.

FAQ

Why a cheaper output price does not make custom agent projects cheaper

Only the infrastructure line. If a project's model usage sat at a few hundred dollars a month, a large cut in token price changes very little of a six-figure build. Delivery work, integration and acceptance testing dominate the budget.

Should every step of an agent flow use the highest thinking level?

No. Set the level per step and test it. Tagging and routing steps usually work at low or medium, and you can measure whether the top setting changes the answer.

Is the Terminal-Bench score a reason to choose this model?

It is a reason to trial it. The score shows the model can finish terminal-driven tasks in a test harness, and it does not speak to your repository, your tests or your review process.

What does the Orchestrator pairing save in practice?

On the Terminal-Bench run named above, the pair saved 24% of quota and finished 19% faster against Carbon alone. Your saving depends on how much of your flow is routine execution, and on whether the plan-execute split is designed before the build starts.

Do I need a new web application to use agentic automation?

Not always. Many flows run against an existing CRM or ERP through an interface. What usually needs building is the agent flow itself and the review step around it.