Google made Gemini 4 Hydrogen the default in AI Mode. The benchmark claims are Google's own.
Hydrogen is better at messy, multi-part questions and brings more images and tables into Search. The comparisons with Claude and GPT rest on Google's own framing.
Google made Gemini 4 Hydrogen the default model in AI Mode on November 12, and it’s now available in Google Search for all users worldwide. The pitch is that Search should follow messy questions: several parts in one sentence, a typo or two, and very specific requests. Google calls it “the biggest model upgrade to Search yet.”
What changes when you type a question
Google’s Search post lists four changes. Hydrogen follows instructions more closely, so each part of a multi-part question should get an answer. Its reasoning is meant to improve query fan-out, the step where Search runs several searches at once on subtopics, so it reaches a wider set of sites. Answers should look more like a page than a paragraph, with more images, structured layouts, comparison tables and custom tools. Language handling is supposed to improve across every supported AI Mode language. [1]
The Verge adds two details. It says the answers also include graphs and maps, and it quotes Robby Stein, Google’s VP of product for Search, saying Hydrogen is “now the new default model in AI Mode” and is being used for the hardest questions in AI Overviews. [3]
My read: the AI Overviews part may matter more for ordinary searches than the AI Mode part, because Overviews sit on the main results page. Google’s Search post doesn’t mention Overviews at all, and neither Google post says how often Hydrogen runs there.
The benchmark numbers, and whose they are
The model post has the only hard numbers. Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, one point above Gemini 4 Argon at 50. At its top xhigh setting, Hydrogen reaches 54. [2]
| Model and setting | Index score, per Google |
|---|---|
| Gemini 4 Hydrogen, high (default) | 51 |
| Gemini 4 Hydrogen, xhigh | 54 |
| Gemini 4 Argon, setting not given | 50 |
My read: a one-point gap is close enough that I’d call Hydrogen and Argon roughly level on this index. The xhigh score is the one that makes the headline look bigger. The post doesn’t say how many runs sit behind each score or which thinking setting it used for Argon.
Our Gemini 4 Argon price story cited Artificial Analysis putting Argon at 53, not 50. Both figures could be right if they used different Argon settings, but Google doesn’t say which one it used, so don’t treat either as a correction of the other.
The bigger claim is about rival models. Google says Hydrogen “matches or outperforms some of the competing models in its class.” The Verge says that comparison includes Claude Sonnet 5.5 and GPT-6.1. Google’s post doesn’t name those models in the text I read, and I didn’t find an independent test of Hydrogen against them while writing this. Until someone publishes numbers, I’d read that line as Google’s claim, not a settled result.
What the post leaves out
Google says Hydrogen connects you to “a broader range of helpful links.” It doesn’t say how many of those links get clicks, how the sites are picked, or whether publishers will get more visits or fewer. That’s the same problem Reddit raised in its shareholder letter, which said search referrals fell by more than a million a week in October. Our Reddit story covers what it blamed.
My read: a wider set of links on the page doesn’t mean more traffic for the sites on it. A model that answers more of the question itself can send fewer people onward, which is the kind of drop Reddit is describing. Google’s post makes the case for better answers and says nothing about that trade-off.
Where it sits in the Gemini 4 family
Google’s model post describes Hydrogen as built for “rapid execution across diverse agentic workflows,” and calls Carbon and Argon its frontier models for deep, long-horizon reasoning. [2] The Search post doesn’t say which model handles which part of Search beyond Hydrogen itself, so that split is a description of the model family, not of Search.
My read: the split suggests the Search upgrade is about faster, clearer answers to everyday questions, not the heavy analysis Carbon and Argon are meant for. For anything that matters, open the links it gives you and check the summary against them.
Note on sources
I read Google’s Search post and its model post in full. I read The Verge story up to the sentence naming the rival models, which cut off at “GPT-6.1” in my copy, so I haven’t seen the rest of that sentence. I didn’t open Robby Stein’s X post, which The Verge quotes, or the Artificial Analysis page. Every Hydrogen benchmark in this story is Google’s own figure.