Google Rolls Out Gemini 4 Hydrogen in Search Worldwide, Calling It the Biggest Upgrade to Search Yet

Google has put a new artificial intelligence model called Gemini 4 Hydrogen into Google Search for every user in the world, and the company says it is the largest upgrade to its search engine yet. The change, announced on Thursday, November 12, 2026, replaces the fast model that powered Search answers until now, Gemini 3.8 Flash, with a quicker and more capable successor that Google says finishes tasks two and a half times faster. For the billions of people who type questions into Google every day, the change is meant to show up as better answers to long, messy and complicated questions, with more links to websites, more tables and more images inside the answers themselves. For the websites that depend on Google for visitors, it is another step in a shift that has been building for years, from a page of blue links toward answers written by the search engine itself. And for Google’s rivals in the race to build the best AI, it is the latest move in a contest where new models now arrive every few weeks and every launch is framed as the biggest yet. This is the story of what Gemini 4 Hydrogen is, what it changes in Search, what Google claims it can do, and what is still unproven, and why the Google Gemini 4 Hydrogen Search November 2026 rollout matters to everyone who searches, publishes or builds on the web.

1. The Announcement

1.1 A Thursday Rollout That Reached Everyone at Once

Google DeepMind launched Gemini 4 Hydrogen on Thursday, November 12, 2026, and the rollout began the same day. Unlike many AI launches, which start with a small group of paying subscribers and reach everyone else months later, Hydrogen went straight into Google Search for all users around the world. The company said the model was available globally in Search from the day of the announcement, with no plan, region or sign-up standing between ordinary searchers and the new system. That matters because Search is the product through which most people will meet the model. A new model inside a developer tool reaches developers. A new model inside Search reaches everyone, including people who have never heard of Gemini and never will. The decision to put Hydrogen in front of every Search user at once also tells its own story about how confident Google is in the model. A company does not hand its most-used product to a new system on day one unless it believes the system is ready, and it does not do so globally unless the infrastructure behind it can carry the load. The launch covered more than Search. On the same day, Hydrogen arrived in the Gemini app, in Google Sheets, in Google AI Studio, in the Gemini API for developers, in Gemini for business, and in Antigravity, Google’s coding platform, for paying AI Pro and Ultra subscribers. It was, in other words, a launch across the whole of Google’s AI estate at once, with Search as the headline. The timing also fits a pattern. Google has been releasing new Gemini models in quick succession through the autumn of 2026, and each one has arrived with a claim attached. Hydrogen’s claim is speed. It is built to execute tasks quickly, to carry out multi-step work on a user’s behalf, and to do so at a lower cost than the heavier models in the same family. That positioning, fast and cheap rather than slow and brilliant, is deliberate. Most of what people ask a search engine to do does not need the deepest reasoning a model can offer. It needs an answer that is right, fast, and easy to read. Hydrogen is Google’s bet that the model built for that job should be the one that powers Search.

The announcement itself came through Google’s own channels, including the Keyword blog and a post from Robby Stein, the vice president of product for Google Search. Stein’s post set out four changes that Hydrogen brings to the answers people see in Search, and each of them is about the quality of the answer rather than the speed of the model. The first is better handling of complicated questions. The second is a technique Google calls query fan-out, which connects people to a broader range of helpful links. The third is more visual answers, with tables and images. The fourth is better answers across all supported languages in AI Mode, the conversational search experience where the change is clearest. Taken together, the four changes describe a search engine that is trying to do more of the work itself: to understand a question in full, to break it into parts, to gather the pieces from several places at once, and to present the result as a finished answer rather than a list of places to look. That is the direction Google has been moving in for several years, and Hydrogen is the engine it has chosen to carry it further.

1.2 “The Biggest Model Upgrade to Search Yet”

Robby Stein called Hydrogen “the biggest model upgrade to Search yet,” and the phrase is worth taking seriously, not because every launch claims to be the biggest, but because of what it signals about how Google sees its own search engine. Search is the company’s oldest and most important product. It is also the product most at risk from the rise of AI chatbots, which answer questions directly instead of sending people to websites. By calling this the biggest upgrade yet, Google is telling its users, its advertisers and its rivals that it does not see Search as a finished product being slowly replaced. It sees it as a product that can be rebuilt from the inside, one model at a time, until the answers it gives are as good as anything a chatbot can offer. The claim also sets a test. If Hydrogen really is the biggest upgrade to Search yet, then the difference should be visible to ordinary users within days, not months. A searcher asking a long question with several parts should get an answer that deals with every part. A searcher who types with mistakes should still be understood. A searcher comparing two services should see a table, with figures, and links to the sources behind them. Early checks on the day after the rollout suggest that is roughly what is happening. Tests run in London on Friday, November 13, found AI Mode answering multi-part questions in full, coping with misspelt and mixed-language questions, and often replying with a comparison table whose source links pointed to the companies’ own pages. Those are small, early observations, not a verdict, but they point in the direction Google promised. The phrase also carries a warning for publishers. Every time Google makes its own answers better, the share of searches that end without a click on a website grows a little larger. Google’s answer to that concern, built into the Hydrogen announcement, is that the new model links to more websites, not fewer, and that its query fan-out technique surfaces pages people might not otherwise have found. Whether that promise holds in practice is one of the open questions of this launch, and it will be answered not by Google’s blog post but by the traffic figures of the websites that live or die by Search.

