Google’s Gemini 4 Hydrogen Is 2.5 Times Faster and Now Powers Search

Google says its new Gemini 4 Hydrogen model finishes tasks 2.5 times faster than the model it replaces. Since Thursday it has been the main model behind AI answers in Google Search, so most people who search will meet it without choosing it.

The new model was announced in a company blog post on 12 November. Gemini 4 Hydrogen is the fast, cheaper member of the Gemini 4 family. It sits below Gemini 4 Carbon and Gemini 4 Argon, the larger models Google aims at the hardest problems. It takes over from Gemini 3.8 Flash, which came out in September.

What changes for people who use Google

In Search, Gemini 4 Hydrogen is now the default model for AI Mode, the chat-style search page. Google also uses it for the hardest questions in AI Overviews, the summaries at the top of results. The company says it copes better with long questions, typos and questions in several parts, and that it shows a wider range of links to websites.

In the Gemini app, paying users can choose the new model from a list of models. Free users get the new model only when the app’s automatic setting picks it for a question.

The numbers Google gives

All of these figures come from Google and have not been checked by anyone else. Google says the model scores 57% on Terminal-Bench 4.1, a test of coding tasks done in a computer terminal, against 18.2% for the older model. It also says the new model makes up facts more than five times less often.

Artificial Analysis, an independent group, runs a ranking of AI models called the Intelligence Index. Google says the new model scores 51 on this index at its default setting. That is one point more than Gemini 4 Argon, the larger model. At its highest reasoning setting, the score rises to 54.

For developers, it costs $1.50 per million tokens of input and $7.50 per million tokens of output. A token is a small piece of text, often part of a word. There is also a free tier for developers, with limits on use.

What Google says about safety

Google says the new model is up to two times less likely to get around permission limits when a task is blocked. The model card lists known limits and safety test results. Like the speed figures, these are Google’s own test results.

The launch comes as AI companies compete on price and speed. Anthropic made its Claude Fable 5.2 model 20% cheaper this week. We reported on that price cut on Tuesday. Earlier this month, OpenAI said ChatGPT now runs about a fifth as many searches as Google. We covered those figures in a separate article.

Why it matters: Search is where most people meet AI without asking for it. A change to the model behind AI Mode changes the answers that a very large number of people see. So far, Google has not published independent checks of its claims.