Gemini 4 Hydrogen: Google’s Fast New AI Is Now in Search

If you searched for something on Google this week, there’s a decent chance a brand-new AI model helped write the answer. It’s called Gemini 4 Hydrogen, and Google started rolling it out on November 12. It isn’t Google’s smartest model. But it might be the one you’ll use the most, because it’s going into Search, the Gemini app and a pile of developer tools at once.

So what is it, and why should you care? Let’s walk through it.

So what actually changed?

Google already had two big Gemini 4 models. There’s Gemini 4 Carbon, the heavy hitter, and Gemini 4 Argon, which we covered last month. Both are built for deep, long reasoning on hard problems.

Gemini 4 Hydrogen is the fast sibling. In its official announcement, Google calls it “our daily driver for agentic workflows.” In plain English, an agentic workflow is a task where the AI takes several steps on its own, like reading files, running code and checking its work.

Google also dropped the old “Flash” name, as 9to5Google pointed out. Hydrogen is the upgrade over Gemini 3.8 Flash, which only launched in September. It can take in up to 1 million tokens and write up to 1 million tokens back. A token is roughly a chunk of a word, so that’s a lot of room for long documents and big codebases.

Where you’ll actually bump into it

This is the part that matters for most people. Hydrogen isn’t hidden in some lab. Instead, it’s showing up in tools you probably already use.

Laptop and smartphone on a desk showing an AI assistant chat interface
  • Google Search: Google says it’s rolling out to all Search users. According to 9to5Google, it’s the default model in AI Mode, and it handles “your hardest questions” in AI Overviews.
  • The Gemini app: It sits in the model picker under “Everyday help.” Free users get it through automatic model selection when it fits the prompt.
  • Work tools: It’s coming to Gemini for business and to Gemini in Google Sheets for eligible subscribers.

So what will you notice? Google says Search should now understand messy, multi-part or typo-filled questions better. Answers should also look more visual and link to a wider range of sites.

In the Gemini app, Google promises better instruction-following. In its words, “email drafts, study guides, rewrites, and documents come back ready to use on the first try.” That’s a bold claim. Still, if it’s even half true, you’ll spend less time saying “no, shorter” to a chatbot.

Why Gemini 4 Hydrogen is a big deal for developers

Here’s where it gets interesting. Gemini 4 Hydrogen costs $1.50 per million input tokens and $7.50 per million output tokens through the API. Google says that price stays flat across the whole 1 million token window. Long prompts don’t suddenly cost more, which is nice.

Then there are the benchmark numbers Google is pushing:

  • 57% on Terminal-Bench 4.1, a coding test, versus 18.2% for Gemini 3.8 Flash
  • 51 on the Artificial Analysis Intelligence Index at its default setting, edging out Gemini 4 Argon’s 50
  • 2.5 times faster than Gemini 3.8 Flash at the high setting
  • More than 5 times fewer made-up claims than Gemini 3.8 Flash

In other words, the “cheap and fast” model now scores about as well as last month’s flagship on one big index. That’s the headline for anyone building AI products.

There’s a neat trick for coders too. In Google’s Antigravity coding tool, a feature called Orchestrator lets Carbon plan the work while Hydrogen handles the routine steps. According to Interesting Engineering, Google says this gets close to Carbon’s success rate while using 24% less quota and finishing 19% faster on average. For now, Orchestrator is only for Google AI Ultra subscribers.

The fine print, and our honest take

Let’s be fair here. Every number above comes from Google. Interesting Engineering flagged that these are Google’s own evaluations, and nobody has checked them independently yet. So treat the charts as a sales pitch until outside testers weigh in.

Google also shared safety numbers. For example, it says Hydrogen cuts “agentic overreach” by up to 81%. That’s when an AI agent does more than you asked, like editing a file you never mentioned. It’s also up to 2 times less likely to sneak past permission limits when a task gets blocked.

Honestly, we find those safety stats more exciting than the speed ones. Fast agents are great. But fast agents that go rogue in your inbox are a nightmare, so any progress there counts.

Okay, but should you care about Gemini 4 Hydrogen?

If you just use Google Search, you don’t need to do anything. Gemini 4 Hydrogen is simply arriving in the background. The bigger story is about price.

When a cheaper model scores close to a flagship, building with AI gets more affordable. That matters a lot for small teams, because every API call eats into a tight budget. For startups in Armenia and other emerging markets, a model like this could turn a “maybe next year” AI feature into a “let’s try it this month” one.

We’ve watched Google move fast with Gemini this year, from new frontier models to tighter defaults in schools. So here’s our open question. If Gemini 4 Hydrogen really is good enough for most jobs, how many people will ever need the expensive models again?

Scroll to Top