Google is releasing Gemini 4 Hydrogen, a fast model for coding and AI agents. It is available worldwide from today for developers, businesses and consumers, including in Google Search.
Gemini 4 Hydrogen is the third model in the Gemini 4 family, alongside the heavier models Carbon and Argon. According to Koray Kavukcuoglu, SVP at Google DeepMind and Chief AI Architect of Google, those two are intended for deep reasoning, while Hydrogen carries out tasks quickly.
The model has a context window of one million tokens and costs 1.50 dollars per million input tokens and 7.50 dollars per million output tokens. That price applies to the full context window.
Performance and four thinking levels
On the Artificial Analysis Intelligence Index, Hydrogen scores 51 at the default level, compared with 50 for Argon. At the highest level it reaches 54. The model has four thinking levels: low, medium, high (default) and the new xhigh.
According to Google, those who switch from Gemini 3.8 Flash get tasks completed 2.5 times faster. The score on Terminal-Bench 4.1 rises from 18.2 to 57 percent and the number of fabricated claims drops by more than a factor of five. In early tests, partners saw Hydrogen run two to 2.8 times faster, using twelve to 23 percent fewer tokens than 3.8 Flash.
On arena.ai’s Agent Arena, Hydrogen ranks fifth. It scores 58.8 percent on HLE-Diamond, 52.3 percent on Vals Tax Agent Bench and 55.8 percent on Vals Legal Research Bench, where it shares first place.
Orchestrator for Google Antigravity
For developers, Google is introducing Orchestrator in Google Antigravity. It dynamically switches between models such as Carbon and Hydrogen within a single session. On Terminal-Bench 4.0, Orchestrator achieved nearly the same success rate as Carbon, with an average of 24 percent lower quota consumption and 19 percent faster completion.
With parallel AI agents, Carbon can take on the coordination, while Hydrogen powers the worker and verification agents. They build and test changes at the same time in isolated workspaces.
Safety and Search
Google says Hydrogen reduces excessive intervention in agent tasks by up to 81 percent. The model reportedly bypasses permission boundaries up to half as often and scores 96.7 percent on the internal honesty test for agents. According to Google, it is more than 2.3 times more resilient against indirect prompt injections on the Gray Swan IPI benchmark.
In Google Search, Hydrogen should better understand questions with multiple parts or typos. It delivers more visual answers and a broader range of links. In the Gemini app, the model follows instructions, tone and formatting more closely, and draws context from Gmail, Google Drive and linked apps such as Airtable and Dropbox.
Availability
Developers can find Hydrogen in the Gemini API, Google AI Studio and Android Studio. Paying subscribers to Google AI Pro and Ultra get access in Google Antigravity. Businesses use it via Gemini for business, Antigravity and, with a suitable subscription, Gemini in Google Sheets.
Consumers can select the model in the Gemini app and get it automatically in Google Search. Gemini in Google Sheets launches for subscribers to Google AI Pro and Ultra. You can find the model card with all the detailed information about Gemini 4 Hydrogen here.