Google launches Gemini 4 Hydrogen for faster coding and agentic workflows

A model built for rapid execution

Google has introduced Gemini 4 Hydrogen, a model specifically designed for speed and efficiency in everyday coding and agentic workflows. While the company's frontier models, Carbon and Argon, are built for deep reasoning on complex problems, Hydrogen focuses on rapid execution. It is intended to stand alone for daily work or serve as a high-efficiency engine within larger systems.

The model is being released across a wide range of platforms, including the Gemini API, Google Search, and the Gemini app. Paid subscribers to Google AI Pro and Ultra will also see it integrated into Google Antigravity. This broad rollout ensures that both developers and general consumers have access to the new model's capabilities immediately.

Hydrogen is priced at a flat rate across its entire 1 million-token context window. It costs $1.50 per million input tokens and $7.50 per million output tokens. This pricing structure is designed to help teams manage costs while scaling their multi-model workloads and complex data processing tasks.

Optimizing developer workflows

For software development teams, Hydrogen is positioned as a disciplined execution model for writing code, running tests, and verifying changes. It is designed to pair with frontier models like Carbon, which handle architectural judgment. This multi-model approach allows Carbon to act as a coordinator while Hydrogen powers worker and verifier agents to build and test changes simultaneously.

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

Hydrogen also shows significant improvements over Gemini 3.8 Flash for iterative repository coding and debugging. It handles issue resolution and codebase Q&A with focused edits that avoid rewriting unrelated files. In early testing of complex enterprise workflows, partners saw the model run 2x to 2.8x faster while reducing token usage by 12% to 23%.

Capabilities for knowledge work

Beyond coding, Hydrogen is built to handle multimodal knowledge work by synthesizing dense files like PDFs, spreadsheets, and audio. It can process up to 1 million input tokens to create polished, cited reports and structured deliverables. The model is designed to follow specific rules and tones closely, making it useful for drafting emails, study guides, and document rewrites.

The model also excels at breaking down complex goals into step-by-step actions by pulling context from Google apps like Gmail and Drive. It can connect to external tools such as Airtable and Dropbox to help users check off items on their to-do lists. For users exploring new ideas, it can generate functional layouts for interactive projects in fewer prompts.

In Google Search, Hydrogen is being used to improve the quality of responses for all users globally. It helps the search engine comprehend natural language queries that might include typos or multiple parts. The model also powers a query fan-out technique that connects users to a broader range of relevant links than were previously discoverable.

Safety and reasoning transparency

Google has integrated its latest safety and security stack into Hydrogen to mitigate misalignment and misuse. The model is designed to be more reliable in agentic tasks, reducing overreach by up to 81% and under-service by up to 79%. It also cuts misaligned actions and task shortcutting by 12x across various coding and tool-use evaluations.

The model is built to be more resilient against indirect prompt injection attacks, showing a 2.3x improvement on the Gray Swan IPI benchmark. It also maintains a 96.7% honesty rate on internal evaluations. These safeguards are intended to ensure that the model remains helpful and secure as it performs more autonomous actions for users.

For everyday interactions, Google tuned Hydrogen to distinguish between real biosecurity risks and legitimate scientific questions. This resulted in keeping false refusals on HealthBench Pro under 0.4%. The model also preserves 99.9% chain-of-thought legibility, which means the reasoning steps remain clear and readable during audited turns without deceptive edits.

Adjustable thinking levels

One of the key features of Gemini 4 Hydrogen is its four distinct thinking levels: low, medium, high, and xhigh. These levels allow users to scale the reasoning depth of the model based on the specific requirements of their task. This progression is intended to be smooth and predictable, ensuring users only pay for and wait for the level of intelligence they actually need.

Hydrogen matches or outperforms competing models in its class across various core developer and enterprise workflows. It is particularly effective at handling multi-step tasks like deep research and data analysis. On the Vals Tax Agent Bench, it leads its class with a score of 52.3%, and it ties for first on the Vals Legal Research Bench with a score of 55.8%.

The model is rolling out globally today across developer, enterprise, and consumer surfaces. Developers can access it via the Gemini API in AI Studio and Android Studio. For consumers, it is becoming available in the Gemini app and Google Search. Business users can find it in Gemini for Google Sheets for eligible subscriptions.