Gemini 4 Hydrogen: Features, Speed, Pricing and How to Use It

Gemini 4 Hydrogen is Google's new fast, low-cost model for everyday coding and agentic work, announced on 12 November 2026. It sits beside Gemini 4 Carbon and Argon and takes over the role the Flash models used to play. This ProCraft.site guide explains what it does, how fast it is, what it costs, how it compares with rival models, and how to get the best results from it.
What Is Gemini 4 Hydrogen?
Gemini 4 Hydrogen is the newest member of Google's Gemini 4 model series. Google DeepMind announced it on 12 November 2026 and describes it as an everyday workhorse for agentic workflows: a model built for fast, cost-efficient tasks rather than the hardest research problems. Google's own pitch is cutting-edge intelligence for everyone, delivered with speed and efficiency for everyday coding and agent work.
Here is what we know from the launch announcement and early coverage:
Key Points
- Naming. Hydrogen replaces the Flash branding. It follows the Gemini 3.8 Flash launch on 2 September 2026 and arrives in the same week that Gemini 4 Carbon became available.
- Purpose. It is built for rapid execution. It can work alone, or act as the execution engine while Carbon or Argon plan and reason.
- Context. It accepts up to 1 million tokens of input and can produce up to 1 million tokens of output.
- Inputs. It works with text, PDFs, spreadsheets, charts, audio and video.
- Reasoning control. It offers four thinking levels: low, medium, high (the default) and a new xhigh level.
- Reach. It is rolling out across the Gemini app, Google Search, the Gemini API, Google AI Studio, Gemini for business and Google Antigravity.
One caution before we go further. Almost every performance number in this article comes from Google's own announcement or from press coverage of it, not from independent testing. Treat the figures as a strong starting point and run your own tests before you commit a product or a budget to any model.
Gemini 4 Carbon, Argon and Hydrogen: How the Family Fits Together
The Gemini 4 series now has three main models, and the names tell you the job each one does. Carbon and Argon are the frontier models for deep, long-horizon reasoning. Hydrogen is the fast, light model for rapid execution.
Gemini 4 Carbon is best for the hardest problems, complex software engineering and multimodal knowledge work. It costs $2 per million input tokens and $10 per million output tokens through 30 November 2026, then $4 and $20. It was announced on 28 October 2026 and became available from 10 November. For a deep dive on the flagship model, see our Gemini 4 Carbon guide.
Gemini 4 Argon is best for most software engineering and knowledge work, with faster replies and lower token use. Its price is the same as Carbon. It launched on 13 October 2026.
Gemini 4 Hydrogen is best for everyday coding and agentic workflows, high volume and low latency. It costs $1.50 per million input tokens and $7.50 per million output tokens, flat across the full context window. It was announced on 12 November 2026.
Carbon and Argon details come from Google's model pages as summarized by FourWeekMBA. Hydrogen pricing comes from Google's launch announcement.
The Gemini app shows the same split in its model picker. Quick answers use 3.5 Flash-Lite, Everyday help uses Hydrogen, Detailed research uses 3.1 Pro for AI Pro subscribers, Code and analysis uses Argon, and Most complex work uses Carbon.
The design idea is teamwork rather than competition. In Google Antigravity, a feature called Orchestrator can switch between Carbon and Hydrogen inside a single session. Carbon can coordinate the plan while Hydrogen powers the worker and verifier agents that do the repetitive steps. You get most of the quality of a frontier model while spending far fewer tokens.
Key Features of Gemini 4 Hydrogen
Google's launch message centers on one promise: it is easier to get things done. The model follows instructions more closely, calls the right tools and reaches a reliable result with less back-and-forth. These are the features behind that promise.
Key Points
- Four thinking levels. Low, medium, high and xhigh let you trade speed for depth. High is the default. Google reports that the xhigh level lifts the model's score on the Artificial Analysis Intelligence Index from 51 to 54.
- Better instruction following. Hydrogen sticks to the brief, which matters for long agent runs where small deviations compound.
- Smarter tool calling. It batches file reads and searches in parallel instead of opening files one at a time, and it adapts when a tool call fails or a search returns nothing.
- Fewer made-up claims. Google says it cuts fabricated claims by more than five times compared with Gemini 3.8 Flash.
- Multimodal input. PDFs, spreadsheets, charts, audio and video can all go in the same request.
- Large context. One million tokens of input fits long documents, big codebases and months of notes in a single prompt.
- Connected apps. In the Gemini app it can pull context from Gmail, Google Drive, Airtable and Dropbox.
