Google releases Gemini 4 Hydrogen, focusing on speed and everyday agent tasks
Google has officially released Gemini 4 Hydrogen, positioned as a fast model for handling everyday work and agent tasks, with native support for 1 million Token input and up to 1 million Token output. Google says it scores 57% on Terminal-Bench 4.1, compared with 18.2% for Gemini 3.8 Flash; on the Artificial Analysis Intelligence Index it scores 51 under the default High reasoning setting and 54 under the XHigh setting. API pricing is $1.50 per million Tokens for input and $7.50 for output. It is now available to users at all tiers of Google AI Studio, Gemini API developers, Gemini for Business enterprise users, and Gemini app users, and is integrated into Google Search.
Why read this
SelectedAI score 91
The original text lays out the division of labor and pricing for Hydrogen, Carbon, and Argon, so readers can judge how multi-model collaboration affects agent deployment costs.
Topics
Body · AI translation
Previously, Google has already launched Gemini 4 Carbon and Gemini 4 Argon, aimed respectively at complex reasoning and highly difficult tasks. Hydrogen, meanwhile, takes on a broader range of everyday work, hoping to provide faster response speeds and lower usage costs while maintaining a relatively high level of intelligence.
Google says that compared with Gemini 3.8 Flash, launched in September 2026, Gemini 4 Hydrogen has achieved significant improvements in both speed and intelligence. It is no longer just a lightweight model emphasizing low cost and fast response, but has gained the ability to handle complex programming tasks and multi-step agentic work.
In terms of context handling, Gemini 4 Hydrogen natively supports 1 million Token input and can generate up to 1 million Token output. This enables it to handle massive code repositories, lengthy technical documents, and complex multi-turn tasks. Especially in agentic workflows, the model can continuously retain a large amount of contextual information, reducing the need to repeatedly provide background materials when executing long tasks.
Google also particularly emphasized Hydrogen's performance in agentic programming. It can participate in continuous work such as code writing, debugging, testing, and modification, rather than merely generating a piece of code according to user requests. For development tasks that require multiple steps to complete, the model can maintain contextual connections between different stages, making it better suited to taking on everyday software engineering work.
In terms of performance testing, data published by Google shows that Hydrogen scored 57% on the Terminal-Bench 4.1 test, while Gemini 3.8 Flash scored 18.2%. This test mainly evaluates a model's ability to complete actual software engineering tasks in a terminal environment, including command execution, code modification, and troubleshooting.
On the Artificial Analysis Intelligence Index, Hydrogen scored 51 points under the default High reasoning setting, slightly higher than Gemini 4 Argon's 50 points; under the highest XHigh reasoning setting, its score further reached 54 points. Google says that while achieving these results, Hydrogen can also typically complete tasks in a shorter time.
However, Google has not positioned Hydrogen as a flagship model to replace Carbon and Argon. Carbon and Argon remain primarily aimed at complex reasoning, long-duration autonomous execution, and more difficult specialized tasks, while Hydrogen emphasizes fast execution and everyday work efficiency. Google hopes that through collaboration among different models, users will not always have to use the more computationally expensive flagship models to handle simple tasks.
In terms of agentic collaboration, Google's Orchestrator feature can delegate some routine work to Hydrogen for execution, while letting Carbon handle more complex decision-making and reasoning. Google describes this model as reducing Token consumption while maintaining Carbon-level work quality as much as possible. Through task allocation among models, enterprises can lower the operating costs of deploying AI agents at scale.
In terms of pricing, Gemini 4 Hydrogen's API input price is $1.50 per million Tokens, and the output price is $7.50 per million Tokens. Google says the entire 1 million Token context window uses uniform pricing, and different price tiers will not be applied as context length increases.
Hydrogen has adopted a fairly broad release strategy this time. It has already been made available to users at all tiers of Google AI Studio, Gemini API developers, Gemini for Business enterprise users, and Gemini app users, and has also been integrated into Google Search. Users with paid Google AI Pro and Ultra subscriptions can also call the model in Google Antigravity.
In the Gemini app, Hydrogen is positioned as a model suitable for everyday assistance. Users can use it for general Q&A, information processing, and programming assistance, while developers can connect it to their own applications and agent systems through the API.
This release also means that Google is further changing the way the Gemini product line is positioned. In the past, the Flash series mainly emphasized speed and cost advantages, while the Pro series took on more complex tasks. Now, Hydrogen even surpasses Argon in some intelligence index tests, indicating that Google hopes to give lower-cost models the ability to handle fairly complex tasks as well.
From the perspective of product layout, the Gemini 4 series is forming a clearer division of labor: Carbon and Argon handle highly difficult reasoning, while Hydrogen handles fast execution and everyday agentic tasks. Through multi-model collaboration, Google is trying to solve an increasingly practical problem: as AI agents begin to continuously execute dozens or even hundreds of steps, relying solely on the strongest model to complete all work is not only slower but also brings high computational costs.
The launch of Gemini 4 Hydrogen is precisely an attempt by Google to find a new balance among model intelligence, execution speed, and operating costs.