Your AI bill shows tokens. It hides the cost of rework

Your agent is supposed to find and fix an error in your online shop. While you wait for its answer, another job is already running in the background. Then the bill arrives, and the number that matters most is missing: what did the completed task actually cost? Gemini 4 Hydrogen is designed to handle coding and repeat agent tasks quickly and at a relatively low cost. In one benchmark, it edges ahead of Argon. But a low token price and a strong test score do not tell you whether you will save money in day-to-day work. So what should you measure?
A low price meets an expensive question
For a small business, the task might be fixing a bug in a booking system. In a larger organisation, it could be running recurring checks across many internal processes. Either way, what matters is whether the job gets done reliably.
According to the published API rates, Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. Tokens are small pieces of text that a model processes or generates. These prices make Hydrogen easier to budget for, but they do not tell you how many attempts, corrections or staff minutes it takes to get a usable result.
That is the difference between a cheap call and cheap work. A model can cost little per amount of text and still become expensive if it gets tasks wrong or needs frequent corrections. What do the early benchmarks tell us?
What the score of 51 really tells you
In the Artificial Analysis Intelligence Index v5.0.1, Hydrogen scores 51 in high mode. Argon scores 50, while Carbon scores 60. For businesses, that is an interesting signal: the smaller model narrowly beats Argon in this test, but remains behind Carbon.
The index combines nine evaluations. According to the source, it does not directly measure a model’s productivity or safety in a business setting. A score shows how a model performs on defined tests. It does not tell you whether it will categorise your invoice data correctly, make the right change to your code or handle a sensitive request appropriately.
Another comparison also needs context. Google reports a score of 57 per cent for Hydrogen on Terminal-Bench 4.1. AI4Business reports 63.9 per cent for Claude Sonnet 5.5 on the same benchmark. These results come from fixed test sets. They are not a guarantee of how either model will perform on your work. You will only know how closely your day-to-day tasks match those tests by running your own trial.
Speed only counts when the job is done
A benchmark can take a long time while your team waits for a change to go live. Artificial Analysis measured an average of around 8.3 minutes for Hydrogen in high mode and 13.5 minutes in xhigh mode. Carbon took 27.4 minutes in high mode, while Argon took 30 minutes.
That sounds like a clear time advantage. But the measured time covers the test task, not the whole process from request to approval in your business. If a member of staff has to check the suggestion first, undo a wrong change or reword the request, that time moves into another column. It does not disappear.
For a business with only a few development tasks, waiting for a good answer may matter more than small differences in token prices. A company with many similar tasks will also want to know how often Hydrogen gets it right first time. The important measurement starts after the benchmark.
The cost includes the model’s thinking
A price per million tokens looks clear and straightforward. But the cost of an AI task depends on how much text goes in and how much comes back out. Hydrogen charges separately for input and output. According to AI4Business, output tokens also include tokens used for internal reasoning, the intermediate steps involved in solving a task.
That can make a difference when a job requires more thought. The xhigh mode does not automatically cost more per token, but a task may use more output tokens. Longer conversations also increase the amount of context supplied, meaning the information the model receives to do the job.
The source gives this example: ten million input tokens and one million output tokens cost $22.50 at the published rates. This excludes cache, external tools and any other possible fees. It is a useful starting point for budgeting, but it is not yet the cost of a completed task.
What this means for your business
Keep the first test small enough that a mistake will not take anyone by surprise. Choose a recurring task where a person would check the result anyway. For an online retailer, that might be a clearly scoped code fix. A professional services firm could look at recurring changes to internal scripts.
Decide what counts as success before the test starts. A task is only complete when the result is technically right and does not create an unreasonable amount of rework. Record four things for every attempt:
- Input and output costs, including the model mode used.
- Time to a checked result, not just the time to the first answer.
- Errors, retries and the minutes spent on human corrections.
- The share of tasks Hydrogen handles without switching to a more capable model.
