Gemini 4 Hydrogen rolls out in Google Search and the free Gemini app

Google introduced Gemini 4 Hydrogen on November 12 and began rolling it out globally in Google Search and the Gemini app. Free Gemini users can get an answer from Hydrogen when the app’s automatic model picker assigns it to their question. In Search, Google says the model helps AI Mode handle questions with multiple conditions and assemble answers with tables, images and links.

The Gemini app’s help page says the free plan still routes most straightforward questions to Flash-Lite. Auto can choose Hydrogen for prompts that call for more reasoning. If a free user switches Auto off, the app uses Flash-Lite. People with personal Google AI Plus, Pro or Ultra subscriptions can choose Hydrogen for chats. Model availability for work and school accounts depends on their Workspace license. The app lists the model that handled a prompt in the response details, including when Auto made the choice.

How AI Mode searches a question in parts

The Search change starts with what Google looks up before it writes an answer. AI Mode already divided a question into related topics and ran several searches at once. Google says Hydrogen improves how the system interprets the request, including typos and detailed conditions, so it can search for more relevant subtopics. A travel question with limits on time and luggage, for example, can call for searches on both transport options and baggage rules.

The resulting answer can include a comparison table, images and links to the pages it draws on. Google lists South Korea and Korean among the regions and languages supported by AI Mode. Its help page says AI Mode may return a set of web links when it lacks confidence in an AI-generated answer. Interactive visuals that users can adjust with sliders or switches currently require sign-in and are available in English.

Four reasoning levels in the Gemini API

Hydrogen is generally available through the Gemini API, with standard requests on the free tier. Paid standard pricing is $1.50 per million input tokens and $7.50 per million output tokens. The output rate also covers tokens the model spends reasoning before it writes an answer. Grounding API responses with Google Search is a paid-tier feature, with search queries counted separately from model tokens. The consumer AI Mode in Search has its own availability rules.

Developers can set reasoning effort to low, medium, high or xhigh. High is the API default. The new xhigh setting spends more time and tokens on a response. The model accepts text, images, audio and video. Its context window can hold up to about a million tokens across input and output combined. The developer documentation sets a synchronous output limit of 65,536 tokens. Longer responses use background execution or continuation.

In Google’s model card, Hydrogen scored 57.0% on Terminal-Bench 4.1, a test of software tasks, compared with 63.9% for Anthropic’s Claude Sonnet 5.5. Hydrogen scored 71.6% on Agents’ Last Exam v1.1, a test of professional tasks carried out on a computer. Sonnet 5.5 scored 60.4% there. The tests measure performance in defined work environments. Search accuracy and response time in the Gemini app are different measurements.

Vals AI also tested Hydrogen at xhigh. It placed sixth out of 37 models on the Vals Index with a score of 68.49%, and 11th out of 64 on Terminal-Bench 4.0. The Terminal-Bench result comes from a different version of the test than the one in Google’s model card. The broader Vals Index and the terminal task test are reported separately.

Google is also bringing Hydrogen to the Gemini feature in Google Sheets, starting with AI Pro and Ultra for consumers and eligible business subscriptions for workplace accounts. Its Antigravity development app offers the new model to personal AI Pro and Ultra subscribers.