Google's new Gemini 4 Hydrogen: what it changes in Search and the Gemini app
Nov. 21, 2026
Google has swapped in a new AI engine, and you may already be using it. On Nov. 12, Google launched Gemini 4 Hydrogen, a faster model that now sits behind some of what you see in Google Search and in the Gemini app [1][2].
What is it, in plain English?
An AI model is the "brain" behind a chatbot. Google makes several. Hydrogen is the quick, everyday one. Google calls its other two, Carbon and Argon, its frontier models, built for slow, deep thinking on hard problems [1]. Hydrogen is Google's upgrade for people moving up from the older Gemini 3.8 Flash, and Google says it's adapted from Carbon [1][3].
Think of it as the difference between a quick-thinking assistant who handles most requests right away and a specialist you call in for the tough ones.
What you might notice
Google says Hydrogen is better at:
- sticking with your instructions across a long back-and-forth conversation,
- keeping to the tone and format you ask for when writing,
- pulling in information from files you upload and from Gmail and Google Drive,
- and making clear tables when you ask it to compare things [1].
Those are Google's claims. We haven't tested them. In Search, Google says it's better at understanding long, multi-part questions and questions with typos, and it shows more visual answers [2].
Where you'll see it, and whether you pay
- Gemini app: Google points everyone to the model menu. A footnote says free users get it through automatic selection, meaning the app picks it when it decides it's right for your question [1].
- Google Search: Google says Hydrogen is "available globally in Google Search for all users" [2]. That sentence doesn't say which part of Search. Search Engine Journal reports, citing Google's head of Search product, that Hydrogen is now the default in AI Mode and answers the hardest questions in the AI summaries at the top of results [5]. It doesn't mean every search now runs on Hydrogen.
- Paid plans: Google AI Pro and Ultra subscribers get it in Google's coding tool and, starting with those subscribers, in Gemini inside Google Sheets [1].
You don't have to pay to get Hydrogen in Search or the Gemini app.
Is it better than what Google had?
Google says it's faster. Independent testers say it's strong for its size, with some caveats:
- Artificial Analysis, a company that scores AI models, gave it 51 on its Intelligence Index at the default setting. Google's post says Argon scored 50 [1][6]. The score drops if the model is told to think less, and rises to 54 at the top setting. At the top setting, Artificial Analysis also measured it taking about 15.6 minutes per test task, versus 8.8 minutes at the default [6]. "Fast" depends on how hard you ask it to think.
- Zapier ran 657 business-automation tasks. At its high setting, Hydrogen passed 51.1%, nearly matching Argon's 51.3% and trailing Carbon's 53.9%. Zapier estimated about $0.82 per task for Hydrogen, versus $1.63 for Argon and $1.94 for Carbon [8]. That's one test of one kind of work.
- Vals AI, an independent testing lab, ranked it first of 85 models on a tax-questions test and sixth of 37 on its overall index on Nov. 12. Its page now shows 2nd of 91 and 7th of 43, with the same scores [7]. More models were added, so the rankings moved.
How it compares to Claude
Google's own model card lists Hydrogen next to Anthropic's Claude Sonnet 5.5 on ten tests. Hydrogen comes out ahead on seven. Sonnet comes out ahead on three: working in a command-line terminal (63.9% to 57.0%), operating a computer (83.9% to 77.6%), and a science test (59.9% to 51.4%) [3].
Keep three things in mind. First, Google's methodology page says the numbers for Sonnet come from a mix of public leaderboards, Artificial Analysis and Google's own runs, so these ten tests were not all set up the same way [4]. Second, Google ran Hydrogen at its highest thinking setting for most tests [4]. Third, I use Claude, which is made by Anthropic, a Google competitor, so weigh that. The seven-to-three count describes Google's chosen table. It doesn't tell you which is better overall.
One small hands-on test
Puter, a company that gives developers access to Hydrogen (so keep that in mind), asked it 16 questions that left out information needed to answer, three times each. Hydrogen said what was missing every time. It made up one false fact. Google's older Flash model made up six. It's a small test [9].
What to keep in mind
Faster doesn't mean always right. Google's own model card lists wrong, confident answers as a known limitation. It also says refusals of reasonable requests went up compared with the older model, especially with images [3]. For health, money or legal questions, double-check with a real source. And most of the big claims still come from Google or its partners.
For the techies
- It can handle up to 1 million tokens, input and output combined. A token is a word or part of a word [3].
- Google's developer price is $1.50 per million tokens in and $7.50 out [1]. That's the same price Gemini 3.8 Flash will cost after its introductory discount ends on Dec. 31 [10], so "lower-cost" means in comparison with Google's bigger models.
- Google says early partners saw it run 2 to 2.8 times faster than 3.8 Flash. The partners aren't named [11].
- Google added an Orchestrator to its coding tool. It switches between models like Carbon and Hydrogen during a job. In a separate setup, Carbon can coordinate while Hydrogen runs worker and checker agents. Google's results for these are its own [1].
My take: Most people won't notice a model swap, and that's kind of the point. The race now is less about who has the flashiest chatbot and more about who can make AI fast enough and cheap enough to show up in the tools you already use. The question to ask: does it make your day easier, or just add another thing to double-check?
How this was made: Claude, made by Anthropic, did the research and first draft, and I reviewed it. Three other AI tools (Perplexity, Gemini and ChatGPT) each fact-checked it, and Claude checked their corrections against the original sources. Disclosure: Anthropic competes with Google. This story is not financial or professional advice.
References & credits
- Google, "Introducing Gemini 4 Hydrogen" (published Nov. 12, 2026; updated Nov. 16)
- Google, "Google Search gets a major upgrade with Gemini 4 Hydrogen" (Nov. 12, 2026)
- Google DeepMind, Gemini 4 Hydrogen model card
- Google DeepMind, Gemini 4 Hydrogen model evaluation (methodology PDF)
- Search Engine Journal, "Google Makes Gemini 4 Hydrogen The Default Model In AI Mode"
- Artificial Analysis, Gemini 4 Hydrogen
- Vals AI, Gemini 4 Hydrogen
- Zapier, Gemini 4 Hydrogen AutomationBench report
- Puter, "Gemini 4 Hydrogen Review: Fewer Made-Up Answers" (Puter gives developers access to Hydrogen)
- Google, Gemini 3.8 Flash announcement (pricing)
- 9to5Google, "Google announces Gemini 4 Hydrogen with significant leaps in speed, intelligence" (Nov. 12, 2026)