Google Puts Gemini 4 Hydrogen Into Search for Every User in Its Biggest Upgrade Yet

Google has put its newest artificial intelligence model at the heart of the world’s most-used search engine. On Thursday, 12 November, the company announced Gemini 4 Hydrogen, a fast, low-cost model built for everyday coding and for the autonomous “agent” software that carries out multi-step tasks, and began rolling it out globally the same day across Google Search, the Gemini app, and its developer tools. Robby Stein, Google’s vice president of product for Search, called it “the biggest model upgrade to Search yet”, according to The Verge, and said Hydrogen is now the default model in AI Mode and is being used for the hardest questions in AI Overviews.

The launch matters because of where the model sits. Search is not one more Google product. It is the front door to the internet for billions of people, and the place where Google’s AI ambitions meet ordinary users every day. Until now, the company’s most powerful new models have arrived in stages, first for researchers and paying developers, then for everyone else. Hydrogen breaks that pattern: it went to everyone on day one, including every Search user, free of charge and with no waitlist, according to Google.

“Today, we’re introducing Gemini 4 Hydrogen, bringing cutting-edge intelligence to everyone — with speed and efficiency for everyday coding and agentic workflows,” Google DeepMind wrote in its announcement, published on 12 November. The post describes Hydrogen as the execution engine of the Gemini 4 family: while its larger siblings, Gemini 4 Carbon and Gemini 4 Argon, are built for deep, long-horizon reasoning on the hardest problems, Hydrogen is built for rapid execution across diverse agentic workflows.

What shipped on Thursday

Hydrogen is rolling out globally across Google’s developer, enterprise and consumer surfaces, according to the company. Developers can use it through the Gemini API and Google AI Studio on all tiers, as well as in Android Studio. Businesses can pick it in Gemini for business and in Gemini in Google Sheets, starting with Google AI Pro and Ultra subscribers. In the Gemini app, it appears in the model drop-down, and free users receive it through automatic model selection when Google considers it appropriate for their prompt. Paid Google AI Pro and Ultra subscribers can also try it in Google Antigravity, the company’s agentic coding environment.

But the largest audience by far is Search. Google says Hydrogen is now available globally in Google Search for all users, and that it significantly upgrades the service. The company lists four changes. First, the model is meant to understand a question as the user actually thinks it, even when the question has several parts, contains hard-to-understand typos, or is highly specific. Second, answers are meant to be more visual, with tables, images and interactive tools that are easy to scan. Third, Google says it is upgrading its “query fan-out” technique, in which one question triggers several related searches at once, so that answers connect users to a broader range of helpful links than before, including sites they might not previously have discovered. Fourth, Hydrogen becomes the default model in AI Mode, Google’s conversational search experience, and handles the hardest questions that trigger AI Overviews.

Stein’s description of the launch as the biggest model upgrade to Search yet is the company’s own framing, and it is worth pausing on what it implies. Google has been rebuilding Search around AI answers for more than two years, from the first AI Overviews to the launch of AI Mode. Each step changed what users see when they type a question. This one changes the engine underneath all of them at once, and it does so for every user rather than for a paying subset.

Hydrogen replaces Gemini 3.8 Flash as Google’s everyday model, according to technology news site Technology.org, which covered the launch on 13 November. Google’s own announcement frames it the same way in substance, calling Hydrogen “a major standalone upgrade for teams migrating from Gemini 3.8 Flash”. The Flash name had become shorthand for the fast, cheap tier of Google’s model range; Hydrogen now carries that role, with a new name drawn from the periodic table rather than from a speed metaphor.

Speed, price and four thinking levels

Google’s pitch for Hydrogen rests on three numbers: speed, price and a new way of dialling reasoning up and down. The company says the model finishes tasks about 2.5 times faster than Gemini 3.8 Flash at its default setting, and that partners in early testing saw runs between two and 2.8 times faster while using 12 to 23 per cent fewer tokens. It is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire one-million-token context window, according to Google. The model can take in up to one million tokens and produce up to one million tokens in a single session, which Google says lets it synthesise dense PDFs, spreadsheets, charts, audio and recorded video into cited reports.

Hydrogen also introduces four “thinking levels” — low, medium, high and a new xhigh tier — that let a user or a developer trade reasoning depth against latency and cost. High is the default. The idea is that not every task needs the same amount of thought: a quick rewrite needs less than a tangled debugging job, and the price should follow the work. Google says the levels scale in a smooth, predictable progression, so users only wait and pay for the reasoning depth a task requires.

