Gemini 4 Hydrogen: What It Is and Who Should Use It
Gemini 4 Hydrogen is Google's fast, cheap Gemini 4 model, out November 12. Specs, price, benchmarks, where it falls short, and who should switch.

Google just shipped the Gemini 4 model most people will actually use.
It is not the smartest one. It is the one that finishes the job before you check your phone.
Gemini 4 Hydrogen went generally available on November 12, 2026. It is the third model in the Gemini 4 family after Argon and Carbon, it costs $1.50 per million input tokens and $7.50 per million output tokens, and unlike its bigger siblings it is open on Google's unpaid developer tier, to Gemini app users through auto model selection, and to everyone who uses Google Search.
Gemini 4 Hydrogen is the right default for everyday coding and agent work, and the wrong choice for your hardest problems.
That is the whole decision. The rest of this post is the evidence, the price math, the places Google's own numbers need a squint, and a test you can run on your own work before you rewire anything.
What Gemini 4 Hydrogen is, in plain terms
Google calls Hydrogen its "daily driver for agentic workflows." In the launch post from Google DeepMind, Koray Kavukcuoglu draws a clean line. Gemini 4 Carbon and Gemini 4 Argon are the frontier models for "deep, long-horizon reasoning on your hardest problems." Hydrogen is "built for rapid execution across diverse agentic workflows."
Think of it as the model that does the work after the expensive model makes the plan.
If you read the Gemini 4 Argon post, you already know the family. Argon was the first Gemini 4 model, announced September 30 and opened to the public on October 13. Carbon followed as the strongest of the three, and it reached developers on November 10. Hydrogen is the speed and price tier, the successor to the Flash line, and Google's migration guide names it the recommended upgrade from Gemini 3.8 Flash.
Here is what it is on paper, from Google's own documentation.
| Spec | Gemini 4 Hydrogen |
|---|---|
| Model ID | gemini-4-hydrogen |
| Release date | November 12, 2026 (generally available) |
| Context window | 1,048,576 tokens, shared across input, thinking and output |
| Output | 65,536 tokens per request, up to about 1M with background execution |
| Input types | Text, images, audio, video, PDFs and Office files |
| Output type | Text |
| Thinking levels | low, medium, high (default), xhigh |
| Price, paid tier | $1.50 per million input, $7.50 per million output |
| Cached input | $0.075 per million with implicit caching |
| Unpaid tier | Yes, rate limited |
| Knowledge cutoff | March 2026 |
Two details matter more than they look. The context window is one shared pool, so a long prompt leaves less room for thinking and answer. And the price is flat across the whole million tokens, with no surcharge for long prompts. You can read the full spec sheet on Google Cloud's model page, and the developer notes in the Gemini API guide for Hydrogen.
The one number Google leads with
Google's headline claim is that Hydrogen beats Argon on the Artificial Analysis Intelligence Index. At its default high thinking level it scores 51 against Argon's 50, and it reaches 54 at the new xhigh setting.
That sounds backwards. A cheap model beating the bigger model it sits under.
It is less strange than it sounds once you read what the index measures. Artificial Analysis averages a basket of evaluations, and a basket can favor a model that is tuned for tool use and instruction following over one tuned for long, deep reasoning. Argon and Carbon still exist for the tasks where a single point on an average hides a gap in the hard cases.
So take the number for what it is. Hydrogen is not smarter than Argon in every way. On a broad average of tests it is now level with it or a point ahead, at a lower price and a faster pace, and that is a real shift in who needs the expensive tier.
One caution before the benchmark table. Almost every figure below comes from Google's launch post and docs. Google ran or chose those evaluations. Independent numbers from outside labs tend to arrive in the days after a launch, and the first week is when you should hold vendor charts a little loosely.
What changed from Gemini 3.8 Flash
The fairest comparison is the model Hydrogen replaces. Google gives you the deltas directly.
| Measure | Gemini 3.8 Flash | Gemini 4 Hydrogen |
|---|---|---|
| Terminal-Bench 4.1 | 18.2% | 57% |
| Speed at default setting | Baseline | About 2.5 times faster |
| Fabricated claims | Baseline | More than 5 times fewer |
| Default thinking level | medium | high |
| Thinking levels | 3 | 4, adding xhigh |
| Explicit context caching | Supported | Not supported, implicit only |
| Sampling controls (temperature, top p, top k) | Supported | Ignored |
The jump on Terminal-Bench is the big one. Terminal-Bench measures how well a model works through tasks in a command line, which is a fair stand in for the messy, multi-step coding work agents do. Going from 18.2% to 57% is not a tune up. It is a different class of tool for that job.
Google's second claim is the quieter one, and it is the one I would test first. Hydrogen cuts fabricated claims by more than five times against 3.8 Flash, which Google attributes to the model flagging missing context instead of guessing. If that holds on your data, it matters more than any benchmark. A fast model that invents file names and function signatures wastes the time it saves.
Then there are the changes that will break something if you ignore them. Google's migration checklist is blunt about them.
- Hydrogen rejects the
minimalthinking level with an error. - It ignores any
temperature,top_pandtop_kyou send. Temperature is fixed at 1.0. - It does not support explicit context caching, so remove
cached_contentfrom requests. Implicit caching runs on its own for prompts of at least 4,096 tokens and discounts cached input by 95%. - It rewards declaring your tools natively. Asking it in a system prompt to print raw JSON tool calls with no declared tools can return a
MALFORMED_FUNCTION_CALLerror. - A low hard coded output cap like 8,192 will truncate answers, because thinking tokens count against it.
If you run an app on Gemini today, that list is your afternoon.
Four thinking levels, and when each one pays
Thinking level is the main dial on this model, and Google changed the default from medium to high. That alone can change your bill and your latency, so it deserves a table.