There is also a competitive edge to the wording. Google’s rivals have spent the year launching models of their own, and the race is now measured in weeks, not years. In that race, the company that powers the world’s most-used search engine has an advantage no rival can match: it can put its newest model in front of billions of people overnight, without asking them to download anything, pay anything or change any habit. That is what Hydrogen’s rollout does. It turns an AI launch into a change in the daily experience of searching, and it does so at a scale that no chatbot launch can match. The question is whether the experience is good enough to change how people feel about Search itself. For years, the complaint about Google has been that its results are cluttered, commercial and hard to trust. A model that answers well, cites its sources and shows its working could begin to answer that complaint. A model that answers confidently and wrongly would make it worse. The first days of Hydrogen will be watched closely for both.

2. What Gemini 4 Hydrogen Is

2.1 The Fast Member of a New Family

Gemini 4 Hydrogen is the fast, execution-oriented member of Google’s new Gemini 4 family. It sits alongside two heavier models, Gemini 4 Carbon and Gemini 4 Argon, and its job is different from theirs. Carbon is the frontier model, built for the hardest reasoning tasks and the longest horizons. Argon is the heavy model that arrived first in AI Mode for paying subscribers. Hydrogen is the model built for everyday work: coding, agentic workflows in which an AI agent carries out multi-step tasks on a user’s behalf, and consumer tasks where speed and cost matter more than depth. It is adapted from Carbon, which means it inherits much of the frontier model’s capability, but it is tuned and priced for rapid execution rather than for the longest, hardest problems. That trade-off is the whole point. In the world of AI models, there has always been a tension between how clever a model is and how fast and cheap it is to run. The cleverest models are slow and expensive. The fastest models are often too simple for serious work. Hydrogen is Google’s attempt to close that gap, to build a model that is fast enough to run inside Search for every user on earth, and clever enough to handle the coding and multi-step tasks that used to need a heavier model. The numbers Google published suggest it has made real progress on that front. On the Artificial Analysis Intelligence Index, a widely watched benchmark, Hydrogen scores 51 at its default thinking level, one point ahead of Argon at 50, and 54 at its highest setting. A fast model that outscores a heavier model on a general intelligence index is a striking result, and it is the number Google has put at the centre of its case for Hydrogen. The result comes with a caveat that applies to every benchmark a company publishes about its own model: the figures are Google’s own, they have not been independently reproduced, and benchmarks measure what their makers choose to measure. But the direction is clear. Google is no longer treating speed as a reason to accept a weaker model. It is treating speed as a property a strong model can have.

The name itself is part of the message. Google has named the Gemini 4 models after elements, and Hydrogen, the lightest element, is the lightest model in the family. The naming is not just branding. It tells developers and users at a glance where each model sits. Carbon is the heavy, frontier core. Argon is the stable, heavy-duty option. Hydrogen is the light, fast one that gets everywhere quickly. That clarity matters because the Gemini family is growing crowded, with several models, several tiers and several surfaces, and users need a simple way to know which one they are getting. In Search, they will not choose at all. Hydrogen is simply the model that answers, free of charge, for everyone. In the Gemini app and the API, developers and power users can pick the model that fits the job, and Hydrogen’s price, $1.50 per million input tokens and $7.50 per million output tokens, makes it the economical choice for high-volume work. The flat pricing across its full 1-million-token context window is also part of the pitch. A million tokens is roughly 750,000 words, enough to hold a whole codebase, a long document set or a lengthy conversation in memory at once, and charging the same rate whatever the length removes a source of surprise in the bill. For a model whose job is to do everyday work quickly and cheaply, that is the right shape.

2.2 Four Thinking Levels, From Low to Extra-High

Hydrogen offers four thinking levels: low, medium, high and extra-high, known as xhigh. The thinking level controls how much reasoning the model does before it answers. At low, it answers quickly with little deliberation. At high, the default, it thinks more before responding. At xhigh, a new top level, it thinks hardest of all. The idea is that not every question deserves the same amount of thought. A simple question should get a quick answer. A hard question should get a careful one. By letting the model, or the developer using it, choose the level, Google is trying to spend its computing power where it matters and save it where it does not. The benchmark scores track the thinking levels closely. At the default high setting, Hydrogen scores 51 on the Artificial Analysis Intelligence Index. At xhigh, it reaches 54. That three-point gain is the reward for thinking longer, and it is also a measure of the cost. Every extra point of intelligence comes with extra time and extra computing, and the existence of four levels is an admission that the right balance depends on the job. For Search, where answers must arrive in a fraction of a second, the default high level is the setting that matters most, and the fact that it already outscores Argon is the result Google wants people to notice. The xhigh level is there for the harder tasks, the coding problems and agentic workflows where a user is willing to wait a little longer for a better answer.

