Google Launches a Faster, Cheaper AI Model for Everyday Coding and Routine Tasks

Google has introduced Gemini 4 Hydrogen, an AI model built to work quickly. It was announced on 12 November 2026. Google designed it for "agentic" tasks, where an AI carries out a series of steps on its own, and for everyday coding work.
Its announcement places Hydrogen in the Gemini 4 family as the fast model. Two larger models, Gemini 4 Carbon and Gemini 4 Argon, are offered as "frontier" models for deep, long reasoning. A frontier model is one of the most capable models available.
Access is broad. Hydrogen is available in every tier of Google AI Studio, the Gemini API, Gemini for business, the Gemini app and Google Search. Paying subscribers to Google AI Pro and Ultra can also use it in Google Antigravity, Google's workspace for building software with AI agents.
The model card, published by Google DeepMind in November 2026, is a technical fact sheet for the model. It states that Hydrogen is adapted from Gemini 4 Carbon. Hydrogen accepts text, images, audio and video files as input.
Its context window, the amount of material it can hold in view at one time, is 1 million tokens. Input and output share that space. A token is a small piece of text, such as a word or part of one. By default, replies returned right away are capped at 64K tokens. Background tasks can extend output to the full 1M.
Pricing is flat at $1.50 per million input tokens and $7.50 per million output tokens, according to the blog post.
Google also introduced Orchestrator in Google Antigravity. Orchestrator switches between models, such as Carbon and Hydrogen, within a single working session. On Terminal-Bench 4.0, a standard test of AI agents on command-line tasks, Google reports that Orchestrator reached task success rates close to Carbon alone. It used 24% less quota on average and completed tasks 19% faster.
For teams that split work by cost and speed, this is the most important part of the release. It lets the choice between the large model and the fast one be made step by step, rather than once for a whole project. In my view, that is the detail to watch.
On speed, Google says Hydrogen finishes tasks 2.5 times faster than Gemini 3.8 Flash at its high default setting. On Terminal-Bench 4.1, the blog post gives Hydrogen 57%, against 18.2% for Gemini 3.8 Flash.
Google also reports that Hydrogen cut fabricated claims, meaning confident statements that are false, by more than five times compared with Gemini 3.8 Flash. It also outperforms similar models on SWE-sweep, a test of finding and fixing bugs without breaking existing code.
On the Artificial Analysis Intelligence Index, Hydrogen scored 51 at its default high "thinking" level, one point ahead of Gemini 4 Argon at 50. Thinking level controls how much effort the model spends before it answers. At its highest setting, xhigh, Hydrogen reached 54. The gap between the default and maximum settings is larger than the gap between Hydrogen and Argon. Cost and waiting time will therefore change depending on the setting a team chooses.
The model card compares Hydrogen with Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra across a wider set of tests. Hydrogen leads on several of them.
It scored 13.5% on SWE-sweep, against 3.1% for Claude Sonnet 5.5 and 5.8% for GPT-6.1 Terra. On Agents' Last Exam v1.1 it scored 71.6%, compared with 60.4% and 69.8%. On AutomationBench-AA it scored 76.1%, against 71.8% and 60.2%.
On Vending-Bench 2, which measures how well an AI runs a simulated vending business, Hydrogen reached a bank balance of $14,549. The other two reached $11,165 and $9,107. It also led on the MKA Climb WSR accuracy measure, at 53.4% against 52.0% and 43.8%, and on Chartography, at 66.6% against 61.6% and 63.2%.
Hydrogen trails on other tests. On the cheating-adjusted version of Terminal-Bench 4.1, it scored 57.0%, against 63.9% for Claude Sonnet 5.5. On the OSWorld-2.1 offline subset, it scored 77.6%, against 83.9% for Claude Sonnet 5.5 and 74.2% for GPT-6.1 Terra. On Terminal-Bench-Science 0.1 it scored 51.4%, against 59.9% for Claude Sonnet 5.5 and 58.3% for GPT-6.1 Terra.
In my view, the more useful way to read this release is as a question of system design, not a head-to-head contest. A fast model with a large memory and a low price suits high-volume agent work, where the number of small steps matters more than the depth of any one step. Orchestrator then passes the hard tasks to Carbon. Think of a front-desk clerk who handles most requests quickly and calls a specialist for the difficult ones. Whether the quota and speed savings hold outside Google's own Terminal-Bench 4.0 tests is something teams will need to check with their own work.
The broader context here is that AI agents are moving from demonstrations into everyday paid use. In that setting, speed, reliable use of tools and cost per finished task count for as much as a leaderboard score. Hydrogen's published figures speak to those points directly. The reduction in fabricated claims matters for the same reason. In an agent, a confident mistake at step three carries into every later step.
What the release makes possible is a cheaper, faster default for routine coding and office work, with the largest model kept for the hard cases. A single session can now manage that split automatically.


