Google rolls out Genie 4 world model in Project Genie for AI Ultra
Google is rolling out Genie 4, the Google DeepMind world model it calls “our most realistic and interactive world model yet”, in a new version of Project Genie, its browser tool for generating and exploring interactive worlds. Eligible Google AI Ultra subscribers can use it from November 18, 2026, and Google AI Pro members can join a waitlist. Genie 4 is DeepMind’s first real-time world model to generate sound along with the picture, and you can change a world with a text prompt while you walk through it.
What’s new
DeepMind first showed Genie 4 on October 27, 2026 as a research preview for trusted testers. The rollout post by Diego Rivas (Google DeepMind) and Pavan Kumar (Google Labs) puts it in front of paying users:
- Prompts mid-session. While you navigate a world, you can type a new prompt to change it: shift the weather, add a character or redirect the action.
- Native, real-time audio. Genie 4 pairs its visuals with sound that adapts to the scene, from ambient soundscapes to character dialogue. DeepMind’s model page says: “Genie 4 is our first real-time world model to generate audio natively.”
- More realistic movement. DeepMind says physics and human movement, such as aerodynamics and collisions, look more realistic, and that the model draws on Gemini Omni for how the world looks and on other Gemini models to keep a simulation coherent.
- Project Genie agent. You describe a world in plain language and the agent uses Google’s multimodal models to make the first visuals and custom background music. Gemini can also join a session to watch the screen, suggest where to go next or trigger events in the world.
- Community gallery. Curated and community-built worlds that you can remix and publish. Google’s examples include “Living Memory”, which turns personal photos into spaces you can explore, and “Dino Field Guide”, built from 360-degree panoramas.
Genie 3 stays in Project Genie’s model picker and gets its own upgrades: sessions of up to 3 minutes per generation (up from 60 seconds), world chaining to link several worlds into one longer journey, new worlds seeded from 360-degree panorama images, and in-world events from text prompts.
Google does not give Genie 4’s resolution, frame rate, session length or a price beyond the Ultra subscription.
What it means for developers
- No public API yet. DeepMind’s model page says Genie 4 is coming to “early access partners” in the Gemini API and Google AI Studio, with a waitlist. Today’s launch is the consumer Project Genie app at labs.google/projectgenie, not an endpoint you can call.
- Agents and robots are the bigger story. DeepMind says Genie 4 can work with Gemini 4 Argon and its Gemini Robotics research, with text captions steering a simulation toward finishing a task. A world model that reacts to actions is a place to train and test agents without a real room or robot; our post on Gemini Robotics 2.5 covers the robot side.
- Everything is watermarked. All video and audio made in Project Genie carries SynthID watermarking, and every input and output goes through safety screening, so clips you export can be identified as AI-generated.
- Expect drift. Google calls Genie 3 and Genie 4 frontier research. It lists camera latency, intricate physical interactions, real-time audio precision and keeping a scene consistent over time as open problems, and DeepMind adds that details and object permanence can drift in longer sessions or after many prompt edits.
How it compares
Genie 4 sits next to Google’s video models rather than replacing them. A video model such as Gemini Omni 1.2 Flash renders a fixed clip from a prompt, while Genie generates the world frame by frame as you move through it and lets you change it mid-session. Against Genie 3, the new model adds sound, more realistic physics and movement, and on-the-fly prompting; the only session length Google gives is Genie 3’s 3 minutes per generation.