The thinking levels also say something about where AI models are heading. The first generation of chatbots gave one answer at one speed. The new generation adjusts how hard it thinks, and the adjustment is becoming a visible part of the product. Developers using the API can set the level for each call. Users of the Gemini app will see the benefit without ever touching a setting, as the model spends more effort on the questions that need it. For Search, the levels are invisible entirely. The searcher asks a question and gets an answer, and the machinery behind it is Google’s problem, not theirs. That invisibility is the point of putting Hydrogen in Search. The model is not a product the user chooses. It is the engine inside a product the user already has. The knowledge cutoff for Hydrogen is March 2026, which means the model’s training data runs to that date. Anything that happened after March 2026 is not in its training, and for questions about recent events the model must rely on what Search itself can find. That is a normal limitation for a model of this kind, and it is one reason why putting it inside Search, where fresh information is always at hand, makes sense. A model that can look things up does not need to have memorised them.

3. What Changes in Search

3.1 Harder Questions, Answered in Full

The first change Google promises is that Search will handle complicated questions better. That sounds modest, but it is the change most people will notice first, because complicated questions are the ones where today’s search engines most often fail. A simple question, like the capital of a country or the score of a match, has a simple answer, and any search engine can give it. A complicated question has several parts, and each part needs its own answer. It might ask for a comparison between two services, with prices, conditions and a recommendation for a particular kind of user. It might be written in a hurry, with typos, half-finished words and a meaning that is clear to a human but hard for a machine. It might be very specific, asking for something that only a handful of pages on the web actually contain. Google says Hydrogen is built for exactly these cases. Its post says Search now handles “questions with multiple parts, hard-to-understand typos, or highly specific requests,” and that long, conversational questions get an answer to each part. The early evidence supports the claim in outline. Tests run on Friday, November 13, in London found AI Mode handling misspelt and mixed-language questions well, answering every part of a multi-part question, and often replying with a table. One test asked for a comparison of two payment services, and the answer came back as a table whose source links pointed to the companies’ own pricing pages, with figures that matched what those pages said. That is a small sample, and one test is not a verdict, but it is the kind of result Google’s announcement promised, and it is the kind of result that, if it holds up at scale, will change how people use Search. The deeper point is about what a search engine is for. The old search engine found documents. The new one answers questions. The difference is the difference between a librarian who points at a shelf and a librarian who reads the books for you. Hydrogen pushes Google further toward the second kind of librarian, and the quality of its answers will decide whether people trust it.

There is a risk in that shift, and it is worth stating plainly. An answer that deals with every part of a question is only useful if every part is right. A model that is confident and wrong about one part of a five-part answer has still given a wrong answer, and the user may not notice which part is wrong. Google’s answer to that risk is to show its sources, with links to the pages behind each claim, so that a user who doubts an answer can check it. That is a good design, and it is one of the reasons the query fan-out technique matters. But it depends on users actually clicking through, and on the sources being right in the first place. The early tests suggest the source links do point to the right pages, and that the figures in the answers match the figures on those pages. That is encouraging, but it is early. The real test will come over the coming weeks, as millions of people ask millions of questions, including questions about subjects where the web itself is confused, contradictory or wrong. A search engine that answers directly takes on a responsibility that a search engine that merely lists links never had. Hydrogen is Google’s bet that it is ready for that responsibility, and the coming weeks will show whether the bet was right.

3.2 Query Fan-Out and More Links

The second change is a technique Google calls query fan-out, and it is the most important idea in the Hydrogen announcement for anyone who publishes on the web. When a person asks a complicated question, Hydrogen does not run a single search. It breaks the question into subtopics and runs several searches at the same time, one for each part, then gathers the results and builds a single answer from them. That is query fan-out. The result, Google says, is that Search connects people to “a broader range of helpful links than before,” including websites they might not otherwise have found. The idea is simple, and the effect, if it works, is significant. A single search for a complicated question tends to return the pages that match the question as a whole, which are usually the big, general pages. Several searches, one per subtopic, return the specific pages that answer each part, which are often smaller, more specialised pages that would never have ranked for the whole question. That means more kinds of websites can be cited inside a single answer, and more publishers get a chance to be seen. It also means the answer itself is built from more sources, which should make it more reliable, since a claim that appears in several independent places is more likely to be right than a claim that appears in one. The early tests again support the outline of the claim. The comparison-table answers produced on November 13 carried source links to the companies’ own pages, and the figures matched those pages. The answers were not built from a single source. They were built from several, and the links showed it.

