Google launches Gemini 4 Hydrogen for coding and agentic tasks

A new layer for rapid execution

Google has officially released Gemini 4 Hydrogen, a new model specifically built for everyday coding and agentic work. Unlike the Carbon and Argon models, which are designed for deep, long-horizon reasoning, Hydrogen focuses on being a fast execution layer. It is intended to handle tasks that require high reliability and quick responses rather than just complex logic.

The rollout is happening globally across a wide range of platforms. This includes every tier of the Gemini API and Google AI Studio, as well as paid versions of Android Studio and Gemini for business. Users who subscribe to paid Google AI Pro and Ultra plans can also access it within the Antigravity environment.

Google is positioning this model as a versatile tool for various professional needs. It is designed to support repository coding, debugging, deep research, and data analysis. By focusing on these areas, Google aims to provide a model that can handle multimodal work while remaining efficient for high-volume tasks.

Balancing cost and performance

One of the most notable features of Hydrogen is its ability to let users choose their thinking levels. Users can select between low, medium, high, and xhigh settings. This allows them to manually balance the quality of the output against the latency and the cost of the operation.

The pricing structure for Hydrogen is set at $1.50 for every million input tokens and $7.50 for every million output tokens. This pricing remains flat across a 1 million-token context window. A context window refers to the amount of information the model can process at one time, which in this case allows for significant data inputs.

In performance testing, Google noted that Hydrogen reached a score of 51 on the Artificial Analysis Intelligence Index when using the default high setting. This actually beats the score of 50 achieved by the Argon model. When pushed to the xhigh thinking level, Hydrogen reaches a score of 54, demonstrating its ability to scale up when needed.

Improving agentic workflows

The model is a key part of Google's push toward agentic work, which involves AI systems performing multi-step actions independently. To support this, the Antigravity platform is gaining a new feature called Orchestrator. This system can automatically switch between the Carbon and Hydrogen models depending on the needs of a single session.

In this setup, Carbon acts as the coordinator to manage parallel agents. Meanwhile, Hydrogen is used to power the worker and verifier agents within isolated workspaces. This division of labor allows for more efficient task completion without wasting computing resources on simple steps.

According to Google, this combined setup approached the task success rate of Carbon while using 24% less quota. It also resulted in 19% faster average completion times. This suggests that using Hydrogen for the 'heavy lifting' of worker tasks is more efficient than using a larger reasoning model for every single step.

Impact on Search and reliability

For the general public, Hydrogen will be integrated into Google Search worldwide. This update aims to provide more intuitive answers and a sharper understanding of user queries. It also enables visual responses and broader links through a process called expanded query fan-out, which helps the engine find more relevant connections.

Google also highlights significant improvements in reliability and safety. Internal evaluations showed that Hydrogen resulted in up to 81% less agentic overreach and up to 79% less under-service. These metrics indicate that the model is better at staying within the bounds of a user's instructions without doing too much or too little.

The model also shows a 96.7% honesty rate in internal tests and demonstrates greater resilience to indirect prompt injection. Prompt injection is a technique where a user tries to trick an AI into ignoring its original instructions. These safeguards are critical for enterprise users who need to ensure the AI remains predictable and secure.

Benchmarking and speed

When compared to previous models, Hydrogen shows significant gains in speed and accuracy. Google reports that Hydrogen is 2.5 times faster at the high thinking level than Gemini 3.8 Flash. It also performed significantly better on the Terminal-Bench 4.1, scoring 57% compared to the 18.2% score of the older model.

Beyond speed, the model is designed to be much more accurate in its outputs. Google claims that Hydrogen cuts fabricated claims by more than fivefold compared to the older Flash model. This reduction in hallucinations—when an AI makes up facts—is a primary goal for the new release.

Hydrogen can process 1 million tokens as input and generate up to 1 million tokens as output. It scored 53.4% on the Multimodal Knowledge Agent test and 71.6% on the Agents' Last Exam v1.1. These scores help establish Hydrogen as a capable model for complex, multi-step interactions that require both memory and precision.