1. Context and Key Points

Google has unveiled Gemini 4 Hydrogen, the third model in the Gemini 4 family, in a post on its official blog The Keyword on 12 November 2026. It is not another frontier model: it is the family's fast model, built to execute — write code, run tests, verify changes, move agents — rather than reason for hours over a problem. Carbon and Argon remain Google's frontier models for deep, long-horizon reasoning; Hydrogen is the engine that does the repetitive work, fast and cheap.

The announcement comes just weeks after the arrival of Carbon and Argon, and it completes the map of the Gemini 4 family: two models to think deeply and one to execute at full speed. There is a second reading to the move. Hydrogen is the first model in the performance segment that Google sells under a codename instead of the "Flash" label. The Flash brand, which for years identified Google's fast and cheap models, disappears from the front: the fast model is now called Hydrogen, and that says a lot about how Google wants this generation to be perceived.

The figures Google publishes in the announcement are direct. Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default high setting. It offers a native 1-million-token input context and up to 1 million output tokens. It costs $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across the whole context window and no long-context surcharge. And on the Artificial Analysis Intelligence Index — an independent index — it scores 51 points at its default high level, one point above Argon (50), and 54 at its maximum level, the new xhigh. All of these figures are Google's claims as published in the announcement; the index is independent, but the values come via Google and have not yet been reproduced by outside labs.

Availability is immediate and broad: Hydrogen is in AI Studio, in the Gemini API, in Gemini for Business, in the Gemini app, in Google Search and in Antigravity for paid Pro and Ultra subscribers. It is the default model for AI Mode and the one that answers the hardest questions in AI Overviews. In other words, the Gemini 4 generation is no longer just for developers with an API key. From this launch onward, anyone searching on Google or opening the Gemini app will meet Hydrogen.

2. Technical Highlights

The figure Google puts in the headline is speed: Hydrogen completes tasks 2.5 times faster than Gemini 3.8 Flash at its default high setting. That is not a marginal improvement. In agentic workflows — where a model writes code, runs a test, reads the error, fixes it and tries again — the speed of each iteration decides whether the loop is usable or unbearable. A model that finishes every turn in less than half the time changes the whole economics of agentic work, because it multiplies the number of attempts that fit into the same time budget.

The second technical fact is context: 1 million tokens of native input and up to 1 million tokens of output. The symmetry matters. Many models boast huge input windows and then cap output at a fraction of that; Hydrogen can return answers as long as the context it receives. To refactor a whole repository, rewrite a long document or generate a complete report in a single pass, that one-million-token output is the difference between doing it in one go and doing it in pieces.

The third element is reasoning control. Hydrogen introduces four adjustable thinking levels — low, medium, high and xhigh — with high as the default. Users pay in time and tokens only for the depth they need: low for quick answers, xhigh for the problems that demand the maximum. It is at xhigh that Hydrogen reaches its 54 points on the Artificial Analysis Intelligence Index, against 51 at high and Argon's 50. The very existence of a fourth level, above the usual high, signals that Google wants this model to stretch when the task asks for it, without forcing a switch to a frontier model.

There is a lineage detail that Google puts in writing in the model's card: Hydrogen is adapted from Gemini 4 Carbon. It is not a small model trained separately, but the same base with different post-training and a different serving profile. That is the technical explanation for why a "fast" model can beat Argon, the frontier model, on an intelligence index: it shares Carbon's brain, but it is tuned and served to execute. The knowledge cutoff is March 2026, stated in the model's own card, with the warning that some domains are less up to date.

3. Sector Impact

The Gemini 4 family is divided by job, not by size. Carbon and Argon are the models for deep, long-horizon reasoning, the hardest problems; Hydrogen is the model for fast execution: writing code, running tests, verifying changes, calling tools and repeating the loop. Google says it plainly: teams can reserve a frontier model for architectural judgement and hand the repetitive execution to something faster and cheaper. Hydrogen can work alone, or run underneath Carbon and Argon as the engine.

That split has an immediate consequence for the rest of the industry. Until now, the conversation about models was ordered on a single scale — bigger and smarter against smaller and faster — and every lab competed for the same throne. Google breaks that scale into two axes: depth and speed. An execution model that beats a frontier model on an intelligence index forces the question of whether the frontier is still where it was, or whether the distinction between "thinking" and "doing" has become more important than the one between "big" and "small".

The second blow is one of branding. By selling its fast model under a codename — Hydrogen — instead of the Flash label, Google retires the word that had identified its fast range for years. The message is that this model is not a cut-down version of something else: it is a full member of the Gemini 4 family, with its own name and its own place. For rivals, the signal is clear: the performance segment is no longer the cheap step in the catalogue, but a category with a name of its own.

And there is a third fact worth not overlooking: Google has not retired Gemini 3.8 Flash. It stays in the catalogue at $0.75 per million input tokens. Hydrogen does not replace Flash by decree; it lives alongside it. The user who only needs the cheapest model possible still has it, and the one who needs execution speed and intelligence now has a new option. Google's fast range, instead of narrowing, widens.