On Google’s own benchmarks, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, one point above the larger Gemini 4 Argon, and reaches 54 at its highest xhigh setting. On Terminal-Bench 4.1, a coding test run from the command line, Google reports a score of 57 per cent, against 18.2 per cent for Gemini 3.8 Flash — more than a tripling. The company also says the model cuts fabricated claims by more than five times compared with its predecessor, leads its class on HLE-Diamond with 58.8 per cent, ranks first overall on the Vals Tax Agent Bench at 52.3 per cent, ties for first on the Vals Legal Research Bench at 55.8 per cent, and ranks fifth on Agent Arena, a live benchmark of agent performance.

Every one of those figures comes from Google itself, and none has yet been independently replicated. That is not a small caveat. Model benchmarks are chosen, run and reported by the companies that build the models, and the numbers are not directly comparable across vendors. Technology.org, in its coverage of the launch, made the point plainly: “All these figures come from Google. Independent testing will show how well they hold up.” Google’s own published comparisons also show Hydrogen trailing rivals such as Claude Sonnet 5.5 and GPT-6.1 Terra on science and mathematics, on computer use and on Terminal-Bench 4.1, according to Silicon Report, cited by Automater. The honest reading is that Hydrogen looks strong for its price and speed, and that the claims should be treated as claims until independent testing catches up.

The third Gemini 4 model — and the first to reach everyone at once

Hydrogen is the third model in the Gemini 4 family, and its rollout is a deliberate contrast with the two that came before it. Gemini 4 Argon was announced on 30 September and went first to vetted cyber defenders in Google’s Fairwind programme before reaching general availability on the paid API on 13 October. Gemini 4 Carbon followed on 28 October, again starting with a small group of cyber defenders, and reached paying developers in the second week of November. Carbon is the family’s strongest model by Google’s own measures, scoring 59 on the Artificial Analysis Intelligence Index, nine points above Argon, and it is built on Argon. But for its first weeks, almost nobody outside the security community could use it.

Hydrogen skips the queue entirely. It was announced and released on the same day, 12 November, to consumers, developers and businesses at once. The difference is not an accident. Argon and Carbon are frontier models aimed at deep reasoning and long-running agent work, the kind of tasks where a mistake can be expensive and where Google wants trusted testers first. Hydrogen is the everyday model, the one that has to work for billions of search queries, and Google’s strategy is to put it in front of everyone immediately and learn from the traffic.

The family structure also reveals how Google now thinks about AI work. The announcement introduces a feature called Orchestrator in Google Antigravity, which dynamically switches between models inside a single session rather than using one model from start to finish. Carbon can act as the coordinator while Hydrogen powers worker and verifier agents, building and testing changes simultaneously across isolated workspaces. Google says that on Terminal-Bench 4.0, Orchestrator reached near-Carbon task success rates while consuming 24 per cent less quota and completing tasks 19 per cent faster. The pitch is that the future of coding is not one giant model doing everything, but a team of models, each doing the job it is cheapest and fastest at.

The same day Google put AI on your face

Hydrogen was not the only Google launch on 12 November. On the same day, the company opened pre-orders for its first Gemini-powered intelligent eyewear, made with Warby Parker, Gentle Monster and Samsung, with sales beginning on 10 December. The Warby Parker frames start at $395 and the Gentle Monster frames at $449, rising to $489 for some designs, according to 9to5Google and Tech Times. The glasses have no display — that version is promised later, with no date — but they carry a camera, speakers and microphones, and they are the first devices to support a new “Hey Gemini” hotword.

Through the glasses, Gemini can answer questions about what the wearer is looking at, give turn-by-turn navigation, translate speech in real time across more than 70 languages, and read or translate text from signs and menus. The glasses run on a Qualcomm Snapdragon AR1 Gen 1 chip with 2GB of memory and 64GB of storage, carry a 12-megapixel camera, and offer about nine hours of battery life, according to The Gadgeteer, cited by Tech Times. They work with Pixel 7 and newer phones running Android 17, selected Samsung Galaxy phones, and iPhone 13 Pro and later models.

Google also published a privacy plan for the glasses on the same day. The company says camera visuals used by Gemini are not saved in Gemini History, are not used to train AI models, and are not used for advertising. The glasses are designed to blur bystanders’ faces before Gemini processes them, carry an outward-facing light, and have a physical off switch. Age verification is required to activate them. The commitments deserve a careful read, as Tech Times noted in its coverage: conversation text follows Gemini’s normal retention rules, with activity auto-deleted after 18 months and reviewed chats retained for up to three years, so a wearer’s spoken questions may be kept even when the image is not.

The two launches belong together. Hydrogen is the model that makes Gemini fast and cheap enough to run inside everyday products; the glasses are the product that puts it on a person’s face. Google is betting that the next place people meet AI is not a chat window but the world around them, and that the model powering that meeting has to be quick, inexpensive and reliable enough to answer before the moment passes.

