Google Rolls Out Gemini 4 Hydrogen, Its Fast Model for Agentic Work

Mountain View, California — Google has released Gemini 4 Hydrogen, a quicker and cheaper member of its Gemini 4 artificial intelligence family built for coding and agentic tasks, charging developers $1.50 for each million input tokens and $7.50 for each million output tokens.

The model began rolling out worldwide according to a blog post by Koray Kavukcuoglu, SVP, Google DeepMind and Chief AI Architect, Google. Google sees it as the execution layer that sits beneath its two frontier models, Gemini 4 Carbon and Gemini 4 Argon, which it reserves for long, deep reasoning on the hardest problems. The token price stays flat across Hydrogen's whole context window of 1 million tokens.

Hydrogen is also how Google is bringing Gemini 4 to a mass audience. It is available on every tier of Google AI Studio, through the Gemini API, in Gemini for business, the Gemini app and Google Search, and in Google Antigravity for paying Google AI Pro and Ultra subscribers.

Running at its standard high thinking level, Hydrogen scored 51 on the Artificial Analysis Intelligence Index, one point above Argon's 50, Google said, and reached 54 at its top xhigh setting. Users can pick from four thinking levels (low, medium, high and the new xhigh), which the company said lets them trade quality against speed and cost so that they pay only for the reasoning a job needs.

The pitch rests on a division of labor that Google says software teams already use. Frontier models such as Carbon and Argon make design decisions and chase subtle bugs, while a faster model writes the code, runs the tests and checks the changes. Hydrogen can handle that routine work by itself or serve as the execution engine next to the larger models.

To automate the handoff, Google introduced Orchestrator in Antigravity, a feature that switches between models such as Carbon and Hydrogen inside one session. On Terminal-Bench 4.0, it came close to Carbon's task success rate while using 24% less quota on average and finishing 19% faster. When agents run in parallel, Carbon can coordinate while Hydrogen drives the worker and verifier agents that build and test changes in separate workspaces.

Google called Hydrogen a major upgrade for teams moving off Gemini 3.8 Flash. It said the new model finishes tasks 2.5 times faster at its default setting, scores 57% on Terminal-Bench 4.1 compared with 18.2% for the older one, makes more than 5x fewer fabricated claims, and recovers when a tool call fails or a search turns up nothing.

In repository work, the company said Hydrogen makes targeted edits without rewriting files that have nothing to do with the task, and beats similar-tier models on SWE-sweep, a test of finding and fixing bugs without breaking working code. It reads and searches files in parallel batches. In early trials on graphs, SQL and data-heavy enterprise workloads, partners saw it run 2x to 2.8x faster than 3.8 Flash while using 12% to 23% fewer tokens.

Several of the results Google cited touch on professional services. Hydrogen ranks #1 overall on the Vals Tax Agent Bench with 52.3%, ties for #1 on the Vals Legal Research Bench with 55.8%, and leads its class on HLE-Diamond (58.8%) and AutomationBench-AA. It is #5 on Agent Arena, the live ranking run by arena.ai for multi-step jobs such as deep research and data analysis.

The model accepts up to 1 million tokens of input and can generate up to 1 million output tokens. Google said this lets it turn dense PDFs, spreadsheets, charts, audio and recorded video into cited reports, and that it scored 53.4% on the company's new Multimodal Knowledge Agent benchmark and 71.6% on Agents' Last Exam v1.1. In web development, it produces more complete, working layouts with fewer prompts, according to Google.

Early business testers included Box. Yashodha Bhavnani, VP of AI Products at Box, said: “In Box’s evaluations, Gemini 4 Hydrogen’s accuracy significantly improved when compared to Gemini 3.8 Flash, while running about twice as fast and using 12% fewer tokens. We saw the strongest gains in data-heavy workflows, where accurately reading tables and performing complex calculations is essential to producing reliable reports. These results demonstrate meaningful progress in the model’s ability to handle complex enterprise content with greater accuracy and efficiency.”

Spreadsheets are part of the push too. Using Gemini in Google Sheets, Hydrogen can help teams build interactive financial models that shift with different forecasting scenarios, Google said, producing outputs more than 30% faster than earlier models in Sheets canvas. Enterprises with eligible business subscriptions can use it in Sheets, and consumer access starts with Google AI Pro and Ultra subscribers.

Flat pricing and faster execution could make it cheaper for finance teams to automate routine reports and models, though most of the performance figures come from Google's own testing or its partners' evaluations rather than independent audits.

In the Gemini app, Google said Hydrogen sticks more closely to instructions over long conversations, keeps to a user's tone and format, and more often calls the right tools. It can pull context from uploads, Gmail and Google Drive, and Connected Apps such as Airtable and Dropbox. Free users will reach the model through auto model selection when it suits a given prompt.

Hydrogen also drives an upgrade to Google Search for all users worldwide. Google said Search will better read multi-part questions and queries full of typos, return more visual answers that are easier to scan, and use an improved query fan-out technique to point people to a wider range of sites.

On safety, Google measured Hydrogen against 3.8 Flash. It said overreach on agentic tasks falls by as much as 81% and under-service by as much as 79%, misaligned actions and task shortcutting drop 12x, the model is up to 2x less likely to get around permission limits when a task is blocked, and it scored 96.7% on an internal agent honesty test.

The company also reported 99.9% chain-of-thought legibility, with no deceptive reasoning edits in more than 200,000 audited turns, and said Hydrogen is more than 2.3x more resilient to indirect prompt injection on the Gray Swan IPI benchmark. False refusals on HealthBench Pro stay under 0.4%, according to the model card and Google's post.

The launch extends a fast run of Gemini releases. Google introduced Gemini 2.5 in March 2025 as an experimental "thinking model" that works through problems step by step before it answers.

Developers can now build with Hydrogen through the Gemini API in AI Studio and Android Studio, and enterprises can deploy it in Gemini for business, through the API or the model drop-down, and in Antigravity. Consumers can select it from the model menu in the Gemini app.