The scope of the launch reinforces that reading. Hydrogen is the default model for AI Mode and the one that answers the hardest questions in AI Overviews. This is not a model for a niche of developers: it is the model millions of people will meet when searching on Google or opening the Gemini app. When Google puts its new model as the default in Search, it is saying it trusts it for its most visible surface. And in Antigravity, Google's development tool, Hydrogen arrives with an "Extra" thinking tier and with Orchestrator, a new capability that dynamically switches between models — Carbon and Hydrogen — within the same development session.

4. Market Outlook

Price is half the story. Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens, with flat pricing across the whole 1-million-token window and no long-context surcharge. Flat pricing matters as much as the figure: many providers charge more as context grows, and that penalises precisely the use cases — whole repositories, long documents — that a one-million-token context is for. Google removes that surcharge, so the cost of a long task is predictable from the first token.

In relative terms, Hydrogen costs twice as much as Gemini 3.8 Flash's input ($0.75), but promises 2.5 times the speed and superior execution intelligence, to the point of beating Argon on the independent index. The arithmetic favours anyone who lives on agent loops: paying double per token and finishing each task in less than half the time pays off in almost any coding workflow. For teams chaining hundreds of iterations a day — tests, reviews, verifications — the combination of speed and flat pricing makes Hydrogen the natural candidate for everyday work.

The other side of the market is the pressure this puts on the competition. An execution model with a 1-million-token context, flat pricing and availability in Search and in the Gemini app does not compete only in the API: it competes in the experience of millions of users who have never paid for a token. Every hard search Hydrogen solves in AI Mode and every answer it gives in the app is a free demonstration of the model. Rival labs are no longer arguing only about API benchmarks; they are arguing about who puts their model in front of the end user, and Google has just put its own in the busiest place on the internet.

One caution remains, and the announcement itself invites it. All the performance figures — the 2.5×, the 51 and 54 on the index — are Google's claims. The Artificial Analysis Intelligence Index is an independent index, but the published values come through Google's announcement and have not been reproduced outside the company. The history of recent years is full of launch benchmarks that later get qualified in real use. Hydrogen will be judged over the coming weeks, when developers and users put it to work outside the lab and it becomes clear whether the promised speed survives real traffic and whether the index score holds up on tasks the maker did not choose.

5. Next Steps

Trying Hydrogen takes nothing special. It has been available since the day of the announcement in AI Studio and in the Gemini API, at all tiers, as well as in Gemini for Business, in the Gemini app and in Google Search for all users. In the app, the model is picked from the dropdown; free users get it automatically when the system decides their question calls for it. In Antigravity, it is available to paid Pro and Ultra subscribers, with the "Extra" thinking tier and with Orchestrator, which splits the work between Carbon and Hydrogen within the same session according to what each step calls for.

The first practical decision is the thinking level. Low and medium serve for quick answers and routine tasks, where speed is what matters. High is the default and the balance point Google recommends for daily work: it is the setting at which the 2.5× against Flash and the 51 on the index were measured. Xhigh is the maximum level, the one that reaches 54 points, reserved for the problems worth paying more time and more tokens for. The beauty of the system is that there is no need to change models to change depth: the same Hydrogen stretches or shrinks according to the setting.

For developers, the next step is to measure. The flat price and the one-million-token context invite trying the cases that used to be ruled out as too expensive: passing a whole repository through the model, chaining tests and fixes without chopping them up, letting an agent verify its own changes. The honest comparison is not against the announcement's benchmark, but against what each team used to pay and how long it took with Gemini 3.8 Flash or whatever model it used before. If Hydrogen delivers on the promise, the saving is not in the price per token — which is double Flash's — but in the time per task and in the amount of work that fits into a working day.

For everyone else, the change arrives on its own. There is nothing to install and no app to change: Hydrogen is already the default model for AI Mode and the one that answers the hardest questions in AI Overviews. The next time a search demands real reasoning, the answer will come from Hydrogen. And in Antigravity, Orchestrator will keep splitting the work between Carbon and Hydrogen without the developer having to pick a model at every step: the system decides which one fits at each moment of the session.

6. Conclusion and Assessment

Gemini 4 Hydrogen is the missing piece of the Gemini 4 family. Google had two frontier models to think deeply — Carbon and Argon — and lacked one to execute fast. Now it has it: 2.5 times faster than Gemini 3.8 Flash, with 1 million input tokens and up to 1 million output tokens, four thinking levels up to xhigh, and a flat price of $1.50 per million input tokens and $7.50 per million output tokens. And it puts it everywhere at once: the API, AI Studio, the Gemini app, Google Search and Antigravity.

The assessment is positive, with one reservation. The positive side is the combination: speed, symmetrical context, predictable pricing and massive availability. An execution model that beats its own family's frontier model on an independent index, and that reaches millions of users by default, is no minor launch. The reservation is the usual one with launch benchmarks: the figures are Google's and have not yet been reproduced outside the company. The real verdict will arrive over the coming weeks, with daily use.

What is already certain is the change of map. The Gemini 4 family is complete and divided by job: Carbon and Argon for deep reasoning, Hydrogen for fast execution. The Flash label gives way to a codename, and the fast model stops being the cheap step to become a category with a name of its own. The next move is no longer Google's, but the rivals' and the users': to check whether Hydrogen runs as fast as it says.