Google makes Gemini 4 Hydrogen the new default brain behind AI Mode in Search

Google on Thursday made Gemini 4 Hydrogen the default model in AI Mode, the chat-style part of its search engine. The company is pitching it as a model that can untangle long, sloppy or oddly specific questions instead of making people rephrase them.

Robby Stein, Google's vice president of product for Search, called it "the biggest model upgrade to Search yet." He said Hydrogen is "now the new default model in AI Mode & being used for your hardest questions in AI Overviews," the AI summaries that appear above regular results.

The release reaches well beyond Search. Hydrogen went live the same day in the Gemini app, every tier of Google AI Studio, the Gemini API and Gemini for business. In Google Antigravity, the company's coding tool, it is limited to paying Google AI Pro and Ultra subscribers. Google says the Search upgrade is rolling out to users worldwide and works in every language AI Mode supports.

What changes when you search

Google's core claim is that Hydrogen copes better with the way people actually type. The company says it handles "questions with multiple parts, hard-to-understand typos, or highly specific requests." It splits a query into its pieces and answers each one. Picture someone hunting for a laptop that is light, lasts all day on battery and can edit video. Under Google's description, that person can ask for all three at once rather than running three searches.

Behind the scenes, Google says the model runs several searches at the same time across the subtopics of a question. That should pull in a wider range of links than a single query would surface. Answers are also meant to lean less on walls of text and more on images, comparison tables, graphs, maps and interactive tools. In one example Stein shared, the system built an animated driving simulation to explain how a phone can tell which highway lane a car is in when there is no cell service.

The broader link sourcing is the detail worth watching. Publishers and retailers have long worried that AI answers keep people on Google's page. If Hydrogen really cites more sites per answer, it could spread clicks around a bit more. But Google has given no figures on how many links appear or how often people click them. For now that is a promise, not a measured result.

The interactive extras look like the riskier bet. A comparison table for two products is plainly useful. A custom simulation is impressive, but it also gives the model more room to get something wrong in a way that looks authoritative. The company did not address accuracy safeguards for these generated visuals in its announcement.

Speed, benchmarks and the end of Flash

Hydrogen also marks a branding shift. It follows Gemini 3.8 Flash, which launched in September, and Google dropped the "Flash" name rather than carrying it forward. The company calls Hydrogen a "major standalone upgrade" with "significant leaps in both speed and intelligence" over that model. In the Gemini app's model picker, it carries the label "Everyday help."

That positioning sits a little oddly with its job in Search. A model sold as the everyday option is now handling the "hardest questions" in AI Overviews. The likely explanation is cost and speed: at Search's scale, Google needs a model that is quick and cheap enough to run on a huge share of queries. Still, the phrasing suggests Google now believes its mid-tier model is good enough for its toughest consumer queries. That is a confident claim, and outside testing has yet to bear it out.

On benchmarks, Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default "high" thinking level. That edges a model called Argon at 50. At the new "xhigh" setting, the score rises to 54. The company also says Hydrogen "matches or outperforms some of the competing models in its class," naming Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra. These are Google's own numbers and comparisons. The word "some" in that sentence is doing real work: it stops short of claiming Hydrogen beats rivals across the board.

Users get more control over how hard the model thinks, with low, medium, high and xhigh options. In Antigravity, Google lists an "Extra" tier alongside Low, Medium and High. Google also says language understanding improved in every language AI Mode supports.

What developers and app users get

For developers, Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens. It accepts up to 1 million tokens of input natively and can produce up to 1 million tokens of output. A context window that size lets a developer feed in large codebases or long document sets in a single request.

In the Gemini app, Google says Hydrogen sticks more closely to the tone, format and rules a user sets. It produces cleaner tables and visual breakdowns, and the company claims drafts and study guides should come back usable on the first try. Google also says the model is better at picking the right tools, such as Connected Apps, for a task. Free users don't get to pick Hydrogen outright. Instead, the app's Auto setting sends a prompt to Hydrogen when it judges that model the right fit.

The most interesting piece for engineering teams is Orchestrator, a new Antigravity feature. Instead of running one model for a whole session, it switches between models such as Carbon, which became available earlier this week, and Hydrogen. In parallel workflows, Carbon acts as the coordinator. Hydrogen powers the worker and verifier agents that build and test changes at the same time in separate workspaces. Google bills the result as "Same Carbon-level quality, fewer tokens used."

That design makes practical sense. Much of an agent's work is routine, and paying top-model prices for every step is wasteful. Handing that grunt work to a cheaper, faster model while a heavier one makes the judgment calls is how token bills come down. Whether quality truly holds steady is the question teams will answer with their own projects, not Google's announcement.

Taken together, Hydrogen looks like Google's attempt to make one fast, reasonably priced model the workhorse across nearly everything it sells, from Search to the developer API. For ordinary searchers, the test is simple: whether a rambling, typo-filled question finally gets a straight answer, with sources worth clicking.