Choosing a model changes the working day too
A cheaper model can ease pressure on a team if it takes on recurring execution work. AI4Business describes Hydrogen as a model for coding, tool calls and AI agent tasks. Carbon and Argon are intended for more demanding reasoning tasks, according to the same source. That division of work makes sense as an idea, but it is not a ready-made rule for your business.
Your team needs to know when it can trust a result and when a person or another model should take over. Otherwise, the work that was meant to ease the load becomes extra checking for staff. Decide who will assess errors and who is responsible for changes.
According to AI4Business, Hydrogen may only be available through a global endpoint in some Gemini Enterprise configurations, and administrator approval may be required. On Google Cloud, the model is classed as a short-lived version, with at least 45 days’ notice planned before it is retired. Before making a process dependent on it, IT and the relevant business team need to plan together for access and possible changes.
Give next Monday a clear measure
For a limited trial, select a handful of real, anonymised tasks and run each one with Hydrogen and your current solution. Assess the results against the same criteria. For code, it is not enough for a suggestion to look plausible. It needs to make the change you asked for without creating more work.
Google reports 24 per cent lower quota use and 19 per cent faster task completion for a combination using its Antigravity orchestrator system, compared with its own reference configuration. That is a vendor claim about this setup, not a general result for businesses. Treat it as a reason to run a test, not a forecast of your bill.
By next Monday, your team should be able to see which tasks Hydrogen can handle and when a person needs to step in. If that is not clear from a simple checklist, you do not yet have the basis for a sound cost decision. Only then can you judge whether the lower price also brings less work.
Frequently asked questions
How much does Gemini 4 Hydrogen cost?
AI4Business lists the price as $1.50 per million input tokens and $7.50 per million output tokens. According to the source, output tokens also include those used for reasoning, or internal steps. The actual bill depends on how many tokens a task uses and whether there are additional costs for cache, external tools or other services.
Is Gemini 4 Hydrogen better than Gemini 4 Argon?
It depends on what you measure. In the Artificial Analysis Intelligence Index v5.0.1, Hydrogen scores 51 in high mode and Argon scores 50. The index combines nine evaluations and, according to the source, does not directly measure productivity or safety in a business setting. A trial using your own tasks will tell you more than the narrow score difference.
How fast is Gemini 4 Hydrogen?
Artificial Analysis measured an average of around 8.3 minutes for Hydrogen in high mode and 13.5 minutes in xhigh mode. These are benchmark results for specific tasks. They do not predict how long a job will take in your business once it has been checked and approved.
How do I calculate the cost of an AI model per task?
Record the input and output tokens used for each task, then apply the relevant rates. Add the cost of retries, external tools and human review. This gives you a cost per checked, successfully completed task, rather than just a token price.
Will Gemini 4 Hydrogen remain available to businesses?
AI4Business describes Hydrogen on Google Cloud as generally available but classed as a short-lived version. According to the source, at least 45 days’ notice is planned before it is retired. Some Gemini Enterprise configurations may require a global endpoint and administrator approval. Check these conditions before making it part of a business process.
Interested?
Want to know what this means for your company?
Sources
- Gemini 4 Hydrogen: Google accelera gli agenti AI. Ecco costi e benchmark - AI4Business · ai4business.it, 2026-11-13
- Google cree haber dado con el santo grial de la IA: Gemini 4 Hydrogen es rápido, bueno y (casi) barato · xataka.com, 2026-11-13
- Release: llm-gemini 0.37 · simonwillison.net, 2026-11-12
This article was researched and written with AI assistance and automatically checked against the sources listed before publication. More on our AI transparency
Free guide · PDF
The AI Compass for small and mid-sized businesses
9 jobs your team can hand over by Monday.
Get it free
Klarframe AI Readiness Check
How ready is your company for AI?
About ten minutes, instant result with your stage, a 90-day plan, suitable solutions and notes on the EU AI Act.