- Surgical coding edits. For debugging, it makes focused changes and avoids rewriting unrelated files.
Together, these features point to a model designed less for impressive one-off answers and more for dependable repeated work: the kind of tasks you run hundreds of times a day inside tools and pipelines.
Speed and Efficiency: How Fast Is Hydrogen?
Speed is the headline claim. Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash when both run at their default high setting. Faster answers matter most in loops, where an agent calls a model dozens of times to finish one job.
The launch post adds several more specific figures:
Key Points
- Enterprise data work. Early partners saw runs that were 2 to 2.8 times faster, using 12% to 23% fewer tokens than 3.8 Flash, on graph, SQL and data-heavy workflows.
- Antigravity Orchestrator. On Terminal-Bench 4.0, the Carbon-plus-Hydrogen setup used 24% less quota and finished tasks 19% faster on average, while coming close to Carbon's success rate.
- Google Sheets. Outputs in the Sheets canvas arrive more than 30% faster than with earlier models.
Efficiency has two sides: time and tokens. A model that answers quickly but needs several retries is not really fast. Google's claim is that Hydrogen is quick and also lands closer to the right answer on the first attempt, which is why it reports fewer tokens as well as faster runs.
For teams, the practical effect is lower latency for users and a lower bill per completed task. If you run support bots, code assistants, data cleanup or report generation at volume, those two gains stack. All of these figures are company-reported, so measure your own latency and cost on real workloads before you rely on them.

Gemini 4 Hydrogen Benchmarks and Rival Comparisons
The benchmark picture is strong for a fast, low-cost model, but it is not a clean sweep. Hydrogen leads its class on several agent and knowledge tests, and it trails larger rival models on others.
Results reported by Google
Key Points
- Artificial Analysis Intelligence Index: 51 at high, 54 at xhigh (Argon: 50).
- Terminal-Bench 4.1: 57% (Gemini 3.8 Flash: 18.2%).
- HLE-Diamond: 58.8%.
- Vals Tax Agent Bench: 52.3%, ranked first.
- Vals Legal Research Bench: 55.8%, tied for first.
- Agents' Last Exam v1.1: 71.6%.
- Multimodal Knowledge Agent: 53.4%.
- Agent Arena: ranked fifth.
Comparison with rival models
The press outlet OfficeChai compared Hydrogen with other models using Google's own table. Only the figures stated in its text are shown here.
Key Points
- Terminal-Bench 4.1: Gemini 4 Hydrogen 57.0%, Claude Sonnet 5.5 63.9%.
- OSWorld-2.1 (offline subset): Gemini 4 Hydrogen 77.6%, Claude Sonnet 5.5 83.9%.
- Terminal-Bench-Science 0.1: Gemini 4 Hydrogen 51.4%, Claude Sonnet 5.5 59.9%.
Google also reports that Hydrogen beats GPT-6.1 Terra on eight of the nine benchmarks where Terra has a reported score. On SWE-sweep, a test of finding and fixing bugs without breaking existing code, Hydrogen scores 13.5% and is described as well ahead of the other models compared. On Vending-Bench 2, a simulated business test, it earns $14,549. On HLE-Diamond, Gemini 4 Argon scores 69.1% against Hydrogen's 58.8%.
How to read this. Hydrogen wins where speed, cost and agent behavior matter, and larger models still lead on some terminal, science and computer-use tasks. The right choice depends on your workload, so test two or three models on your own tasks.
Agentic Workflows: Where Hydrogen Is Designed to Shine
An agentic workflow is a task where the model plans steps, uses tools, checks its own work and keeps going until the job is done. Hydrogen is built for exactly this pattern, and Google's launch focuses on three strengths.
Execution at speed. Agents often spend most of their time on routine steps such as reading files, running searches, calling APIs and filling tables. A fast, inexpensive model is ideal for these steps. Google positions Hydrogen as the worker, while Carbon or Argon act as the planner when a task needs deeper thought.
Resilience. Real tools fail. Searches return nothing, files are missing and commands error out. Google says Hydrogen adapts when a tool call fails instead of repeating the same mistake, and that it calls the right tools and skills more often in the Gemini app.
Safer behavior. Google published several safety figures for agent use, all compared with Gemini 3.8 Flash:
Key Points
- Up to 81% less overreach, meaning doing more than the user asked.
- Up to 79% less under-service, meaning doing less than the user asked.
- Misaligned actions and task shortcutting reduced by 12 times across tool-use and coding evaluations.