The third and fourth changes follow from the first two. Google says Search will now give more visual answers, with tables and images, and better answers across all supported languages in AI Mode. The tables matter because they are the natural shape for a comparison, and a comparison is the natural shape for many of the hardest questions people ask. A table that sets out prices, features and conditions side by side is easier to read than a paragraph that tries to do the same job, and it is easier to check, because each cell can carry its own source. The images matter for the same reason. A question about what something looks like, or how a thing is put together, is answered better with a picture than with words. The language improvement matters because Search is a global product, and the quality of its answers has always been better in English than in other languages. Google says Hydrogen improves language understanding across all supported AI Mode languages, which, if true, would narrow a gap that has disadvantaged non-English speakers for years. Taken together, the four changes describe a search engine that is trying to be more helpful, more transparent and more even-handed than the one it replaces. The test, as always, is not the announcement but the experience, and the experience will be measured in the weeks ahead by the people who use Search every day.

4. The Numbers Google Cites

4.1 Two and a Half Times Faster, Says Google

The headline number in Google’s announcement is speed. The company says Hydrogen completes tasks two and a half times faster than Gemini 3.8 Flash, the model it replaces. That is a large claim, and it is the claim on which much of the launch rests. If a model is two and a half times faster, then a search engine built on it can answer more questions, in less time, at lower cost, and a developer building on it can do more work for the same money. Speed is also the property that matters most in Search, where a delay of a fraction of a second is felt by every user, and where the model must answer billions of questions a day without slowing down. But the claim needs to be read carefully, because it is Google’s own figure, it has not been tested independently, and at least one independent check has already cast doubt on it. A check by orcarouter.ai, published on November 13, 2026, found that the two-and-a-half-times speed claim does not survive the same outside test setup, while confirming the index score. That does not prove Google’s figure is wrong. It proves that speed depends on how you measure it, on what hardware you use, on what tasks you choose, and on what you count as finished. A model can be faster on Google’s infrastructure and slower on someone else’s, faster at one kind of task and slower at another, faster when measured one way and slower when measured another. The honest way to report the claim is to attribute it, which is what this article does. Google says Hydrogen is two and a half times faster than Gemini 3.8 Flash. An independent check found otherwise in its own setup. The truth, as it usually is with such claims, will emerge over time, as more people run the model on more tasks and publish what they find. What is not in doubt is that speed is the property Google chose to lead with, and that it chose to lead with it because speed is what Hydrogen is for. A model built for everyday execution has to be fast, or it has no reason to exist. The question is how fast, and the answer to that question is still open.

The speed claim also matters for a reason that has nothing to do with benchmarks. Search is the most demanding deployment in the world for an AI model. It must answer questions from billions of people, in dozens of languages, at all hours, without delay, and it must do so at a cost that does not eat the business. A model that is slow or expensive cannot live inside Search, however clever it is. A model that is fast and cheap can. By putting Hydrogen in Search for every user on the first day, Google is making a practical claim as well as a marketing one: that this model is fast enough and cheap enough to carry the world’s biggest search engine. That is a claim that cannot be faked for long. If the model were too slow, Search would slow down, and users would notice within hours. If it were too expensive, the cost would show up in Google’s accounts within a quarter. The rollout itself is therefore a kind of evidence, even if it is not the kind that fits in a benchmark table. It shows that Google believes the model can do the job, and it puts that belief to the test in public, every second of every day. The independent checks will follow, and they will matter, but the largest test of Hydrogen’s speed is already running, and it is called Google Search.

4.2 Benchmarks: 51, 54 and 57 Percent

The other numbers in the announcement are benchmarks, and they deserve the same careful reading. On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at its default high thinking level, one point ahead of Argon at 50, and 54 at its highest xhigh setting. The index is a composite measure, built from several tests, and it is widely watched in the AI industry as a rough guide to how capable a model is. A fast model scoring 51, ahead of a heavier model at 50, is a result that cuts against the usual assumption that speed and capability are a trade-off. It suggests that Google has found a way to make a light model clever, rather than merely making a clever model light. The three-point gain from high to xhigh is also informative. It shows that thinking longer helps, and it puts a number on how much it helps. Three points on this index is a meaningful gain, and it is the reward a user gets for waiting a little longer. But the numbers are Google’s own, and they have not been independently reproduced. The independent check that questioned the speed claim did confirm the index score, which is a point in Google’s favour, but one confirmation is not a consensus. Benchmarks also measure what their makers choose to measure, and a model can be tuned to do well on a particular index without being better in general. That is not an accusation. It is a fact about how benchmarks work, and it is why wise readers treat them as one piece of evidence among several, not as a verdict. The Terminal-Bench figures follow the same pattern. On Terminal-Bench 4.1, a test of how well a model completes real coding tasks in a terminal, Hydrogen scores up to 57 percent, against 18.2 percent for Gemini 3.8 Flash. That is a very large jump, more than three times the score of the model it replaces, and it is the number that will matter most to developers. Coding in a terminal is the kind of multi-step, tool-using work that Hydrogen is built for, and a jump from 18 to 57 percent suggests a model that can genuinely do the job rather than merely attempt it. But again, the figures are Google’s, and they have not been independently reproduced. The pattern across all of the numbers is the same. Google has published an impressive set of figures for its new model, an independent check has confirmed one of them and questioned another, and the rest await verification. That is the normal state of affairs on the day after an AI launch, and it is the reason this article reports the figures as Google’s claims, not as established facts.

