Google Releases Genie 4: Real-Time World Simulation, Now with Sound and Causality

Google DeepMind has released Genie 4, officially positioned as the “most realistic, interactive world model.”

Sundar Pichai posted a real-time demo on X: a first-person motorcycle view stopped at a red light on a coastal highway, with a gloved hand gripping the handlebar. After entering “traffic light green,” Genie 4 turns the signal green, the engine RPM rises, and the motorcycle accelerates away along the coastline. The scene, sound, and camera are all generated in real time.

The full prompt for this demo is very specific: a sunny coastal highway, a white stop line, an overhead red light, and a first-person motorcycle perspective. The user only adds one line, “traffic light green,” and Gemini reasons out the next multi-step events: the light turns green, the engine starts, and the motorcycle pulls away.

For readers encountering world models for the first time: it generates a real-time environment that you can enter and interact with, which is fundamentally different from pre-rendered video. Users change the scene with text, and the AI responds in real time. Genie 4 builds on this by adding native audio, physics simulation, and Gemini-driven event reasoning.

The official blog lists Genie 4's core capabilities:

Real-time control: Change the weather, move the camera, or add new objects or characters using natural language. Gemini adapts the input to the current scene and continuously monitors the frame between actions to keep the world progressing.
Native audio: Ambient sound, music, and character dialogue are generated along with the world. Characters can respond to events in real time—something Genie 4 does for the first time.
Physics simulation: Supports realistic physics such as aerodynamics and collisions. Character movements and reactions are more natural than in the previous generation.
Multi-step events: Gemini 3.8 Flash acts as a guiding agent in the loop, observes the frame, breaks goals into steps, and then confirms whether each step is complete.

The architecture behind it is a collaboration among the Gemini family: Gemini Omni anchors the simulation to how the real world looks and behaves, while other Gemini models handle reasoning. The first batch of trusted testers has already used it for film and TV previsualization, reconstructing scenes from historical photos, rapid iteration on game prototypes, and robotics simulation. Directors can step into real-time generated scenes and adjust camera positions, lighting, and dialogue on the spot; photojournalism teams can turn an old photo into an explorable dynamic environment.

On the access side, Genie 4 is available to trusted testers as a research preview and will later roll out to Project Genie for Google AI Ultra subscribers, as well as to early access partners of the Gemini API and AI Studio. The official blog provides a waitlist for Project Genie.

People in the comments are already impatient. One commenter said, “GTA 7 will arrive before GTA 6”; another developer built a terminal-based world model in the browser with a 360M-parameter model to play with while waiting.

Google is also showcasing the robotics direction. Rajat reposted a video on X: Gemini 4 Argon breaks user instructions into real-time steps, Genie 4 renders the corresponding visuals, and Apptronik's Apollo humanoid robot plays checkers. Because Genie 4 can be controlled in real time with text, this path points toward rapid generation of robot training environments.

The official announcement also clearly states the limitations: Genie 4 is still an early research model. It can maintain coherence within short sessions, but over longer periods and with more complex edits, visual details and object permanence drift. Complex physics, real-time audio accuracy, text rendering, spatial precision, and long-term consistency are all still being improved. All generated content will be watermarked with SynthID.

In the world-model race, Google is focusing on “real-time” and “native audio.” Genie 4 is still some distance from a reliable, commercially usable general-purpose simulator, but it can already let AI generate a world you can step into, make a few changes to, and hear respond.