On Thursday, Google put a brand-new model at the center of the product billions of people use every day. Gemini 4 Hydrogen is now live in Search globally for all users, and Robby Stein, Google's VP of product for Search, calls it "the biggest model upgrade to Search yet." Stein also said Hydrogen is now the default model in AI Mode and handles the hardest questions in AI Overviews — meaning the AI answers most people see first are now written by the newest Gemini.
Hydrogen is the fast member of the Gemini 4 family, and the numbers explain why Google is in a hurry to ship it. Google says it completes tasks 2.5 times faster than Gemini 3.8 Flash at its default high thinking setting, and it replaces Flash as Google's everyday model. It scores 51 on the Artificial Analysis Intelligence Index at that default setting — one point ahead of Argon, the model Google charges developers for — and reaches 54 at its new "xhigh" thinking level. The price for developers is flat: $1.50 per million input tokens and $7.50 per million output tokens across its full one-million-token context window.
What changes in Search is less about a new box to type in and more about what the box does with a messy question. Google says Hydrogen handles multi-part questions, typos, and highly specific requests better, and its improved "query fan-out" runs simultaneous searches across a question's subtopics, then links to a broader range of sites — including ones you might not have found otherwise. Answers get more visual too: more images, comparison tables, and custom tools built for the query, in every language AI Mode supports.
The timing is the tell. Just two days earlier, Google opened Gemini 4 Carbon — its strongest model — to paying developers and Ultra subscribers. Hydrogen is the other half of that move: the everyday model, free, in front of everyone at once.
Two things to watch. First, the promise of more links to more sites is Google's claim, not a measured result — no publisher traffic data came with the announcement, and AI answers have been eating referrals all year. Second, every benchmark in the launch is Google's own, and its own comparison still has Hydrogen trailing Claude Sonnet 5.5 and GPT-6.1 Terra on science and math. The real test starts now, in a billion search boxes.
— via Google Blog