Google Genie 4: The AI That Builds Playable Worlds in Real Time

Google Genie 4 is DeepMind’s new flagship world model, and it is the closest thing yet to a holodeck: fully 3D, explorable environments generated in real time from plain language, with physics, characters and native sound. You type what you want (a storm rolling in, a new character walking on screen, the camera swinging around), and the world reshapes itself live while objects collide, water flows and crowds react. Google announced it as “our most realistic and interactive world model yet”, and it is rolling out first as a research preview, then to Google AI Ultra subscribers in Canada and 139 other countries through Project Genie.

Our team has followed the Genie line since the first playable demos, and this is the jump that turns it from a research demo into something Canadians will actually get to use. Here is what Genie 4 can do, what it cannot do yet (Google is unusually honest about this), and what it will cost you here.

What is Google Genie 4?

Genie 4 is a world model: instead of generating a static image or a fixed video clip, it generates a persistent, explorable 3D environment that keeps running while you interact with it. Describe a landscape and you are standing in it. Change the weather and watch what happens. Add a character and they move, behave and speak naturally. Google’s own examples show scenes shifting into fully explorable 3D stained-glass landscapes or a 1980s Japanese train station, mid-session, from a single prompt edit.

Three things separate it from a video generator. The worlds obey physics-like rules: objects collide, water flows, trees sway in wind. The worlds have memory: what you built or changed stays coherent as you move through the scene. And the audio is native: leaves crunch underfoot, crowds cheer, characters speak with you or with each other, all generated in sync with the simulation rather than dubbed on afterwards.

How Genie 4 actually works

Genie 4 is not one model doing everything. Google builds it from across the Gemini family, and each part has a job:

That division of labour matters: Gemini interprets your prompt behind the scenes and formulates a dynamic plan inside the changing world, which is what makes complex event direction with plausible cause and effect possible. Direct a scene and things happen because of what came before, not as random effects.

What Genie 4 cannot do yet (the honest list)

Google publishes the limitations on the model page itself, which we appreciate, because most launches hide them:

• Visual drift: details and object permanence can shift during long runs, after complex prompt edits, or across big scene changes.
• Minutes, not hours: sessions run a few minutes of continuous interaction.
• Imperfect real places: real-world locations cannot be simulated with perfect accuracy yet (relevant to the Street View worlds below).
• Text and audio: clear on-screen text only appears if you supply it, and real-time audio accuracy is still active research.
• Multi-agent scenes: modelling many independent agents plus a user in a shared world remains an open challenge.

When can Canadians try Genie 4?

The rollout has three lanes. Genie 4 is live now as a research preview for trusted testers. It reaches early-access partners through the Gemini API and Google AI Studio. And it lands with everyday users through Project Genie, which Google’s Canadian plans page lists as included with a Google AI Ultra subscription across more than 140 countries, Canada among them in both English and French.

So the practical answer for Canada: you need Google AI Ultra, which starts at $139.99 per month. What that subscription unlocks, tier by tier, is in our guide to Google AI Ultra in Canada. The Street View side of Genie, where AI rebuilds real places into explorable worlds, has its own guide here: Project Genie and the Street View worlds.

What Genie 4 is for

Google’s own framing points at four audiences:

• Robotics research: paired with Gemini Robotics, Genie 4 generates unlimited simulated environments on demand, an “infinite playground” for training embodied AI like the SIMA 2 agent.
• Filmmaking: real-time camera control across looks like 35mm film grain, 16mm documentary and anamorphic lens flares.
• Education and world building: walk students through a historical period or a fictional landscape that changes as they ask.
• Agents: the same worlds double as training grounds, which is exactly how the Gemini 4 Argon agents that beat Minecraft fit into the picture.

FAQ