Google ships Gemini 4 Hydrogen across Search as its new default model

Google has made Gemini 4 Hydrogen the default model in AI Mode, rolling it out today across search, the Gemini app, all tiers of Google AI Studio, the Gemini API, Gemini for business, and Google Antigravity for AI Pro and Ultra subscribers. Robby Stein, VP of product at Google Search, called it "the biggest model upgrade to Search yet" and said it is "now the new default model in AI Mode & being used for your hardest questions in AI Overviews."

The change targets a specific failure mode in current assistants: questions that arrive malformed or overloaded. Google says the model handles "questions with multiple parts, hard-to-understand typos, or highly specific requests," and it addresses each part of a compound query in a single response rather than answering only the fragment it parsed most confidently.

What changes in how responses are built?

The retrieval layer has changed too. Instead of running one query against an index, Gemini 4 Hydrogen conducts "simultaneous queries across all the relevant subtopics," which Google says returns a wider range of links in each response.

The output layer now carries more structured visuals: images, comparison tables, graphs and maps. One demonstration from Stein shows the system generating an interactive animated driving simulation in response to a question about how phones determine which lane of a highway a car occupies in areas without cell service — a shift from citing sources to producing a purpose-built artifact that answers the question directly.

Google also reports improved language understanding "across all supported AI Mode languages."

How much control do users get over reasoning depth?

Gemini 4 Hydrogen exposes four thinking levels: low, medium, high and a new "xhigh" mode. The setting lets users trade response latency for reasoning depth, matching a pattern that has become standard across frontier model interfaces. No specific benchmarks for each tier were published.

Where does the model sit against competitors?

Google claims the model "delivers significant leaps in both speed and intelligence" compared to Gemini Flash 3.8, and that it "matches or outperforms some of the competing models in its class," naming Claude Sonnet 5.5 and GPT-6.1 Terra. The company did not publish the underlying benchmark results, so the comparison rests on its own characterization.

For practitioners, the practical shifts are three:

• Compound and typo-laden queries now get decomposed and answered part by part, which changes how query intent should be modeled in analytics.
• Multi-query retrieval widens the set of surfaces and links a single response can cite.
• Four thinking levels mean the same question can return materially different latency and depth depending on user configuration.

The rollout is live today on all listed surfaces, and its behavior in AI Overviews will be the first place most users encounter it.