Google puts Gemini 4 Hydrogen in Search, the model that chains the tasks

The facts

In Search, Flash is no longer the one answering. Hydrogen is. Google published it on 12 November 2026, under the byline of Koray Kavukcuoglu, SVP at Google DeepMind and chief AI architect. The page, reread on the 13th, still says “today”. Robby Stein, VP of product for Search, calls it “the biggest model upgrade to Search yet”. The Verge, on 12 November, quotes him one step further: Hydrogen is now the default model in AI Mode, and the one used for the hardest questions in AI Overviews.

Google sets two families against each other. Carbon and Argon, for long-horizon reasoning. Hydrogen, for rapid execution of agent workflows: chaining tool calls, writing code, recovering when a search comes back empty. Four thinking levels: low, medium, high (the default) and a new xhigh tier. Published API price: $1.50 per million input tokens, $7.50 per million output, flat across the million-token context window. VentureBeat, the same day, adds a cached-input rate of $0.075, and notes there is no launch promotion on the standard tier. According to the same article, Gemini 3.8 Flash stays at $0.75 / $3.75 until 31 December 2026.

The rollout, as Google describes it, is worldwide: Search for all users, the Gemini app, the API in AI Studio at every tier, Gemini for business. In the app, free accounts go through automatic model selection “when it is the appropriate model”; paid subscribers get Hydrogen in the dropdown. Antigravity and Gemini in Sheets start with Pro and Ultra subscribers. Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high setting, and that on Terminal-Bench 4.1 it moves from 18.2% to 57% of tasks: those are publisher figures. VentureBeat ties Hydrogen’s score of 51 (high) on the Artificial Analysis Intelligence Index to independent testing, one point ahead of Argon, 51 against 31 for Flash. TechCrunch confirms the Search rollout “to all Search users globally”. The sources we read describe no France-specific setting, and do not list the AI Mode languages.

Editorial view

The phrase walking through the announcement is “cutting-edge intelligence to everyone”. The gesture is narrower: Google has chosen the model that answers in Search. Not Carbon. Not Argon. The one that chains. For a French-speaking freelancer, the news is not another API. It is that a two-part question typed in Search, with a typo, is now handled by this model. The API has a price. The line that changes the bill is not Artificial Analysis’s 51. It is Flash’s rate until 31 December.

What this changes in practice. Tomorrow morning, in Search, if AI Mode is there, it is Hydrogen. In the free Gemini app, you do not get the menu: the model is chosen for you. To wire a coding or document agent, $1.50 / $7.50 per million is a job rate, not an all-day inbox rate. VentureBeat calculates that it comes to fifteen times the lowest input rates of Haiku 5.5, GPT-6 Luna or Muse Ember. Where Hydrogen can help is an agent you launch once, with a human at the end: a file review, a ticket, an extraction from a manual. Elsewhere, Flash until 31 December, or a small model, will do, for less.

Watch-outs. A client PDF, a mailbox or an HR sheet remain personal data. VentureBeat describes an invite-only Contributor tier, where Google may use requests to improve products and train models: for business use in France, that tier is not a pricing footnote. Classifying, extracting or chaining tools remains an AI system under the regulation; the transparency duties are in our marker on the 2 December deadline. On lock-in: Search, the app, the API and Antigravity on the same model means more gestures in a single Google account. The sources we read do not say where Search queries are processed, nor whether French AI Mode is already served by Hydrogen.

Trial, or wait. Worth trying: a multi-part question in Search, then the same in the app, rereading the links. Not worth trying “because it is the new Flash”: an agent wired to mail or a client file, at Hydrogen’s rate, with no cap. Worth waiting for: the end of Flash’s intro price on 31 December, before migrating an agent that already runs. And written confirmation, in French, that Search’s AI Mode has actually switched.

Where it actually helps

Where a human rereads what a model has chained: review a large set of documents on an agent’s cloud machine, assign a GitHub ticket and get back a tested pull request, pull precise data out of technical documentation, keep Gemini as the only AI tool authorised for a team. Search itself does not need a directory case. It needs a question reread, and the links under the answer. 31 December 2026 is the date VentureBeat gives for the end of Flash’s introductory price. That is the date to put next to the model menu, not the slogan of 12 November.