- Up to 2 times less likely to bypass permission boundaries when a task is blocked or impossible.
- More than 2.3 times more resilient to indirect prompt injection on the Gray Swan benchmark.
- A 96.7% honesty rate on Google's internal agent honesty evaluation.
These are Google's internal results, and agent safety still depends on how you deploy the model. Give agents the least access they need, log every action and keep human approval on payments, deployments and anything hard to undo.
Coding With Gemini 4 Hydrogen
Coding is the first use case Google names, and the model's design choices reflect real developer pain. Most developers do not want an assistant that rewrites a whole project to fix one bug. They want a precise change that passes the tests.
What Google says Hydrogen does well
Key Points
- Makes focused, surgical edits for debugging and issue resolution without touching unrelated files.
- Outperforms similar-tier models on SWE-sweep, which tests whether a model can find and fix bugs without breaking existing code.
- Scores 57% on Terminal-Bench 4.1, up from 18.2% for Gemini 3.8 Flash, which shows a large jump in command-line and terminal skills.
- Reads and searches several files in parallel, which shortens the exploration phase on large repositories.
- Recovers when a command fails or a search finds nothing.
Where you can use it
Developers can reach Hydrogen through the Gemini API and Google AI Studio, in the paid tier of Android Studio, and in Google Antigravity for paid Google AI Pro and Ultra subscribers. In Antigravity, the Orchestrator can pair it with Carbon so that the stronger model plans and Hydrogen carries out the routine edits and checks.
Good habits for coding prompts
Specify the goal, the files in scope, the tests that must pass and what the model must not change. Writing a precise specification is largely a prompting skill, and the ProCraft.site prompt engineering guide is a good place to practice it. Always run your tests and review the diff before you merge, no matter how good the benchmark numbers look.

Everyday Use: Gemini App, Search and Sheets
Hydrogen is not only for developers. Google is putting it in front of ordinary users in three places.
The Gemini app. Hydrogen appears in the model picker as Everyday help. Free users get it through automatic model selection when it fits the prompt, so you may be using it without choosing it. Google says it follows instructions more closely and needs less back-and-forth, and it can use Gmail, Google Drive, Airtable and Dropbox for context.
Google Search. Hydrogen is rolling out to all users globally in Search. Google says it handles multipart questions and related-query expansion better and gives more visual responses. For example, a question that mixes comparison, budget and location can be answered in one pass instead of three searches.
Google Sheets. In Gemini in Google Sheets, outputs arrive more than 30% faster. The rollout starts with Google AI Pro and Ultra subscribers and eligible business subscriptions.
Everyday tasks that suit a fast model
Key Points
- Summarizing a long PDF or meeting recording.
- Cleaning and reshaping a spreadsheet.
- Drafting replies from an email thread.
- Comparing products or plans from several pages.
- Turning rough notes into a structured outline.
- Answering quick technical questions with a short code sample.
For these jobs, speed and low cost matter more than winning a hard benchmark, which is exactly the space Hydrogen targets.
Gemini 4 Hydrogen Pricing and Availability
Hydrogen is priced to be the cheap option in the Gemini 4 family. Google lists the Gemini API price at $1.50 per million input tokens and $7.50 per million output tokens, and the price stays flat across the full 1 million-token context window. Many models charge more once a prompt gets long, so a flat rate makes long-document jobs easier to budget.
Prices at a glance
Key Points
- Gemini 4 Hydrogen: $1.50 per million input tokens, $7.50 per million output tokens.
- Gemini 4 Carbon and Argon (through 30 November 2026): $2 per million input tokens, $10 per million output tokens.
- Gemini 4 Carbon and Argon (from 1 December 2026): $4 per million input tokens, $20 per million output tokens.
- Claude Sonnet 5.5 (list price): $2 per million input tokens, $10 per million output tokens.
Carbon and Argon prices come from Google's model pages as summarized by FourWeekMBA. The Sonnet 5.5 list price is reported by OfficeChai. Google Cloud billing listings also show a cached-input rate of $0.075 and a flex-input rate of $0.75 per million tokens for Hydrogen. Those rows were reported before the public launch, so the final figures could differ.
A quick cost example. Imagine a job that sends 1 million input tokens and receives 200,000 output tokens. At Hydrogen's list prices that costs about $3.00. The same job costs about $4.00 on Sonnet 5.5 and about $8.00 on Carbon after 1 December. This is an illustration of list prices only. Real bills depend on caching, thinking level and retries.
Where you can get it
Key Points
- Gemini app. Hydrogen is in the model dropdown as Everyday help. Free users get it through automatic model selection when it fits the prompt.