The pricing numbers are different in kind, because they are not claims about performance but facts about cost, and they can be checked directly. Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1-million-token context window. Input tokens are the words and code a user sends to the model. Output tokens are the words and code it sends back. The flat rate means the price does not rise as the conversation or the document gets longer, up to the million-token limit. A million tokens is roughly 750,000 words, which is enough to hold a large codebase, a set of long documents or a very long conversation in the model’s memory at once. For developers, that combination, low prices and a long memory, is attractive, because it means they can build applications that handle large amounts of material without worrying that the bill will jump when the material grows. For Google, the pricing is a statement of intent. It is saying that Hydrogen is a model for high-volume, everyday work, and that it wants developers to use it a lot. The price is also a clue to the economics of the model itself. A company can only charge $1.50 per million input tokens if the model is cheap to run, and cheap to run is the same property that makes it fast enough for Search. The numbers, taken together, describe a single idea: a model that is good enough, fast enough and cheap enough to be used everywhere, by everyone, all the time. Whether the performance numbers hold up is still to be seen. The price is already a fact.

5. Where Else Hydrogen Lands

5.1 The Gemini App, Sheets and the API

Search is the headline, but Hydrogen’s rollout reaches well beyond it. On the same day, the model arrived in the Gemini app, in Google Sheets, in Google AI Studio, in the Gemini API and in Gemini for business. Each of those surfaces gives the model a different job, and together they show how Google sees Hydrogen fitting into the daily work of its users. In the Gemini app, it becomes the model that answers when a person opens the app and asks a question, writes a draft or works through a problem. It is free to use there, as it is in Search, which means the fastest model in the new family is also the one that costs ordinary users nothing. In Google Sheets, it becomes a helper inside the spreadsheet, able to work with tables, numbers and formulas, and to carry out the multi-step tasks that spreadsheets so often require. A spreadsheet is a natural home for a model built for execution, because spreadsheet work is execution: clean the data, run the calculation, build the table, check the result. A model that can do those steps quickly and reliably is worth more in a spreadsheet than a model that can write a poem about them. In Google AI Studio and the Gemini API, Hydrogen becomes a tool for developers, who can build it into their own applications, set its thinking level for each call, and pay the published price for what they use. The API is where the model’s speed and price matter most directly, because developers measure both in their own systems and will quickly discover whether Google’s claims hold up in practice. The arrival in all of these places at once is a deliberate choice. Google is not testing Hydrogen in one product and rolling it out to others later. It is putting it everywhere at once, and trusting that the model is ready. That confidence is itself a signal, and it is one that developers and users will read alongside the benchmarks.

The business version, Gemini for business, deserves a mention of its own, because it is where the model’s speed and cost meet the needs of companies that use AI at scale. A business that runs thousands of AI tasks a day cares about two things above all: how much each task costs, and how long it takes. A model that is two and a half times faster and priced at $1.50 per million input tokens answers both questions at once, if the claims hold. It also matters that Hydrogen is adapted from Carbon, the frontier model, because businesses want the reassurance that the cheap, fast model they are using is not a cut-down toy but a close relative of the best model Google can build. The knowledge cutoff of March 2026 is a limitation businesses will need to keep in mind. A model whose training ends in March 2026 does not know what happened after that date, and for business questions about recent events, it will need to look things up rather than recall them. In Search, that is no problem, because Search is itself a way of looking things up. In the API and the business product, it is a constraint that developers will need to design around, by giving the model access to fresh information when the task requires it. That is normal for models of this kind, and it is not a flaw. It is a fact about how they are built, and it is one of the reasons the model’s arrival in Search, where fresh information is always available, is the most natural of all its new homes.

5.2 Antigravity for Pro and Ultra Subscribers

Hydrogen is also coming to Antigravity, Google’s coding platform, for paid AI Pro and Ultra subscribers. Antigravity is the environment where developers do their most serious work, building software with an AI agent that can carry out multi-step tasks, write code, run tests and fix what breaks. It is, in other words, exactly the kind of agentic workflow that Hydrogen is built for, and its arrival there is a test that matters. Developers who pay for Pro or Ultra are the most demanding users Google has. They will notice quickly if the model is slower than promised, if it makes mistakes in code, or if its answers are worse than the heavier models it sits beside. They will also notice quickly if it is good, because a fast model that writes good code and fixes its own mistakes is the most valuable tool a developer can have. The Terminal-Bench figures, up to 57 percent against 18.2 percent for Gemini 3.8 Flash, are aimed squarely at this audience. Terminal-Bench tests how well a model completes real coding tasks in a terminal, which is close to what developers actually do, and a threefold improvement on that test, if it holds up, would be felt immediately in daily work. The pricing also matters here. Developers who run many tasks through an API watch the bill closely, and a model that costs $1.50 per million input tokens lets them run far more tasks for the same money. The combination of speed, capability and price is the pitch, and Antigravity is where it will be judged by the people best placed to judge it.