| Level | What it costs you | Where it belongs |
|---|---|---|
| low | Fewest tokens, fastest | Chat, classification, high volume processing |
| medium | Fewer tokens than high | Interactive coding, live prototyping, tool calls where speed shows |
| high (default) | Balanced | Most coding, agents and multimodal document work |
| xhigh | Most tokens, slowest | Hard multi-step problems where high came up short |
Google's own warning on xhigh is worth repeating. It can overthink simple prompts, and it raises both latency and token use. Reserve it for the problem that failed at high. Do not make it your default because the benchmark at 54 looked good.
Here is the practical move. Build a small set of real tasks, run them at medium, high and xhigh, and pick the lowest level that clears your quality bar. Google recommends exactly this, with a high output token limit so the cap does not cut answers short. Output tokens include thinking tokens, and you pay $7.50 per million of them, so a higher level costs money on every reply.
How Hydrogen compares with Carbon, Argon and the rest
The family is easy to mix up, so here is the version Google's docs give you, with the use each one is built for.
| Model | Built for | Price per million tokens (in / out) |
|---|---|---|
| Gemini 4 Carbon | The hardest architecture, debugging, science and security work | Higher, see Google's pricing page |
| Gemini 4 Argon | Most advanced long-horizon tasks with fewer tokens than Carbon | Introductory $2 / $10, then $4 / $20 after November 30 per Google |
| Gemini 4 Hydrogen | Everyday coding, agents, multimodal documents, parallel subagents | $1.50 / $7.50 |
| Gemini 3.5 Flash-Lite | Classification, routing, sub-second latency | Lowest in the line |
Google's guidance is to start with Hydrogen for everyday work and move up only when a task needs deeper reasoning. Keep Flash-Lite for the high volume plumbing where milliseconds matter more than judgment. The Argon pricing change is covered in the Gemini 4 Argon post and the Gemini pricing guide, which tracks the dates when each number moves.
The interesting part is that Google no longer wants you to pick just one.
The pitch is a planner and an executor. Carbon (or Argon) reads the codebase, resolves ambiguity and writes the plan. Hydrogen then edits files, updates packages and runs tests. If Hydrogen hits friction, it hands the problem back up. That is the idea behind Orchestrator, a new option in Google Antigravity that pairs the two inside one session.