- Google Search. It is rolling out to all users globally.
- Gemini API and Google AI Studio. Available to all tiers.
- Android Studio. Available in the paid tier.
- Gemini for business. Available through the API or the model dropdown, and in Gemini in Google Sheets for eligible business subscriptions.
- Google Antigravity. Available to paid Google AI Pro and Ultra subscribers.
The launch post does not mention Vertex AI, so check your cloud console before you plan an enterprise deployment around it. Access also varies by product and subscription, and the coverage does not name specific regions, so confirm what your own account can use.

Who Should Use Hydrogen, and When to Pick Carbon or Argon
Choosing a model is a cost and quality decision. A simple rule works well: start with the cheapest model that passes your tests, and move up only when it fails.
Key Points
- High-volume support, tagging, extraction or summaries: start with Hydrogen, because speed and price matter more than peak reasoning.
- Everyday coding, bug fixes and code review: start with Hydrogen, then Argon if it struggles, for surgical edits and strong terminal results at low cost.
- Long agent runs with many routine steps: use Hydrogen as worker and Carbon as planner, because Orchestrator-style setups cut tokens and time.
- Hard research or tough engineering problems: start with Argon or Carbon, for higher scores on the hardest reasoning tests.
- Quick questions in the Gemini app: use automatic selection, because Google routes the prompt to a fitting model.
Good fits for Hydrogen
Key Points
- Developers who run many small coding tasks and want low latency.
- Startups and small teams that watch their token bill closely.
- Operations teams that process documents, spreadsheets and emails in bulk.
- Product teams that build chat or agent features where every second of delay costs users.
When to move up
Choose Argon or Carbon when a wrong answer is expensive or the task needs many steps of careful reasoning. Argon scores 69.1% on HLE-Diamond against Hydrogen's 58.8%, which shows the gap on hard research questions. If a Hydrogen run fails your tests even at xhigh thinking, that is your signal to escalate.
If you are still comparing tools and vendors, the ProCraft.site AI directory is a handy place to browse AI products side by side before you commit.
Limitations and Things to Watch
A balanced review needs a section on what could go wrong. These are the main caveats we found.
Key Points
- The numbers are mostly Google's. The benchmark results come from Google's launch post and press coverage of it. Independent testing may show different results, especially on your own data. Coverage from Superpower Daily raises the same point: the key question is whether the gains hold beyond Google's benchmarks.
- It does not win everywhere. In the OfficeChai comparison, Claude Sonnet 5.5 scores higher on Terminal-Bench 4.1 (63.9% against 57.0%), OSWorld-2.1 (83.9% against 77.6%) and Terminal-Bench-Science 0.1 (59.9% against 51.4%). Argon also leads on HLE-Diamond.
- Index versions are not directly comparable. The 56 reported for Sonnet 5.5 on the Artificial Analysis Intelligence Index comes from an earlier index version, so it should not be set against Hydrogen's 51 and 54.
- Access is uneven. Some features depend on your plan. Antigravity and the Sheets rollout start with paid Google AI Pro and Ultra subscribers, and the launch post does not mention Vertex AI.
- Small inconsistencies in the announcement. The launch text refers to Gemini 3.8 Flash, while one image caption says 3.5 Flash. It also mentions both Terminal-Bench 4.0 and 4.1. These look like editing slips, but they are a reminder to read the model card for exact details.
- Prices can change. Early billing listings and final public prices may differ, and the Carbon and Argon prices double on 1 December 2026.
- Fewer errors is not zero errors. A fivefold cut in fabricated claims is impressive, but you should still verify facts, numbers and citations before anything is published or shipped.
A sensible testing plan. Collect 20 to 50 real tasks from your own work. Run them on Hydrogen at high thinking, then at xhigh, then on one larger model. Score the results for accuracy, speed and cost, and pick the cheapest option that meets your quality bar.
Prompt Tips for Gemini 4 Hydrogen
A fast model rewards clear instructions. Because Hydrogen follows instructions closely, vague prompts produce vague results, and precise prompts produce precise ones. These habits work well with any modern model, and they pay off even more with an agent that acts on your words.
