Google releases Gemini 4 Hydrogen, a fast model for everyday coding and agent work

Google DeepMind released Gemini 4 Hydrogen on 12 November 2026. Google describes it as a model built for "rapid execution across diverse agentic workflows". Its larger models, Gemini 4 Carbon and Gemini 4 Argon, are for long reasoning on hard problems. Hydrogen is the fast one, and Google is giving it to the widest audience it has: the Gemini API, AI Studio, Gemini for business, the Gemini app, Google Search, and Google Antigravity for paid AI Pro and Ultra subscribers.

The important part is how Google wants people to use it. Hydrogen is not only a cheaper chat model. Google pitches it as the model that writes code, runs tests and checks results, while a larger model plans the work and makes the hard judgment calls.

Numbers and price

According to Google, Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default "high" thinking level. That is one point above Argon's 50. At the new "xhigh" setting it reaches 54. The model has four thinking levels: low, medium, high and xhigh. Google says they let users "only wait and pay for the reasoning depth a task requires".

The price is the same across the full 1 million-token context window: $1.50 per million input tokens and $7.50 per million output tokens.

Google compares Hydrogen with Gemini 3.8 Flash, the model it is meant to replace for many teams. Google says Hydrogen finishes tasks 2.5 times faster at the default setting. It scores 57% on Terminal-Bench 4.1, against 18.2% for Gemini 3.8 Flash. Google also says it makes up claims more than five times less often, and that it "adapts cleanly when a tool call fails or a search returns empty".

Google also lists results on domain benchmarks: 52.3% on Vals Tax Agent Bench, which it says ranks first overall, and 55.8% on Vals Legal Research Bench, which it says ties for first. All of these figures come from Google. Independent results will follow as outside testers run the model.

Splitting the work between models

With the model, Google is launching Orchestrator in Google Antigravity, its agentic development environment. Orchestrator does not run one model from start to finish. It switches between models such as Carbon and Hydrogen within the same session. Google reports that on Terminal-Bench 4.0, Orchestrator came close to Carbon's task success rate while using 24% less quota on average and finishing 19% faster.

For parallel work, Google describes a setup where Carbon coordinates and Hydrogen runs the worker and verifier agents, which "build and test changes simultaneously across isolated workspaces". This is a team structure: one expensive model to plan and review, and several cheaper ones to do the work and check each other's output.

Google backs this up with demos. In Antigravity, Hydrogen built an e-commerce storefront with a WebGL2 shader and a live product configurator in one shot. It turned a YouTube keynote and a design template into a Google Slides deck, then ran its own audit-and-fix loop to correct overlapping text. In Google Sheets, Google says it builds financial models that adjust to forecasting scenarios and produces outputs more than 30% faster than earlier models in Sheets canvas. These are demonstrations picked by Google, not measured results on customer work.

Safety and rollout

Google says Hydrogen uses its Gemini 4 safety stack. It names two focus areas. The first is agent work: less overreach and fewer misaligned actions, and better resistance to prompt injection and attempts to bypass permissions. The second is everyday use: more dependable conversations and fewer false refusals on legitimate medical and scientific questions. Details are in the model card.

The rollout is global and starts today. Developers get it through the Gemini API, AI Studio and the paid tier of Android Studio. Businesses get it through Gemini for business, Antigravity and Gemini in Google Sheets for eligible subscriptions. In the Gemini app, free users get it through automatic model selection. Google Search uses it for all users. The AI Studio documentation lists what changes for developers.

What it means for work

Hydrogen is aimed at the large middle of software work: writing code to an agreed design, running tests, checking changes and fixing small failures. That is much of what junior and mid-level engineers spend their days on. Orchestrator makes the split between planning and doing an automatic setting, so teams may stop choosing a model by hand for each task. Outside coding, the Sheets and Slides demos target routine office work, such as building scenario models and turning recorded talks into formatted decks. The tax and legal research benchmarks point to analysts in those fields, but benchmark scores of 52% and 56% also show how much work still needs a person to check it. What remains open is whether Google's speed and reliability figures hold up on real codebases, and who answers for the result when one model plans the work and another carries it out.