Google's Antigravity team reports that on Terminal-Bench, Orchestrator reached near-Carbon task success while consuming 24% less quota and finishing 19% faster. Two small flags on that claim. The Antigravity post names the benchmark Terminal-Bench 4.1, and the main announcement names it Terminal-Bench 4.0 for the same 24% and 19% figures. And the numbers are Google's, from Google's agent, on Google's chosen test. They are a direction, not a guarantee for your repo.
The idea behind it, though, is sound, and you can use it without Antigravity. Spend the expensive model on the planning and the verification. Spend the cheap one on the typing.
Where Hydrogen is strong, according to the evidence so far
Google lists a set of results that are worth reading one by one, because they point to what the model is actually good at.
On agent work, Hydrogen ranks fifth on Agent Arena, arena.ai's live benchmark for multi-step tasks like deep research and codebase navigation. On domain reasoning, Google reports 58.8% on HLE-Diamond, first place on the Vals Tax Agent Bench at 52.3%, a tie for first on the Vals Legal Research Bench at 55.8%, and a lead in its class on AutomationBench-AA. On multimodal knowledge work, it scores 53.4% on Google's new Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1.
Look at the pattern. Tax, legal research, automation, documents with charts and tables. This is paperwork with tools attached.
That tracks with the design. Hydrogen batches file reads and searches in parallel instead of opening one file at a time, and Google says early partners saw it run 2 to 2.8 times faster than 3.8 Flash while using 12% to 23% fewer tokens on complex graph, SQL and data heavy workflows.
If your day is spreadsheets, PDFs and a long list of small tool calls, this is the model aimed at you.
There is also a safety story, and it is more concrete than usual. Google says Hydrogen reduces overreach in agentic tasks by up to 81%, is up to twice as unlikely to bypass permission boundaries when a task is blocked, and is more than 2.3 times more resilient to indirect prompt injection on the Gray Swan benchmark. Its internal agent honesty score is 96.7%. These are internal and vendor reported figures. Still, an agent that stays inside its permissions is the property you want in a model that holds the keys to your files.
Where Hydrogen will let you down
You asked for a verdict, so here is the other half of it.
Hydrogen is a speed and efficiency model. Google says so itself, and it keeps Carbon and Argon around for "deep architectural judgment when designing systems or untangling subtle bugs." Expect it to trail them on the hardest reasoning, the sprawling refactor and the obscure bug three layers down.
The output is text only. It reads images, audio, video and documents, but it does not generate images or audio the way a separate model would.
It has no explicit context caching, no tuning and no RAG Engine support on Google Cloud, and it does not support the Live API. If your product leans on any of those, the migration is not a one line change.
And it is a short-term release. Google's documentation says Hydrogen is a short-term generally available release with a minimum 45 day retirement notice. Read that as a promise to warn you, not a promise to keep the model forever. Pin your model ID, build an eval you can rerun, and plan for a successor.
Some readers will push back here. A model that is this cheap and this fast beats the expensive ones on a published index, so why not use it for everything?
Because an average hides the tail. You do not feel the average. You feel the one task out of twenty where the cheap model confidently takes the wrong path, and you feel it more when an agent acts on it. Use Hydrogen as the default and escalate on failure. That is a better rule than either extreme.
Who should switch today
The answer depends on what you do all week, so here it is by job.
| If you are | Do this |
|---|---|
| A developer on Gemini 3.8 Flash | Move to Hydrogen. Google calls it the recommended upgrade and the gain on agentic coding is large. Fix the migration list first. |
| Running agents with many small tool calls | Test it as your worker model. Pair it with a stronger model for planning and checking. |
| Processing PDFs, spreadsheets, charts and recordings | Try it. Multimodal knowledge work is where Google's numbers are strongest. |
| Handling hard architecture, security or research problems | Stay on a frontier model for those, and use Hydrogen for the surrounding execution. |
| Using Gemini casually in the app | Do nothing. If auto model selection is on, Hydrogen will already answer the prompts that suit it. |
| Running high volume classification at tiny cost | Keep Flash-Lite. Hydrogen is cheaper than the frontier models, not cheaper than the cheapest. |
If you pay for more than one assistant, you also have the cross-vendor question. Hydrogen's price of $7.50 per million output tokens puts it in the same conversation as the low-cost tiers from OpenAI and Anthropic. The honest answer to "which is better" is that the winner changes with the task, and the only reliable way to know is to run your own work through each of them. Our Claude vs Gemini and Gemini vs ChatGPT comparisons give you the shape of that decision, and the habit of testing matters more than the model name.
How to try it today, and what to run in Magai while you wait
Start with where it lives. Google offers Hydrogen in AI Studio and the Gemini API on both the unpaid and paid tiers, in Android Studio on the paid tier, in the Gemini app, in Google Search, in Gemini for business, and in Antigravity for paid Google AI Pro and Ultra subscribers. If you want the raw model with the thinking levels exposed, the API and AI Studio are the way in.
Now the part about Magai, stated plainly. Gemini 4 Hydrogen is not in the Magai model picker today, and we will not tell you it is. The Gemini family is in Magai, along with Claude, GPT, Grok, DeepSeek and dozens more, so you can run the same prompt through several models in one thread without losing the conversation when you switch.
That is the test that settles a launch like this one. Here is the five task version, built to take about an hour.
- Pick five real tasks from your last two weeks. One coding fix, one spreadsheet or table cleanup, one long document summary, one messy email to answer and one thing you got wrong the first time.
- Paste the first task into one Magai thread. Add a Gemini model with the
/Geminipill, then read the reply. - Without leaving the thread, ask a Claude model and a GPT model the same thing with
/Claudeand/GPT. Each one sees the same history, so you are comparing answers and not setups. - Score each reply from one to five on whether you would send or ship it, and write down any invented fact. Fabricated details are the failure Google says Hydrogen reduces, so count them.
- Repeat for the other four tasks, then add up the scores.
When Hydrogen shows up in your API console, run the same five tasks through it at medium and high. You will have a baseline from the models you already use and a real number for whether the new one earns a place. Our guide to switching between AI models in one workflow walks through keeping context intact while you do it, and the best AI chatbot post has a scored version of the same five task idea if you want a template.
The price of doing it this way is one hour. The price of switching your stack on a press release is a month of cleanup.
Run the five tasks this week, keep the winner as your default, and let the next launch audition for the job instead of getting it.
More in Artificial Intelligence

Artificial Intelligence
Perplexity vs ChatGPT: Pick by How They Fail
Perplexity vs ChatGPT: one cites sources you can click, the other sounds sure even when wrong. How each fails, what each costs, and a 20 minute test.

Artificial Intelligence
DeepSeek vs ChatGPT: Price the Task, Not the Token
DeepSeek vs ChatGPT: one is a finished app, the other is cheap raw models. Capability gaps, real task costs, privacy, and a five prompt test.

Artificial Intelligence
MCP Client: What It Does and How to Pick One
An MCP client holds your connection to every MCP server. Here is what it does, the five kinds, and how to pick one for your work.