Google DeepMind's Genie 4 Builds Full AI Worlds With Sound

Type a sentence describing a rainy street market at dusk, and Google's newest AI model will build you a version of it you can walk through, listen to, and reshape in real time just by typing something new. That's Genie 4, which Google DeepMind rolled out this week as a research preview, and it's a meaningfully bigger jump than its predecessor.

What a "world model" actually means

Most people have gotten used to AI that writes text or paints a static image from a prompt. A world model is a different animal: instead of spitting out one picture, it generates an entire interactive space, frame by frame, that responds to you as you move through it. Ask it to turn a sunny street into a thunderstorm, or add a character who walks up and starts talking, and the simulation updates on the fly rather than having to be regenerated from scratch.

Genie 4's headline upgrade is sound. Earlier versions were silent — you got a moving scene and nothing else. This one generates ambient noise, music, and even interactive dialogue alongside the visuals, and DeepMind says it can hold a scene together for up to two minutes before things start to drift, twice as long as Genie 3 managed. It still leans on Google's Gemini models under the hood, and physical interactions, legible text, and long-term consistency are listed by DeepMind itself as open problems rather than solved ones.

Why this isn't just a toy for Indian developers and studios

World models aren't being built as a party trick — the real money is in training robots, self-driving systems, and game characters without the cost of building physical test environments or hand-crafted 3D levels. Code24 covered a similar push last month when Odyssey-3's AI world model learned to drive on Indian roads in just 20 hours of simulated training, precisely because generating chaotic Indian traffic conditions synthetically is far cheaper than recording them for months.

That's the opportunity for India here too. The country's gaming industry has been growing fast but is still largely bootstrapped, and building explorable 3D environments by hand is one of the most expensive, time-consuming parts of making a game. A tool that generates a playable world from a text prompt — even an early, glitchy one — could let small Bengaluru or Hyderabad studios prototype ideas that would otherwise need a budget they don't have. The same goes for robotics labs at IITs and startups working on autonomous vehicles, warehouse robots, or drones, all of which need huge amounts of varied simulated environments to train safely before touching the real world.

A model that can generate its own training ground is, in a very literal sense, a model that gets to practice for free.

The catch: almost nobody can touch it yet

Before anyone gets too excited, it's worth being clear about what "research preview" means here. Access is limited to a small group of trusted testers, Google AI Ultra subscribers through something called Project Genie, and early partners using the Gemini API and AI Studio. There's no public launch date, no pricing, and no independent benchmarks from anyone outside Google. A handful of things worth keeping in mind:

• No confirmed India-specific rollout or pricing has been announced
• Scenes still visually drift after roughly two minutes of continuous interaction
• Complex physical interactions and accurate on-screen text remain unsolved, per DeepMind's own notes
• There's no API model ID yet for developers to build against

So this is early-stage research, not a product Indian developers can sign up for tomorrow. Google has a habit of widening access to these previews over a few months, as it did with earlier Genie and Gemini releases, so the realistic window for broader availability is probably sometime in 2027 rather than this year.

What happens once this actually ships

The interesting part isn't the demo — it's what gets built once a tool like this is cheap and reliable enough for small teams to use daily. A world model that can be steered with plain language collapses a chunk of the technical skill that used to gate who could build interactive 3D content at all. For India's game and robotics ecosystem, which has talent but rarely the budgets of Western or Japanese studios, that kind of leveling could matter more than it does in markets where 3D production budgets were never the bottleneck in the first place.