Google has introduced Gemini 4 Hydrogen, a model designed for routine coding and task execution inside AI agents. The announcement appeared on November 12 at 18:00 UTC, or November 13 in Beijing. Google places Hydrogen alongside Carbon and Argon, which handle longer reasoning tasks. [1]
The rollout spans the Gemini API, AI Studio, Gemini for business, the Gemini app and Google Search. Google Antigravity access is available to paid AI Pro and Ultra subscribers. In the Gemini app, free users receive Hydrogen through automatic model selection when Google deems it suitable for the prompt. [1]
Pricing and the cost of a request
Google lists a flat rate of $1.50 per million input tokens and $7.50 per million output tokens across the model's million-token context window. The same announcement introduces Antigravity's Orchestrator, which switches models within a session. Google reports 24% lower quota use and 19% faster completion than Carbon in its Terminal-Bench 4.0 comparison, at near-Carbon task success. The figures are vendor-reported. [1]
At those token rates, a request with 100,000 billed input tokens and 10,000 billed output tokens would cost $0.225: $0.15 for input and $0.075 for output. The example covers model tokens before other services or adjustments. Because output also includes reasoning and function calls, the length of the visible answer is only part of the calculation. [2]
How the context budget works
Google Cloud's model specifications describe a combined limit of 1,048,576 tokens for input and output. Output includes reasoning, function calls and the final response. A long source document therefore reduces the space left for generation. The headline output maximum is not an additional million tokens on top of a full input window. [2]
A single foreground generation segment can produce up to 65,536 tokens. Longer results require the documented continuation mechanism, with total output still constrained by the combined budget. The specifications document support through the Interactions API and GenerateContent v1beta1. Developers handling large reports should account for continuation when designing completion and timeout behavior. [2]
The enterprise model accepts Office and OpenDocument files as well as PDFs, text, images, audio and video. Its four reasoning settings are LOW, MEDIUM, HIGH and XHIGH, with HIGH the default. The settings let developers evaluate the model against their own latency and answer-quality requirements. [2]
What existing applications need to check
The migration guide warns that familiar sampling controls may behave differently. Custom temperature, top-p and top-k values are ignored. Frequency and presence penalties, or a request for more than one candidate, return an HTTP 400 error. Google directs developers to thinking levels for reasoning depth and response schemas for output format. [3]
Tool integrations also have stricter bookkeeping requirements. Function responses must match the preceding calls by identifier, name and order. For parallel calls, the first call carries the thought signature; the guide describes errors when that signature is missing, duplicated or replayed out of order. Existing wrappers need a compatibility check before a model switch. [3]
Teams can compare familiar application tasks using the same inputs, then inspect the answers and actions alongside completion times and token bills. Such checks can reveal failures that a benchmark score omits, such as a tool call that exceeds an application’s permissions or a response missing fields required by downstream code.
The enterprise guide also classifies Hydrogen as a short-term generally available release with at least 45 days’ retirement notice. Teams planning a longer deployment should add model lifecycle monitoring to their initial migration work. [3]
Sources
[1] Google announces Gemini 4 Hydrogen
[2] Google Cloud model specifications for Gemini 4 Hydrogen
[3] Google Cloud developer and migration guide for Gemini 4 Hydrogen
Google's headquarters photographed in July 2016. File photo. The Pancake of Heaven! / Wikimedia Commons, CC BY-SA 4.0. Image source. Image license / permission.
Have a correction or a useful perspective?
Get in touch with the publication.
