Google Releases Gemini 4 Hydrogen: Paired with Carbon, It Cuts Quota Use by 24% and Speeds Up Tasks by 19%
Google has released Gemini 4 Hydrogen, a smaller coding model that searches for bugs, edits lines of code, and runs tests. In Antigravity, it works alongside the larger Carbon: Carbon builds the plan, and Hydrogen carries it out. According to Google's tests, the pair performs almost as well as Carbon alone, but uses 24% less quota (the usage limit) and completes tasks 19% faster. That makes it worthwhile to hand routine agent work to the smaller model.
AI-processed from TG: @neurogen_news (Neurogen); edited by Hamidun News
When an agent writes code all day, the biggest resource drain is not the hard tasks but the routine work: reading a dozen files, finding a bug, fixing one line, running the tests. Paying for all of that at the price of the most powerful model makes little sense. Google is addressing the problem with the release of Gemini 4 Hydrogen, a smaller model in the Gemini 4 family that takes on everyday work, while the larger Carbon model is kept for complex decisions. According to Google itself, this pair performs almost as well as Carbon alone, but uses 24% less quota (the usage limit allocated under a plan) and completes tasks 19% faster.
What Hydrogen is and how it works
Hydrogen can work on its own, or it can carry out a plan drawn up by Carbon. Its skill set is practical: it searches for bugs, makes targeted code edits, and runs tests. For file reading and searching, the model groups operations into parallel calls instead of running them one by one.
Beyond code, Hydrogen processes multimodal material, compiles reports, and builds web application prototypes. Among its tools, Google lists function calling, code execution, computer use, structured output, and Google Search. Reasoning depth can be set to one of four levels: low, medium, high, and xhigh. High is enabled by default.
The technical parameters are as follows: the total context window is 1,048,576 tokens, which includes the request, the reasoning, and the response. A standard response is limited to 65,536 tokens, including reasoning. For very long generation, there is a long decode mode of up to 1M tokens within the overall window; it works through background runs or continuation. The model accepts text, images, PDFs, audio, and video as input, but outputs only text.
Orchestrator: two models in one session
The most interesting part is Orchestrator in the Antigravity environment. It switches between models within a single session. Carbon breaks down a complex task and builds a plan. Hydrogen executes: it edits code and runs tests. If the smaller model hits a difficult spot, control returns to Carbon.
It resembles a team in which a senior engineer lays out the task and an executor does most of the work. The 24% and 19% figures compare the pair with Carbon alone: 24% less quota is spent, and tasks are completed 19% faster. Google does not disclose its measurement methodology in the announcement in detail, and the company measured everything itself, so these numbers are worth verifying for your own projects.
Price and availability
The standard API rate is $1.50 per million input tokens and $7.50 per million output tokens. Reasoning is billed as output, and the rate is the same across the entire context: there is no surcharge for the long window. Google Search is billed separately, which matters if your agent frequently queries the web.
A free tier with quotas is available in AI Studio and the Gemini app. Free users of the app get the model through automatic selection: the system decides on its own when to engage it. In Antigravity, Hydrogen is available on the Pro and Ultra plans, while Orchestrator is available only on Ultra.
What this means in practice
For developers, the takeaway is simple: the pattern of "an expensive model plans, a cheap model executes" becomes part of the product rather than a homemade wrapper. If you build your own agents on the API, you can reproduce the pattern manually: Hydrogen for routine work, the larger model for analyzing hard problems. It is also worth measuring how the reasoning level affects results. If high, the default, spends too many output tokens, and those are billed at $7.50 per million, it makes sense to lower the level on simple steps.
There are limitations too. Output is text-only, with no images or audio. Orchestrator is locked to the most expensive plan. The price of Carbon is not stated in the material, so the real money savings cannot yet be calculated: it is known only that the pair uses 24% less quota. Even so, the release points in a clear direction: value is shifting from a single most capable model to the ability to distribute work across several.
- Nov 21, 2026Jensen Huang Steps Down as NVIDIA CEO After 33 Years, CFO Colette Kress Becomes Interim CEO
- Nov 21, 2026OpenAI Disbands Robotics Team Again: GPT-6 Astra Completed 74.8% of Tasks in Simulation Versus 51%
- Nov 21, 2026Gondola Used Codex to Ship 25% More Code Without Hiring and Doubled Revenue
- Nov 21, 2026AI Warehouse Clerk Broke a Hidden Rule: DeepSeek V4 Pro Failed in 18% of Cases, Gemini 3.8 Flash in 4.8%
Key Questions Answered
What Hydrogen is and how it works?
Hydrogen can work on its own, or it can carry out a plan drawn up by Carbon. Its skill set is practical: it searches for bugs, makes targeted code edits, and runs tests. For file reading and searching, the model groups operations into parallel calls instead of running them one by one. Beyond code, Hydrogen processes multimodal material, compiles reports, and builds web application prototypes. Among its tools, Google lists function calling, code execution, computer use, structured output, and Google Search.
What are the key details on Orchestrator: two models in one session?
The most interesting part is Orchestrator in the Antigravity environment. It switches between models within a single session. Carbon breaks down a complex task and builds a plan. Hydrogen executes: it edits code and runs tests. If the smaller model hits a difficult spot, control returns to Carbon. It resembles a team in which a senior engineer lays out the task and an executor does most of the work.
What are the key details on Price and availability?
The standard API rate is $1.50 per million input tokens and $7.50 per million output tokens. Reasoning is billed as output, and the rate is the same across the entire context: there is no surcharge for the long window. Google Search is billed separately, which matters if your agent frequently queries the web. A free tier with quotas is available in AI Studio and the Gemini app. Free users of the app get the model through automatic selection: the system decides on its own when to engage it.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.