Google puts a cheaper Gemini 4 in front of every Search user: Hydrogen is 2.5x faster and makes up facts 5x less, Google says

AI companies love to talk about their most powerful models. But the model most people actually touch every day is usually the fast, cheap one: it writes the AI answer at the top of a search page and powers quick replies in phone assistants. That is the tier Google just upgraded.

On Nov. 12, 2026, at 1:00 p.m. EST (2:00 a.m. Nov. 13 Beijing time), Google announced Gemini 4 Hydrogen on its official blog, and Google DeepMind followed with a post on X. In the Gemini 4 lineup, Carbon and Argon are the high-end models built for long, deep reasoning. Hydrogen is the everyday workhorse, aimed at coding, digesting documents and running multi-step automated tasks quickly and cheaply.

**What Google is claiming**

All of the following are Google's own figures on benchmarks it selected:

- Speed and reliability: Compared with its previous fast model, Gemini 3.8 Flash, Hydrogen finishes tasks 2.5 times faster at its default setting and produces more than five times fewer fabricated claims. On Terminal-Bench 4.1, a test of whether an AI can complete programming jobs on its own inside a computer terminal, its score rose from 18.2% to 57%. - Overall smarts: On the Artificial Analysis Intelligence Index, an independent composite ranking, Hydrogen scores 51 at its default setting, one point above Google's own pricier Argon (50), and 54 at its highest reasoning setting. In other words, the budget model caught the premium one on that leaderboard. - Price: $1.50 per million input tokens and $7.50 per million output tokens, flat across its full 1-million-token context window. A token is a chunk of text, roughly three-quarters of an English word, so a million tokens is several long novels' worth. The "context window" is how much material the model can hold in mind at once.

Google DeepMind's published model card is more mixed. On Terminal-Bench 4.1, Hydrogen's 57.0% trails Anthropic's Claude Sonnet 5.5 (63.9%) in Google's own comparison table, and on OSWorld-2.1, which tests an AI's ability to operate a computer, it scores 77.6% versus 83.9%. Its leads come mostly in knowledge work, multimodal understanding and some reasoning tests; on HLE-Diamond, for example, it scores 58.8% against Claude Sonnet 5.5's 53.3%.

**The big change: it's now inside Google Search**

Google says Hydrogen is rolling out globally to all Google Search users starting now. Robby Stein, Google's VP of product for Search, called it "the biggest model upgrade to Search yet," promising better handling of messy, multi-part questions and typos, more images and comparison tables, and links to a wider range of websites. In the Gemini app, paying users can pick it from the model menu, while free users get it automatically when the system decides it fits the prompt.

For developers, Google also introduced "Orchestrator" in its Antigravity coding tool. The high-end Carbon plans and troubleshoots, Hydrogen writes the code and runs the tests, and the work goes back to Carbon when Hydrogen gets stuck. Google says this split gets close to Carbon-only success rates while using 24% less quota and finishing 19% faster on average. The idea is to stop paying the most expensive model to do everything, and to run AI more like a team: a manager sets the plan and workers carry it out.

**What it means for you**

The most immediate effect: the AI answers that billions of people see on Google Search will now come from a newer model. If Google's "five times fewer fabricated claims" holds up in everyday use, AI search answers that confidently get things wrong could become noticeably rarer. But those are the company's own test results, and for anything involving health, law or money, it's still worth clicking through to the original source. For developers and businesses, a budget model that matches a premium one means more AI work for the same money, and more pressure on OpenAI, Anthropic and others to make their own cheap tiers faster and cheaper.

Original sources published: Google (The Keyword blog), Nov 13, 2026, 02:00 Beijing time / Nov 12, 2026, 1:00 PM EST; Google DeepMind (X), Nov 13, 2026, 02:02 Beijing time / Nov 12, 2026, 1:02 PM EST