Google launches Gemini 4 Hydrogen for faster coding and agentic tasks

A model built for rapid execution

Google has introduced Gemini 4 Hydrogen, a model specifically engineered for speed and efficiency in everyday coding and agentic workflows. While Google's other frontier models like Gemini 4 Carbon focus on deep, long-horizon reasoning, Hydrogen is designed to handle rapid execution across many different types of tasks. It serves as a versatile tool for users who need quick results without the overhead of more complex reasoning models.

The model is being integrated into a wide range of Google products, including the Gemini app, Google Search, and Google AI Studio. It is also available to paid Google AI Pro and Ultra subscribers within Google Antigravity. By making Hydrogen available across these tiers, Google aims to bring high-speed intelligence to a broader audience of developers and consumers.

One of the primary goals for Hydrogen is to act as a high-efficiency execution engine. It is built to stand alone for standard work or to pair with larger models in multi-model workflows. This allows teams to use the right tool for the right job, balancing the need for deep architectural judgment with the need for fast, disciplined execution in software development.

Efficiency in software development

For software engineering teams, Hydrogen is positioned as a major upgrade over Gemini 3.8 Flash. It completes tasks 2.5 times faster at its high default setting and shows significant improvements in accuracy. Specifically, it more than triples scores on Terminal-Bench 4.1 and cuts fabricated claims by more than 5x compared to its predecessor.

The model excels at iterative repository coding and debugging by providing focused, surgical edits. Instead of rewriting entire files, it handles issue resolution and codebase Q&A while maintaining the integrity of existing code. It also handles large repositories efficiently by natively batching file-read and search calls in parallel rather than inspecting files one by one.

In real-world testing, partners have seen Hydrogen run between 2x and 2.8x faster on complex data-heavy enterprise workflows. These improvements come while simultaneously reducing token usage by 12% to 23% compared to the 3.8 Flash model. This combination of speed and cost-effectiveness makes it a practical choice for scaling developer workflows that require frequent updates and verification.

Multi-model orchestration and agents

Google is also introducing a new feature called Orchestrator within Google Antigravity to manage multi-model workflows. Orchestrator dynamically switches between different models, such as Carbon and Hydrogen, within a single session based on the task at hand. This allows for a more efficient use of resources by using the most appropriate model for each specific step of a project.

In a parallel agent setup, Carbon can act as a coordinator while Hydrogen powers the worker and verifier agents. This structure allows teams to build and test changes simultaneously across isolated workspaces. This approach helps organizations ship complex projects faster while keeping token costs under control by avoiding the use of expensive frontier models for every single sub-task.

Hydrogen has already shown strong performance in various agentic benchmarks. It ranks #5 on Agent Arena for multi-step tasks like deep research and data analysis. It also leads its class on several specific benchmarks, including the Vals Tax Agent Bench and the Vals Legal Research Bench, demonstrating its ability to handle specialized knowledge work.

Consumer and multimodal updates

For everyday users, Gemini 4 Hydrogen will improve the experience in the Gemini app and Google Search. The model is better at following specific instructions, sticking to a user's preferred tone, and pulling relevant context from various sources like Gmail and Google Drive. It is designed to break down goals into steps and provide clear tables or visual breakdowns without losing important details.

In Google Search, Hydrogen helps the system comprehend natural language queries more effectively, even when they include typos or multiple parts. It also improves the query fan-out technique, which allows the search engine to connect users to a broader range of helpful links. Users can expect more visual and easy-to-scan responses that are uniquely relevant to their specific needs.

The model also handles multimodal knowledge work by synthesizing dense PDFs, spreadsheets, and audio into polished reports. It pairs a 1 million-token input window with up to 1 million output tokens to process large amounts of data. This makes it useful for professional workflows that require analyzing complex visual files or long-form recorded video content.

Safety and thinking levels

Hydrogen features four distinct thinking levels: low, medium, high (the default), and a new xhigh tier. This allows users to scale the reasoning depth of the model based on the complexity of the task, ensuring they only pay for and wait for the level of intelligence they actually need. This flexible structure helps balance quality, latency, and cost across different types of work.

Safety is a core component of the new model, which includes built-in training and real-time surface safeguards. Hydrogen reduces overreach in agentic tasks by up to 81% and cuts misaligned actions by 12x in coding evaluations. It is also up to 2x less likely to bypass permission boundaries when faced with blocked tasks, achieving a 96.7% honesty rate on internal evaluations.

The model also maintains high transparency for its reasoning processes. It preserves 99.9% chain-of-thought legibility with zero deceptive reasoning edits across more than 200,000 audited turns. Additionally, it is more than 2.3x more resilient against indirect prompt injection attacks, making it a more dependable choice for secure enterprise and developer environments.