The tiering of the rollout also tells a story about Google’s strategy. Hydrogen is the first of the three Gemini 4 models to reach every Search user free of charge. Argon reached AI Mode for paying Pro and Ultra subscribers on October 13, 2026. Carbon started for Ultra subscribers on November 10. Hydrogen, the fast one, is the one that goes to everyone. That ordering is deliberate. It says that Google’s priority is to put its best everyday model in front of the most people, and to keep its heaviest models for the subscribers who pay for them. It also says something about how Google thinks the AI race will be won. Not by keeping the best model behind a paywall, but by putting a very good model in the product everyone already uses. The paying tiers still get something extra: Carbon for the hardest work, Argon in AI Mode, Hydrogen in Antigravity with the tools that professionals need. But the free tier is no longer a cut-down version of the real thing. It is the real thing, running the model that Google believes is good enough for the hardest everyday jobs. That is a change from the way AI launches used to work, and it is a change that puts pressure on every rival that still keeps its best model behind a subscription.

6. The Family Around It

6.1 Carbon and Argon: The Heavier Models

To understand Hydrogen, it helps to understand the two models it sits beside. Gemini 4 Carbon is the frontier model of the family, the heaviest and most capable, built for deep, long-horizon reasoning, the kind of work where a model must think hard for a long time before it answers. It started reaching Ultra subscribers on November 10, 2026, two days before Hydrogen’s launch, and it is the model from which Hydrogen is adapted. That relationship matters. Hydrogen is not a separate, simpler model built from scratch. It is a version of the frontier model, tuned for speed. That means it inherits much of Carbon’s capability, including the reasoning that makes Carbon good at hard problems, while giving up some depth in exchange for speed. The trade is the right one for everyday work, where most questions do not need the deepest reasoning, but it is worth knowing that the trade has been made. Gemini 4 Argon is the other heavy model. It reached AI Mode for paying AI Pro and Ultra subscribers on October 13, 2026, a month before Hydrogen, and it is the model that Hydrogen now outscores on the Artificial Analysis Intelligence Index, 51 to 50 at Hydrogen’s default setting. That result is worth pausing on. A fast model outscoring a heavy model on a general index is not what the industry expected, and it raises a question about what the heavier models are for. If Hydrogen can do the work of Argon, faster and cheaper, then Argon’s role must be in the tasks where its extra weight still matters, the longest reasoning, the hardest problems, the work where thinking longer is worth the wait. Google has not said much about that boundary, and it is one of the things developers will work out for themselves in the coming weeks, by trying both models on the same tasks and seeing where each one wins.

The family structure also reflects a change in how Google builds models. The old approach was to build one model and use it everywhere. The new approach is to build a family, with different members for different jobs, and to let each one do what it is best at. That is a more complicated way to build, but it is a better way to serve users, because the needs of a search engine, a coding platform and a research tool are genuinely different. A search engine needs speed. A coding platform needs accuracy in code. A research tool needs depth. No single model is best at all three, and a family can cover them all. The naming, after elements, helps users keep track. Hydrogen is light and fast. Carbon is the heavy core. Argon is stable and strong. The names are a promise that each model knows its job, and the launch of Hydrogen is the moment when the lightest member of the family takes on the biggest job of all, powering Search for everyone. Whether the family works as a whole will depend on whether users can tell the members apart in practice, and whether the right model reaches the right task. In Search, the user does not choose. In the app and the API, they do. The coming weeks will show whether Google has made the choice easy enough.

6.2 Goodbye to Gemini 3.8 Flash

Hydrogen replaces Gemini 3.8 Flash, the fast model that powered Search answers until now, and the replacement is a measure of how quickly this field moves. Gemini 3.8 Flash was itself a recent model, and its replacement after a short life is a reminder that in AI, no model stays current for long. The figures Google published make the scale of the change clear. On Terminal-Bench 4.1, Hydrogen scores up to 57 percent, against 18.2 percent for 3.8 Flash. On the speed claim, Hydrogen is said to be two and a half times faster. On the intelligence index, it scores 51 at its default setting, a level that 3.8 Flash never approached. If those figures hold up, the replacement is not a small step. It is a leap, and it is the kind of leap that makes the old model look outdated overnight. That is the nature of progress in this field. Each new model makes the last one look slow, and each launch resets what users expect. The people who will feel the change most are the ones who use Search every day, because the model behind their answers has changed without them doing anything. They will not see a new name or a new button. They will see better answers, or they will not, and that is the only test that matters. Google has not said when Gemini 3.8 Flash will be removed from the API, and that silence is worth noting. Developers who built applications on 3.8 Flash need to know when it will go away, and when they must move to Hydrogen. The company has said that Hydrogen is free in Search and the Gemini app, and that it is available in the API, but it has not given a date for the old model’s retirement. That is a gap in the announcement, and it is one of the questions developers will be asking in the days ahead.

