Google launches Gemini 4 Hydrogen for faster agentic workflows

A model built for execution

Google has introduced Gemini 4 Hydrogen, a new model specifically designed for rapid execution across diverse agentic workflows. While the company's frontier models like Carbon and Argon are meant for deep, long-horizon reasoning, Hydrogen focuses on speed and efficiency. It is intended to handle everyday coding tasks, running tests, and verifying changes in software development.

The model is being rolled out across all tiers of Google AI Studio, the Gemini API, and the Gemini app. It also powers Google Search and is available to paid Google AI Pro and Ultra subscribers. This broad availability means the model will be the engine behind many common interactions with Google's AI ecosystem.

The pricing for Hydrogen remains consistent across its full 1 million-token context window. Users will pay $1.50 for every million input tokens and $7.50 for every million output tokens. These prices are designed to make it a viable option for high-volume tasks that require significant processing power but do not necessarily need the highest level of reasoning.

Multi-model orchestration

One of the primary ways Hydrogen functions is by pairing with more powerful models in a multi-model workflow. For example, a team might use Carbon for high-level architectural judgment while using Hydrogen as a high-efficiency engine for writing code. This allows teams to ship complex projects faster without driving up the total cost of the operation.

To facilitate this, Google is introducing Orchestrator in Google Antigravity. This tool dynamically switches between models like Carbon and Hydrogen within a single session based on the task at hand. In testing on Terminal-Bench 4.0, this approach reached near-Carbon success rates while consuming 24% less quota and completing tasks 19% faster.

In parallel workflows, Carbon can act as the coordinator for a project. Meanwhile, Hydrogen can power multiple agents that perform work and verification to build and test changes at the same time in separate workspaces. This setup helps developers manage complex repositories by processing file-reads and search queries together instead of checking each file individually.

Performance and safety gains

Hydrogen offers significant upgrades over the previous Gemini 3.8 Flash model. It finishes tasks 2.5 times faster at its high default setting and more than triples scores on the Terminal-Bench 4.1. It also cuts fabricated claims by more than 5x and adapts better when a tool call fails or a search returns no results.

The model is designed with a focus on safety and dependable execution at scale. It reduces overreach in agentic tasks by up to 81% and cuts misaligned actions and task shortcutting by 12x in coding evaluations. When faced with blocked tasks, it is up to 2x less likely to ignore security boundaries than previous versions.

Google also emphasizes the model's honesty and transparency. Hydrogen achieves a 96.7% honesty rate on internal evaluations and preserves 99.9% chain-of-thought legibility. This means the reasoning steps the model takes remain clear and readable even across more than 200,000 audited turns without deceptive edits.

Impact on everyday users

For general consumers, Gemini 4 Hydrogen will improve the experience within the Gemini app and Google Search. The model is better at following specific instructions and sticking to a user's desired tone and format. This makes it more reliable for drafting emails, creating study guides, or synthesizing dense material into tables and visual breakdowns.

The model also helps connect information across different platforms. It can pull context from uploads and Google apps like Gmail and Drive, as well as connected apps like Airtable and Dropbox. It is designed to break down goals into steps and call the right tools more frequently to help users check off items on their to-do lists.

Search users will see more intuitive responses and a wider range of relevant links. Hydrogen uses advanced reasoning to understand queries that might contain typos or multiple parts. It also upgrades the query fan-out technique, which helps the search engine connect users to a broader variety of helpful websites than before.