Google DeepMind shared a demo this week of a browser that doesn't fetch web pages. It writes them. Built by Ben Cobley, a Creative Technologist at DeepMind, the Flash-Lite Browser runs inside Google AI Studio and uses Gemini 3.1 Flash-Lite to generate HTML and CSS in real time from prompts, clicks, and navigation context.
The concept is simple: instead of requesting a stored page from a server, the browser asks the model what should exist right now, then streams the result as code. Type a prompt, get a live page. Click a link, get a freshly generated destination. Google DeepMind posted the demo on X on March 24 with a public link anyone can try for free.
3.1 Flash-Lite makes this possible mostly because of raw speed. Google reports 2.5x faster time-to-first-token than Gemini 2.5 Flash, with output hitting around 363 tokens per second according to Artificial Analysis benchmarks. Priced at $0.25 per million input tokens and $1.50 per million output, it's cheap enough to burn tokens on throwaway interfaces. The model launched in preview via AI Studio and Vertex AI in early March.
Don't mistake this for a product. Cobley's project is a side experiment, and it shows. Output is inconsistent, content hallucinates after a few clicks, and styling drifts. Matt Wolfe compared it to WebSim, the indie AI web generator, but without persistent memory. The interesting bit isn't what the browser does today. It's the fit with agentic workflows: an AI assistant that needs a temporary dashboard or tool mid-task could just generate one on the fly, use it, and discard it. No server, no deployment, no CMS.
The demo builds on an earlier DeepMind prototype from mid-2025 that simulated a generative operating system using the older 2.5 Flash-Lite. That version had the same core idea but lacked the speed to feel responsive. Whether generated-on-demand interfaces go anywhere useful depends on whether the hallucination problem gets solved before the novelty wears off.
Bottom Line
Google DeepMind's browser demo generates live web pages via Gemini 3.1 Flash-Lite at 363 tokens per second, but output quality remains too inconsistent for anything beyond experimentation.
Quick Facts
- Model: Gemini 3.1 Flash-Lite, available in preview since early March 2026
- Speed: 363 tokens/sec output, 2.5x faster TTFT than Gemini 2.5 Flash (company-reported and Artificial Analysis-confirmed)
- Pricing: $0.25/1M input tokens, $1.50/1M output tokens
- Creator: Ben Cobley, Creative Technologist at Google DeepMind
- Access: Free via Google AI Studio
