Google DeepMind has shown an experimental interface that behaves like a web browser but never loads a real web page. Instead, when a user types a topic or question, the system uses the Gemini 3.1 Flash-Lite model to write a complete, interactive web page on the spot and renders it in the browser window.
The demo, published in late March 2026 as a developer showcase, is meant to illustrate the speed of Google’s smallest Gemini 3.1 model rather than to launch a product. Google has not described it as a replacement for Chrome or any other browser.
How it works
In a conventional browser, a search returns links to existing pages, and the user picks one to read. In DeepMind’s demo there are no links to follow. The model receives the user’s input and generates HTML, CSS and JavaScript for a page about that subject, often running to thousands of lines of code, within a few seconds. When the user clicks a link or control inside the generated page, the model generates the next page in the same way, so every interaction produces a fresh page.
The key point, which the demo’s own description makes clear, is that the system does not browse the actual internet. It is not pulling text from Wikipedia or a news site. The content comes from the model’s training and reasoning, which means it can be fluent and well-formatted but is not grounded in a specific, checkable source and may contain errors.
Why Flash-Lite
Gemini 3.1 Flash-Lite, released in early March 2026, is the fastest and cheapest model in Google’s Gemini 3.1 lineup, aimed at high-volume tasks where latency and cost matter more than peak reasoning ability. Generating a full page in the time it takes a normal page to load requires exactly that profile. The demo would be impractical with a larger, slower model.
Google positioned the project as a developer tool and a proof of concept. Its code is intended to show developers what the model can do in interactive settings, such as prototyping interfaces or producing on-demand explanatory content.
Limits and open questions
Because every page is created anew, the approach has obvious weaknesses for ordinary use. There is no persistent data, so two users asking the same question will get different pages, and a user who returns later will not see the same page again. Consistency, accuracy and the cost of regenerating content for millions of users are unsolved problems at scale.
Even so, the demo points at a direction several companies are exploring: interfaces that are generated to fit the user’s request rather than designed in advance. Google has already built generative UI features into its Gemini app and AI Mode in Search, where the model assembles custom layouts for some answers. The DeepMind demo takes the idea to its extreme by generating the whole site, and it raises a longer-term question for web publishers about what happens to traffic if AI systems answer with pages of their own rather than sending users to existing ones.