A race measured in fractions of a cent

Hydrogen arrives in the middle of an AI race that has moved from capability to cost. The most powerful models still make headlines, but the models that carry most everyday traffic are the fast, inexpensive ones: they answer search queries, fill spreadsheets and run inside agents that make many calls per task. Prices in that tier have collapsed. Stanford’s AI Index found that the cost of GPT-3.5-level performance fell more than 280-fold between late 2022 and late 2024, according to Technology.org, and a Google Flash model set that record low.

Hydrogen costs more than the older Flash models it replaces, so Google is betting that speed and agent reliability justify the premium. Its flat $1.50 input and $7.50 output pricing undercuts the $2 and $10 introductory rates of Argon and Carbon, which are scheduled to rise to $4 and $20 on 1 December. For developers building agents that call a model dozens or hundreds of times per task, that gap compounds quickly. The same logic explains the four thinking levels: if a cheap, shallow pass can answer a question, there is no reason to pay for a deep one.

The rivals are not standing still. Google’s own comparison tables pit Hydrogen against Claude Sonnet 5.5, GPT-6.1 Terra and Fable 5.1, and the company concedes ground on some of those rows even while claiming the lead on others. Anthropic, OpenAI and the rest are shipping their own fast tiers on similar timetables. What Hydrogen changes is the distribution: no rival can put a new model in front of every search user on earth in a single day, because no rival owns Search. That is the advantage Google is spending, and the reason Stein’s “biggest model upgrade to Search yet” line matters more as a statement of distribution than as a statement of capability.

Safety claims for a model built to act

Because Hydrogen is built for agent work — software that takes actions, calls tools and changes files on a user’s behalf — Google published an unusually detailed set of safety numbers with the launch. The company says the model reduces overreach in agentic tasks by up to 81 per cent and under-service by up to 79 per cent, cuts misaligned actions and task shortcutting across tool-use and coding evaluations by twelve times, and is up to two times less likely to slip past permission boundaries when a task is blocked. It reports a 96.7 per cent honesty rate on an internal agent evaluation, preserves 99.9 per cent legibility of its chain of thought with zero deceptive reasoning edits across more than 200,000 audited turns, and is more than 2.3 times more resilient to indirect prompt-injection attacks on the Gray Swan benchmark. Its updated safeguards are also, according to Google, much more precise at distinguishing real biosecurity risks from legitimate medical or scientific questions, keeping false refusals on HealthBench Pro under 0.4 per cent.

Those numbers are, again, Google’s own, drawn from its internal evaluations. But the direction of travel is significant. As models move from answering questions to taking actions, the failures that matter change: a model that books the wrong flight, deletes the wrong file or follows an instruction hidden in a web page it was asked to read is a different kind of risk from a model that writes a wrong paragraph. Google is saying, in effect, that it has measured those risks and that Hydrogen is safer at them than its predecessor. The claim will be tested in public, quickly, because the model is now in the hands of everyone.

The launch also lands in a week when AI’s power is being challenged in court. On the same day Hydrogen was announced, four advocacy groups sued OpenAI, Google, Anthropic, Meta and SpaceX in a San Francisco court, asking a judge to freeze development of AI systems more capable than those in existence, warning of what they called catastrophic harm. That lawsuit, reported by NEXA News HD on 13 November, is a separate story, but it frames the context in which Hydrogen arrives: the companies building these models are being asked, in courtrooms and in public, who gets to decide how fast they advance. Google’s answer, delivered the same day, is to ship faster and more widely.

Analysis: Why It Matters

Search is the product that decides what AI is for

Most AI launches are judged by developers. This one should be judged by searchers, because Search is where the model will do most of its work. Every query routed through Hydrogen becomes a live demonstration, and every answer shapes what billions of people believe a question’s answer looks like. If Hydrogen is better at understanding a rambling, typo-ridden, multi-part question, then the people who ask questions that way — which is to say, most people — get better answers. If it is better at showing the sources behind an answer, then the web behind Search gets more visible rather than less. And if it is not, then the most widely used AI product on earth has just been rebuilt on a promise that does not hold. The stakes are not abstract. They are the next question you type.

The end of the Flash era, and the agent bet

Hydrogen’s real significance may be what it replaces. The Flash line was built for a world in which AI answered questions. Hydrogen is built for a world in which AI does jobs: writing code, running tests, verifying changes, breaking a goal into steps and carrying them out. Google’s language gives the game away — “agentic workflows” appears in the first sentence of the announcement, and the Orchestrator feature is designed for teams of models working together. The company is betting that the next great use of AI is not a smarter chatbot but a reliable worker, and that the model which wins that bet will be the one that is fast, cheap and careful enough to be trusted with the keys. Hydrogen is Google’s entry in that contest, and its first test is the most public one possible: every search on earth.

