Google releases Gemini 4 Hydrogen: a cost-efficient model for coding agents, now in Search too
Gemini 4 Hydrogen is Google's new model for quick agent and coding work, out now in the Gemini API, Google Cloud and Search. Specs, benchmarks, caveats and what to test.
- Author
Umair Mateen- Published
- Reading time
- 4 min
- Category
- AI
Ideas in this post
Tap one for a 5-second answer. Underlined words in the article work the same way.
0 of 4 understood

Google released Gemini 4 Hydrogen on 12 November 2026, a model it describes as built for cost-efficient agentic coding and debugging. It is available to developers through the Gemini , to businesses through Google Cloud, and it now powers AI answers in Google Search for all users worldwide.
What it is
Hydrogen sits below Google's two , Gemini 4 Carbon and Gemini 4 Argon. According to the model card, it was adapted from Carbon, but where Carbon and Argon are aimed at deep, long reasoning jobs, Hydrogen is meant for quick, repeated steps inside agent workflows. Google's developer post lists coding and debugging across a code repository, interactive web development and workflows that mix text, images and video. Google's own line: "Starting today, developers can build with Gemini 4 Hydrogen."
The main specs, from the and Google Cloud docs:
Quick check · Everyday ↔ real
The restaurant and the API, side by side
Switch between the two and see how each part lines up.
- You, the customer
- The menu
- Your order to the waiter
- The kitchen
- Your dish, or "sold out"
You order from the menu and the waiter brings back your dish, without you ever seeing the kitchen.
| Everyday | Real |
|---|---|
| You, the customer | The client (your app) |
| The menu | The documentation |
| Your order to the waiter | The request |
| The kitchen | The server |
| Your dish, or "sold out" | The response, or an error |
From the explainer What is an API? Explained with a waiter in a restaurant
- : up to 1 million tokens of input and output combined. A token is roughly three-quarters of a word.
- Output: up to 64,000 tokens per normal response, and up to 1 million through a slower background mode.
- Inputs: text, images, audio, video and documents such as PDFs and Office files. Output is text only.
- Thinking levels: four settings that trade speed and token use for more careful reasoning.
- Knowledge cutoff: March 2026, so it won't know about later events unless you give it search or documents.
Where you can use it
- Developers: the Gemini API in Google AI Studio and in Android Studio.
- Businesses: Gemini for business, and Google Cloud's Gemini Enterprise Agent Platform, where it is listed as generally available (GA) rather than preview. It runs in a global endpoint and in US and EU multi-regions, with data residency, customer-managed encryption keys and VPC Service Controls supported. Batch jobs and reserved capacity (Provisioned Throughput) are supported.
- Coding tool: Google Antigravity, for paying Google AI Pro and Ultra subscribers.
- Everyday products: the model card also lists the Gemini app and Google Sheets. In Search, Google says Hydrogen improves handling of multi-part questions and typos, links to a wider range of sites, and adds more tables and visual answers.
How it compares
The model card compares Hydrogen with Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra on Google-run tests. The picture is mixed. Hydrogen scored highest on Agents' Last Exam (71.6% against 60.4% and 69.8%) and on AutomationBench-AA (76.1% against 71.8% and 60.2%). Sonnet 5.5 led on Terminal-Bench 4.1, a test of command-line work (63.9% against Hydrogen's 57.0%), and on OSWorld-2.1, a test of operating a computer (83.9% against 77.6%). These are vendor-reported numbers; run your own tasks before switching.
Limits and caveats
- Pricing: the pages we read today do not state per-token prices. Check Google's Gemini API or Google Cloud pricing pages before you budget for it.
- More refusals: Google's own safety tests show Hydrogen refuses harmless requests more often than Gemini 3.8 Flash, up 5.57 percentage points, especially with image and other multimodal prompts.
- Reliability: the model card warns of occasional slowness or timeouts, and higher thinking levels use more tokens, which costs more.
- Fixed settings: on Google Cloud, temperature and similar sampling settings are fixed, and custom values are ignored. Prompts tuned by adjusting those will behave differently.
- Safety report pending: Google says the full safety and security report will follow in the coming days.
Who should care
Teams running coding agents, internal automations or document-processing jobs where many small model calls add up. A faster, cost-efficient tier is usually where the bulk of agent traffic goes, with a larger model like Carbon kept for the hard steps. If your site depends on Google traffic, note that Search answers now come from this model too, which Google says link to a broader range of sites.
What to do this week
- Pick 20 to 50 real tasks from your current agent or automation and run them on Hydrogen next to the model you use now. Compare accuracy, refusals and total tokens, not just speed.
- Look up the current price on Google's pricing page and estimate monthly cost at your volume, including thinking tokens.
- If you're in the EU or handle regulated data, confirm the EU multi-region endpoint and encryption options match your requirements.
- If your prompts rely on a custom temperature, retest them; the setting is ignored on this model.
- Check your Search Console data over the next few weeks for changes in traffic from AI answers.
The DevDrift helps teams test and route AI models inside their workflows, with approvals and audit trails.
Sources
You’ve read the news. Now understand it.
- Learn · 5 min · Everyday ↔ realWhat is an API? Explained with a waiter in a restaurantAn API is a set of rules that lets one program ask another for data or a task and get an answer back. Explained with a waiter carrying orders between you and the kitchen.Start in 5 seconds →
- Learn · 5 min · Side by sideSmall vs large AI models, explained with a bicycle courier and a lorryLarge frontier AI models handle the hardest work but cost more and run slower; small models are quick and cheap for focused jobs. Explained with a bicycle courier and a lorry.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat is a model card? Explained with the label on a food packetA model card is the document published with an AI model that says what it is for, how it was tested and where it falls short. Think of it as the label on a food packet.Start in 5 seconds →
- Learn · 5 min · Try the dialWhat is a context window? Explained with a clerk's deskA context window is how much text an AI model can look at in one go: your instructions, documents, the chat so far and its own reply. Explained with a clerk working at a desk of fixed size.Start in 5 seconds →