1. State the goal and the finish line. Say what a good result looks like: format, length, audience and any must-have details. 2. Give context once, clearly. Paste the relevant text, files or data. With a 1 million-token window you rarely need to cut context, but you should label each part. 3. Set the boundaries. Tell the model what it may touch and what it must not change, especially in code and spreadsheets. 4. Choose the thinking level on purpose. Use low or medium for simple extraction and formatting, high for normal work, and xhigh when a task is tricky. 5. Ask for checks. Request that the model list its assumptions, run the tests or verify totals before it answers. 6. Break big jobs into steps. Let a stronger model plan and Hydrogen execute, or run the steps one after another yourself. 7. Show an example. One sample of the output you want beats a paragraph of description. 8. Review before you trust. Treat every result as a draft from a fast junior colleague.
A reusable template
Want ready-made starting points? Browse the ProCraft.site prompt library for tested prompts, then try and refine your own versions in the ProCraft.site AI prompt studio.
Frequently Asked Questions About Gemini 4 Hydrogen
What is Gemini 4 Hydrogen?
Gemini 4 Hydrogen is Google's fast, cost-efficient AI model in the Gemini 4 series, announced on 12 November 2026. It is built for everyday coding and agentic workflows and replaces the Flash naming.
How is Hydrogen different from Carbon and Argon?
Carbon and Argon are frontier models for deep, long-horizon reasoning. Hydrogen is built for rapid execution. It can work alone or act as the worker while Carbon or Argon plan the task.
How much does Gemini 4 Hydrogen cost?
The Gemini API price is $1.50 per million input tokens and $7.50 per million output tokens, flat across the 1 million-token context window. Prices could change, so check Google's pricing page before you budget.
Is Gemini 4 Hydrogen free?
Free Gemini app users can get it through automatic model selection when it fits the prompt, and Google AI Studio is available to all tiers. Heavier use through the API is billed per token, and Antigravity access needs a paid Google AI Pro or Ultra plan.
What are the thinking levels?
There are four: low, medium, high (the default) and xhigh. Higher levels think longer and score better on hard tasks, and lower levels are faster and cheaper.
How big is the context window?
Google lists a 1 million-token input window and up to 1 million output tokens.
Is Hydrogen better than Claude Sonnet 5.5 or GPT-6.1 Terra?
It depends on the task. Google reports that Hydrogen beats GPT-6.1 Terra on eight of nine benchmarks where Terra has a score, while Sonnet 5.5 scores higher on several terminal, computer-use and science tests. Hydrogen is cheaper at list price. Test on your own work.
Can Hydrogen read PDFs, images, audio and video?
Yes. Google says it handles PDFs, spreadsheets, charts, audio and video.
Is Gemini 4 Hydrogen good for coding?
Google says it makes focused, surgical edits and scores 57% on Terminal-Bench 4.1, up from 18.2% for Gemini 3.8 Flash. Always run tests and review changes before merging.
Where can I try it?
Open the Gemini app and look for Everyday help in the model picker, use Google Search, or try the model in Google AI Studio.
Conclusion: A Fast Everyday Model Worth Testing
Gemini 4 Hydrogen is Google's answer to a simple question: how do you make strong AI fast and cheap enough to use all day? It brings four thinking levels, a 1 million-token context window, better tool calling and a flat $1.50 and $7.50 price per million tokens. It also arrives in the places people already work: the Gemini app, Google Search, Sheets, AI Studio and Antigravity.
The reported results are strong for a model in its class, with a large jump in terminal skills, leading scores on several agent tests and a claimed 2.5 times speed gain over Gemini 3.8 Flash. At the same time, larger models still lead on some of the hardest benchmarks, and most of the numbers come from Google itself.
The smart move is to test, not to guess. Pick a handful of real tasks, run them through Hydrogen and one larger model, and compare accuracy, speed and cost. If Hydrogen passes, you save money and time every day. If it does not, you have a clear reason to use Argon or Carbon.
For more practical AI guides, prompt templates and tool comparisons, visit ProCraft.site.
Sources
Figures are from the reports below as of 13 November 2026. Most performance numbers are reported by Google and have not been independently verified.
Key Points
- Introducing Gemini 4 Hydrogen for Agentic Workflows (Google blog, 12 November 2026)
- Google announces Gemini 4 Hydrogen, rolling out to Gemini app (9to5Google)
- Google Releases Gemini 4 Hydrogen, Beats Sonnet 5.5 On Most Benchmarks (OfficeChai)
- Gemini 4 Models Explained: Carbon, Argon, Hydrogen Prices (FourWeekMBA)
- Google releases Gemini 4 Hydrogen for coding, Search and the Gemini app (Superpower Daily)
Explore More Tested Prompts & AI Workflows
Browse 6,800+ copy-ready AI prompts and 5,000+ curated AI tools on ProCraft.