Benchmarks are claims, not proof — and Google knows it

It is worth being blunt about the numbers in this story. The 2.5-times speed claim, the 57 per cent coding score, the 81 per cent reduction in overreach: all of them come from Google, and all of them use tests Google chose. That does not make them false. It makes them unverified, and the history of AI launches is full of benchmark claims that looked different once independent testers got hold of the model. The most honest thing in Google’s announcement may be the part it did not trumpet: its own comparison tables show Hydrogen trailing rivals on several measures. A company confident enough to publish its losses is a company that expects the wins to hold up. Readers should still wait for the independent results, which will arrive within weeks, before deciding whether Hydrogen is the leap Google describes or simply the latest step in a race where every step is announced as a leap.

Publishers are watching the links, not the model

For the millions of websites that live or die by Google traffic, the important sentence in the announcement is not about speed or price. It is the promise of a “wider range of relevant links” through upgraded query fan-out. AI Overviews have already changed the economics of the web: a 2025 Pew Research Center study found that users clicked on traditional search results in 8 per cent of visits when an AI summary appeared, compared with 15 per cent of visits without one. If Hydrogen surfaces more links, and more diverse ones, that could soften the blow for publishers. If the extra links are buried under longer AI answers, it will not. Google has published no numbers on how link distribution shifts, and no traffic data. The publishers watching this launch are not asking whether Hydrogen is smart. They are asking whether it sends readers, and they will know the answer from their own analytics long before Google tells them.

AI is leaving the screen

The coincidence of the two 12 November launches is the real story of the day. A faster model and a pair of glasses, announced together, describe a single bet: that AI’s next home is not a window on a phone but the world a person is standing in. The glasses have no display yet, and at $395 they are not a mass-market product. But they are the first Google hardware designed from the start around Gemini, and Hydrogen is the first Gemini model designed from the start to be fast and cheap enough to answer a question about the thing you are looking at before you have walked past it. The search box took twenty-five years to become the way people find things. Google is now building its replacement, and Hydrogen is the engine it has chosen to power it.

What to Watch Next

1. Independent benchmarks: Artificial Analysis and other independent testers will publish their own Hydrogen scores within days or weeks. Watch whether the 51 on the Intelligence Index and the 57 per cent on Terminal-Bench hold up outside Google’s labs.
2. AI Mode behaviour in the wild: users will quickly report whether Hydrogen’s answers are genuinely better at multi-part questions and typos, or whether the upgrade is invisible in everyday searching. Social media and search forums are the fastest early signal.
3. Publisher traffic data: website owners will compare their Google referral traffic before and after 12 November. Any shift in how many links Hydrogen surfaces, and to whom, will show up in analytics before it shows up in Google’s announcements.
4. The Gemini app rollout: free users receive Hydrogen through automatic model selection, which means Google decides when it is used. Watch for the model appearing in the app’s drop-down, and for user reports on whether everyday tasks feel faster or more reliable.
5. Developer migration from 3.8 Flash: teams running agents on Gemini 3.8 Flash must now decide when to move. Watch for reports of migration problems, cost changes, and whether the four thinking levels deliver the savings Google promises.
6. Argon and Carbon pricing on 1 December: the introductory $2 and $10 rates for Argon and Carbon rise to $4 and $20. Hydrogen’s flat $1.50 and $7.50 will look cheaper still — watch whether developers shift workloads toward it.
7. The glasses on 10 December: pre-orders are open now and sales begin on 10 December in the United States, Canada and South Korea. Early reviews will test battery life, the “Hey Gemini” experience, and whether the privacy promises hold up in daily use.

Sources

• Google DeepMind — “Introducing Gemini 4 Hydrogen for Agentic Workflows”, 12 November 2026 (primary announcement)
• Google — “Intelligent eyewear collections with Gemini features”, 12 November 2026 (primary announcement)
• Technology.org — “Gemini 4 Hydrogen Replaces Gemini 3.8 Flash”, 13 November 2026
• TechWyse — “Gemini 4 Hydrogen Changes Google AI Mode Search Results”, 13 November 2026
• AndroidPure — “Gemini 4 Hydrogen Now Powers AI Mode in Google Search for All Users”, 13 November 2026
• SmartScope — “Gemini 4: Hydrogen vs. Argon vs. Carbon Differences, Benchmarks, and Pricing”, 13 November 2026