The replacement also carries a lesson about the pace of change. When a model is replaced this quickly, and this completely, it is a sign that the field is still in its early, fast-moving phase, where each generation brings a large improvement over the last. That phase will not last forever. At some point, the improvements will get smaller, the models will settle, and the competition will move from capability to cost, reliability and trust. Hydrogen’s pricing, its speed and its arrival in Search for free suggest that Google is already preparing for that phase, by building a model that is cheap enough and fast enough to be used everywhere. But the large jump in the benchmark figures suggests that the fast-moving phase is not over yet. A model that scores 57 percent where its predecessor scored 18 percent is not an incremental improvement. It is a step change, and step changes are what keep this field exciting and unsettling in equal measure. For users, the practical advice is simple. The model behind Search has changed, and the answers will change with it. The best response is to use Search as before, to check the answers that matter, and to judge the new model by what it does, not by what its launch announcement says.

7. What It Means for Websites and Publishers

7.1 More Links, More Competition

For the websites that depend on Google for visitors, Hydrogen is both a promise and a threat, and the two need to be weighed together. The promise is in the query fan-out technique. By breaking a question into parts and searching for each part, Hydrogen connects people to a broader range of helpful links than before, including pages they might not otherwise have found. That is good news for smaller, specialised websites, which have always struggled to rank for broad questions but may rank well for the specific subtopic that answers one part of a question. It is also good news for the web as a whole, because an answer built from many sources is better than an answer built from one, and because a search engine that cites more pages is a search engine that still sends people to the web. The threat is the one that has been growing for years. Every time Google’s own answers get better, fewer people need to click through to a website at all. A question answered in full, with a table and a summary, is a question that does not produce a visit. For publishers whose business depends on visits, that is a direct loss, and it is a loss that no number of extra links inside the answer can fully offset, because a link inside an answer is not the same as a visit. The balance between the promise and the threat will be decided by the numbers, and the numbers will take weeks to emerge. Publishers should watch their traffic from Search in the weeks after November 12, and mark that date in their reporting, because it is the day the model behind their traffic changed. They should also watch which of their pages get cited in AI answers, and which do not, because citation is becoming a new kind of visibility, alongside ranking, and it may matter as much.

The change also rewards a particular kind of publishing. If Hydrogen’s query fan-out looks for the best page for each part of a question, then the pages that win will be the ones that answer a specific question well, with clear facts, clear figures and a clear source. That is a reason to write pages that are precise, that state their numbers plainly, and that make it easy for a machine to tell what they say. It is also a reason to be careful about accuracy, because a page that is cited in an AI answer is a page whose claims are repeated to millions of people, and an error in that page becomes an error in the answer. The publishers who do best in this new world will be the ones who treat accuracy as their main asset, because accuracy is what makes a page worth citing. The publishers who do worst will be the ones who chased clicks with vague, sensational or thin content, because a model that reads the page before it cites it has no use for that kind of writing. That is a harsh judgement, but it is also an opportunity. A small website that writes carefully and accurately about a subject it knows well may now be cited where it would never have ranked, and that citation may be worth more than a ranking ever was. The Hydrogen launch does not settle any of this. It opens the question, and the answer will come from the traffic figures of the months ahead.

7.2 What Site Owners Should Watch

There are practical steps that site owners can take this week, and they are worth setting out plainly. The first is to run their customers’ real questions in Google Search, on Thursday, November 12, and Friday, November 13, and to look at what comes back. Does the answer deal with every part of the question? Does it cite their site, or a competitor’s? Are the figures right? The second is to mark November 12, 2026, in their traffic reporting, so that any change in visits from Search can be traced to the day the model changed. The third is to check their pages for the things a model needs: clear headings, stated facts, figures that match their sources, and answers to the specific questions their customers actually ask. The fourth is to watch for the arrival of Hydrogen’s effects in the languages they serve, since Google promises better answers across all supported AI Mode languages, and the improvement may arrive unevenly. None of these steps requires a new tool or a new subscription. They require attention, and a willingness to treat the search engine as something that has changed, rather than something that stays the same. The announcement itself gives site owners no new settings, menus or controls. There is nothing to switch on and nothing to switch off. The change is in the model, and the model is Google’s. The work for site owners is to observe, to measure and to adapt, and to do so quickly, because the model will not wait for them.

