According to an official announcement published by Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google, the company is launching Gemini 4 Hydrogen. The model is designed to provide speed and efficiency for everyday programming tasks and agentic workflows. While the company's frontier models, Gemini 4 Carbon and Gemini 4 Argon, excel at deep, long-horizon reasoning for particularly complex problems, Hydrogen is built for rapid execution across a variety of agent-based workflows. With the launch of the model, Google is bringing Gemini 4 to all tiers of Google AI Studio, the Gemini API, Gemini for business, the Gemini app, Google Search, and paid Google AI Pro and Ultra subscribers in Google Antigravity.
On the Artificial Analysis Intelligence Index, Hydrogen surpasses Argon at its default level, the high thinking level (scoring 51 versus 50), and reaches a score of 54 at its highest thinking setting, xhigh, while often completing tasks faster than other models in its class. The model is priced flat across its entire context window of 1 million tokens: $1.50 per million input tokens and $7.50 per million output tokens.
Efficiencies in Multi-Model Workloads and the Orchestrator Tool
According to the announcement, in real-world software development, teams rely on frontier models like Carbon and Argon for deep architectural judgment when designing systems or untangling subtle bugs, but they still need a fast and disciplined execution model for writing code, running tests, and verifying changes. Hydrogen is built to stand on its own as a model for everyday work, or to operate in combination with Carbon and Argon as a high-efficiency execution engine in multi-model workflows.
To automate this multi-model collaboration, Google is introducing Orchestrator within Google Antigravity. Rather than using a single model from the beginning of the process to the end, Orchestrator dynamically switches between models like Carbon and Hydrogen within the same session. On the Terminal-Bench 4.0 benchmark, the Orchestrator reached task success rates close to those of Carbon, while on average consuming 24% less quota and completing tasks 19% faster. For workflows requiring parallel agents, Carbon can act as a coordinator, while Hydrogen can power worker agents and verifier agents to build and test changes in parallel across isolated workspaces.
Upgrade for Developer and Enterprise Workloads
Beyond multi-model workflows, according to the report, Hydrogen represents a significant standalone upgrade for teams migrating from Gemini 3.8 Flash. It delivers leaps in both speed and intelligence: it completes tasks 2.5 times faster at the default high setting, more than triples scores on Terminal-Bench 4.1 (57% versus 18.2%), reduces fabricated claims by more than 5x, and cleanly adapts when a tool call fails or when a search returns empty.
The model handles core developer and enterprise tasks across several areas:
- Iterative coding and debugging in repositories: The model handles issue resolution, debugging, rapid iteration, and codebase Q&A using targeted edits, without rewriting unrelated files. In the SWE-sweep benchmark, where it must discover and fix bugs without breaking existing code, it outperforms models in a similar tier. To explore large code repositories efficiently, the model naturally batches file-read and search calls in parallel; in early partner testing across data-heavy enterprise workflows, complex graphs, and SQL, it was observed to run 2x to 2.8x faster while reducing token usage by 12% to 23% compared to 3.8 Flash.
- Open-ended agent task execution: On multi-step tasks such as deep research, data analysis, and broad codebase navigation, the model ranks fifth on Agent Arena by arena.ai. It leads its category on HLE-Diamond (with 58.8%), ranks first overall on Vals Tax Agent Bench (with 52.3%), ties for first place on Vals Legal Research Bench (with 55.8%), and leads its category on AutomationBench-AA.
- Multimodal knowledge work: Hydrogen pairs an input of 1 million tokens with an output of up to 1 million tokens to process dense PDFs, spreadsheets, charts, audio, and recorded video into cited reports and structured deliverables. It outperforms other models on the Multimodal Knowledge Agent (MKA) benchmark with a score of 53.4% for end-to-end agentic knowledge work on complex visual files, and on Agents' Last Exam v1.1 with 71.6% for end-to-end professional workflows.
- Web development: The model generates more functional and complete layouts in fewer prompts. An example of this is Geminisland, where using Hydrogen enables adding custom structures that players can explore in first-person to a shared voxel island from a single prompt. Paired with Gemini's front-end design skill in Google Antigravity, the model built an e-commerce storefront in a single prompt featuring a WebGL2 topographic shader, four-stage scrollytelling, a thermal slider, a 3D spec card, and a live finish configurator. In Google Sheets, the model turned a spreadsheet into an animated pixel canvas by shrinking thousands of cells and combining Python, LET formulas, and conditional formatting. In addition, from a YouTube keynote video and a design template in Antigravity, the model created a presentation in Google Slides and ran an autonomous audit-and-fix loop. In Google Sheets, the model generates interactive financial models over 30% faster than previous models in the Sheets canvas.
Four Thinking Levels and Adaptability Across Workloads
Gemini 4 Hydrogen offers four thinking levels: low, medium, high (default), and a new tier named xhigh. These levels provide options across quality, latency, and cost, so that the user waits and pays only for the reasoning depth required for the task. According to Google, across the different thinking levels, the model matches or outperforms some competing models, while often completing tasks faster and using fewer output tokens.
Assistance in the Gemini App and Google Search
According to the announcement, for consumers the model makes it easier to complete tasks, staying on track and following instructions more closely in multi-step conversations. In the Gemini app, it calls appropriate tools and helps reach reliable output in fewer conversational turns.
In writing and research, it maintains defined tone, format, and rules when drafting emails, study guides, and rewrites, and produces tables and visual breakdowns when comparing products. The model breaks down goals step-by-step and pulls context from uploaded files, from Google apps such as Gmail and Google Drive, and from connected apps like Airtable and Dropbox. In the area of prototyping, it was reported that the model built an interactive 3D globe and a Web Audio music engine driven by real-time lightning strike data, created a WebGL2 water and glass visualization based on mathematics and SDF textures, and produced an archival footage-based video about the element Hydrogen.
In Google Search, the model upgrades the system's understanding of user intent even when the query consists of several parts or includes spelling errors, provides visual responses, and upgrades the query fan-out technique to connect to a broader range of links.
Safety Mechanisms and Global Availability
The model includes improvements from the Gemini 4 safety and security stack, combining built-in training with real-time safeguards. According to the data released by Google, the model reduces overreach in agent tasks by up to 81%, reduces under-service by up to 79%, and decreases misaligned actions and shortcuts in tool usage and coding tests by 12x. When encountering blocked or impossible tasks, it is up to 2x less likely to bypass permission boundaries, achieving a 96.7% honesty rate on Google's internal agent honesty evaluation.
Additionally, the model preserves 99.9% chain-of-thought legibility without deceptive reasoning edits across more than 200,000 audited turns. In the Gray Swan IPI test evaluating resistance to indirect prompt injection attacks, the model is more than 2.3x more resilient. In healthcare, safeguards keep the false refusal rate below 0.4% on HealthBench Pro when distinguishing between biosecurity risks and legitimate medical and scientific questions.
The model is available starting today worldwide: for developers via the Gemini API in AI Studio, in the paid tier in Android Studio, and in Google Antigravity for Google AI Pro and Ultra subscribers. For enterprises, it is available in Gemini for business via the API or model selector menu, in Google Antigravity, and in Gemini in Google Sheets for eligible business subscriptions. For general users, it is accessible in the Gemini app, in Google Search, and in Google Sheets for Pro and Ultra subscribers.