Google has a new answer for everyone who thought its best AI models were getting too expensive to actually use. Gemini 4 Hydrogen launched on November 12, and it is the cheapest model in the Gemini 4 family, the first one free users can reach, and, on Google’s own benchmark, slightly faster than the flagship it sits below.
The model is built for the unglamorous work that now dominates AI spending: writing code, running tests, verifying changes, and chaining tools together in agentic workflows. Google DeepMind’s Koray Kavukcuoglu, the company’s senior vice president and chief AI architect, called it “our daily driver for agentic workflows” in the launch post. The heavier Gemini 4 Carbon and Argon models stay reserved for long, deep reasoning. Hydrogen is the execution engine underneath them.
What Gemini 4 Hydrogen costs
Hydrogen is priced at $1.50 per million input tokens and $7.50 per million output tokens, flat across its entire 1 million-token context window, with no long-context surcharge. At current exchange rates that works out to roughly Rs 133 per million input tokens and Rs 665 per million output tokens. There is also a rate-limited free tier on the Gemini API, something neither Carbon nor Argon offers, and cached input costs $0.075 per million tokens through implicit caching.
The price gap matters more than it looks. Carbon, which went generally available on November 10, lists at $2 per million input and $10 per million output on its introductory rate, and that doubles to $4 and $20 on December 1. Hydrogen undercuts Carbon’s list price by about 62 percent on both input and output, and it comes in 25 percent below Argon’s introductory pricing.
| Model | Input per 1M tokens | Output per 1M tokens | Free tier |
|---|---|---|---|
| Gemini 4 Hydrogen | $1.50 | $7.50 | Yes |
| Gemini 4 Carbon | $2.00 intro, $4.00 from Dec 1 | $10.00 intro, $20.00 from Dec 1 | No |
| Gemini 4 Argon | $2.00 intro, $4.00 from Dec 1 | $10.00 intro, $20.00 from Dec 1 | No |
| Claude Fable 5.2 fast mode | $16.00 | $80.00 | No |
For teams running agents at volume, a model that matches the flagship at well under half its standard price changes the economics of production workloads. That is the whole pitch.
Benchmarks: ahead of Argon, behind on the hardest coding test
Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, one point above Argon, and 54 at a new xhigh setting. It ranks fifth on Agent Arena, the live benchmark for multi-step agentic work, and leads its class on HLE-Diamond at 58.8 percent. On Terminal-Bench 4.1, the command-line coding test, Google reports 57 percent, close to Carbon’s success rate while using 24 percent less quota and finishing 19 percent faster. Against the older Gemini 3.8 Flash, Hydrogen finishes tasks about 2.5 times faster and more than triples the Terminal-Bench score, 57 percent against 18.2 percent.
Those are Google’s own figures, and they have not been checked independently yet. The independent picture is more mixed. On the cheating-adjusted Terminal-Bench 4.1 table, Claude Sonnet 5.5 sits at 63.9 percent, above Hydrogen’s 57 percent. And on the independent benchmark index, Claude Fable 5.2 still leads the field, which is the comparison that matters most for buyers choosing a frontier model this month.
Four thinking levels and a 1 million-token window
Hydrogen is adapted from Gemini 4 Carbon, with a 1 million-token input context and up to 1 million output tokens, and a knowledge cutoff of March 2026. It offers four thinking levels, low, medium, high by default, and a new xhigh tier, so developers pay only for the reasoning depth a job needs. The model accepts text, images, audio and video, and can synthesize dense PDFs, spreadsheets, charts and recorded video into cited reports.
Google also wants Hydrogen working alongside its bigger models. A new Orchestrator feature in Antigravity, its agent coding app, switches between Carbon and Hydrogen inside a single session, using the frontier model for architectural judgment and Hydrogen for the repetitive execution. In Carbon’s multi-agent swarm teams, Hydrogen can run the worker and checker agents.
Safety claims aimed at agent deployments
Because Hydrogen is meant to run tools on its own, Google made safety claims aimed squarely at agent deployments. It says the model reduces overreach on agentic tasks by up to 81 percent compared with Gemini 3.8 Flash, cuts misaligned actions and task shortcutting across tool-use and coding evaluations twelvefold, and is up to twice as unlikely to work around permission boundaries when a task is blocked. The model card found no material general-capability increase over Argon, and targeted tests placed it below the biological-information and cybersecurity critical capability levels in Google’s framework.
Where you can use it today
Hydrogen is rolling out globally across Google’s developer, enterprise and consumer surfaces: AI Studio on free and paid tiers, the Gemini API, the paid tier in Android Studio, Gemini for business, the Gemini app, and Google Search, where Google calls it the biggest model upgrade yet. Paid Google AI Pro and Ultra subscribers get it in Antigravity and in Gemini inside Google Sheets. Free Gemini app users reach it through automatic model selection, and it is already in Cursor, the popular AI coding editor.
It is the third Gemini 4 model to ship in just over six weeks, after Carbon and Argon, and the pattern is now clear. Google is splitting its lineup into frontier models for the hardest problems and a cheap, fast execution model for everything else. Carbon’s introductory pricing ends November 30, which makes Hydrogen the obvious default for most developers from December onward, unless the independent benchmarks tell a different story once they land. Google’s new TPU hardware is built for exactly this kind of high-volume agent workload, and Hydrogen is the model that will fill it.