There is also a larger question that site owners, and everyone else who cares about the web, should keep in mind. A search engine that answers questions directly is a search engine that stands between the reader and the publisher, and the terms of that relationship matter. Google says Hydrogen links to more websites, and that its query fan-out surfaces pages people might not otherwise have found. If that is true, and if it stays true, then the new search engine is a better neighbour to publishers than the old one. But the promise is easy to make and hard to verify, and it will be tested by the traffic figures of the coming months. The honest position, for now, is to note the promise, to watch the figures, and to judge the model by what it does. That is the same advice that applies to the speed claim and the benchmark figures. Google has made a set of claims about its new model. Some have been confirmed, one has been questioned, and the rest are waiting. The coming weeks will provide the evidence, and the evidence will come not from Google’s blog but from the experience of the people who use Search, build on the API and publish on the web.

8. The Bottom Line

8.1 What Is Confirmed and What Is Not

What is confirmed is this. Google DeepMind launched Gemini 4 Hydrogen on Thursday, November 12, 2026. It is the fast, execution-oriented member of the Gemini 4 family, adapted from the frontier Carbon model, and it is rolling out globally in Google Search for all users, in the Gemini app, in Google Sheets, in Google AI Studio, in the Gemini API, in Gemini for business, and in Antigravity for paid AI Pro and Ultra subscribers. It replaces Gemini 3.8 Flash. Robby Stein, Google’s vice president of product for Search, called it the biggest model upgrade to Search yet. The model has four thinking levels, low, medium, high and extra-high, with high as the default, and a knowledge cutoff of March 2026. Its price is $1.50 per million input tokens and $7.50 per million output tokens, flat across a 1-million-token context window. It is free in Search and the Gemini app. The family around it includes Carbon, which started for Ultra subscribers on November 10, and Argon, which reached AI Mode for Pro and Ultra subscribers on October 13. Those are the facts, and they come from Google’s own announcement and from the reporting that followed it. What is not confirmed is the set of claims about performance. Google says Hydrogen completes tasks two and a half times faster than Gemini 3.8 Flash, scores 51 on the Artificial Analysis Intelligence Index at its default setting and 54 at xhigh, and reaches up to 57 percent on Terminal-Bench 4.1 against 18.2 percent for its predecessor. Those figures are Google’s own. They have not been independently reproduced. One independent check, by orcarouter.ai on November 13, confirmed the index score and found that the speed claim does not hold up outside Google’s own test setup. That is a mixed result, and it is the right place to leave the performance question for now: promising, partly checked, and not yet settled. Also unconfirmed are the practical questions that matter most to users and publishers. Which parts of Search run on Hydrogen and which do not. When Gemini 3.8 Flash will leave the API. How much traffic the new answers send to publishers. Whether the improvement in answers is the same in every language. Those questions have no answer yet, and they will be answered by experience, not by announcement.

The distinction between confirmed facts and unconfirmed claims matters, because it is the difference between reporting and repeating. A launch announcement is a company’s account of its own product, and it is the company’s job to present that product in the best light. It is a journalist’s job, and a reader’s job, to separate what the company has done from what the company says its product can do, and to hold the second to the test of evidence. That is what this article has tried to do. The launch is real. The rollout is real. The price is real. The performance figures are claims, and they are labelled as claims throughout. The early independent evidence is mixed, with one figure confirmed and one questioned, and the rest still waiting. That is not a reason to dismiss the launch. It is a reason to watch it closely, and to judge it by what happens next. The model is now in front of billions of people, and their experience will be the largest and most honest test it will ever face.

8.2 The Questions That Remain

The questions that remain are the ones that will decide whether Gemini 4 Hydrogen is remembered as a turning point or as another launch in a long line of launches. The first is whether the speed claim holds up outside Google’s tests. If it does, Hydrogen will change what developers can build and what Search can answer. If it does not, the launch will be remembered for a number that did not survive contact with the outside world. The second is whether the benchmark scores, 51 at the default setting and 54 at the highest, translate into answers that are actually better, in the hands of ordinary users asking ordinary questions. Benchmarks measure what their makers choose to measure, and the real test is the daily experience of searching. The third is whether the promise to publishers holds: whether query fan-out really does link to more websites, and whether those links turn into visits, or whether the better answers simply keep more people on Google’s own pages. The fourth is when Gemini 3.8 Flash will leave the API, and how smooth the move to Hydrogen will be for the developers who built on it. The fifth is whether the heavier models, Carbon and Argon, keep a clear role now that the fast model outscores one of them, or whether the family will be reshuffled again in the next launch. And the last is the largest of all. Search is the front door of the web, and a change to the model behind it is a change to the way billions of people find things out. If Hydrogen makes Search better, more accurate and more transparent, then the web becomes a better place to learn. If it makes Search more confident and less checkable, then the web becomes a worse one. The coming days will provide the answers.

Comments

comments