On October 27, 2026, Google DeepMind announced Genie 4, a world model that generates interactive environments in real time. Users move with keyboard controls and steer the scene with text prompts. It is available only to trusted testers for now. Google says it will come to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio, and it has opened a waitlist.
What is new
DeepMind's post lists four advances. Genie 4 is "natively multimodal", generating music, ambient sound and dialogue together with the visuals. Users can change a running simulation in natural language, for example the weather, the camera or a new person or object. DeepMind claims better physics, including aerodynamics and collisions, and more natural human movement. TechCrunch, briefed by the team, reports that Genie 4 keeps a scene consistent for up to two minutes, twice Genie 3's limit.
Gemini as the director
The architectural change matters more than the realism. "Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models provide the complex reasoning needed to help maintain a simulation's coherence and continuity." When a user types a prompt, Gemini adapts it to the scene, and "between your inputs, Gemini also monitors and steers the scene to keep the world progressing cohesively." In a robotics experiment, Gemini 4 Argon broke a task into steps that Genie 4 simulated, showing Apptronik's Apollo humanoid playing checkers.
DeepMind describes this as an early form of "hybrid world models, which combine the logical, structured reasoning of code and language models with the rich, sensory representation of pixel-level simulation." Tejas Kulkarni, a research director at Google, told TechCrunch: "What we'll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time".
Uses and limits
Trusted testers are using it for film pre-visualisation, turning archival photographs into explorable scenes, game prototyping and scenario generation. DeepMind says the stated goal includes "developing training environments for robots and agents". All generated video and audio carries SynthID watermarks.
The limits are stated plainly. "Visual details and object permanence can still shift or drift over time—especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene." Complex physical interactions, audio accuracy, legible text, spatial accuracy and long-term consistency remain open problems.
What remains uncertain. There is no technical report, no benchmark and no public measure of physical accuracy. The two-minute figure comes from a press briefing. Nothing has been published on latency, resolution, compute cost or API terms. It is also unclear how much of the coherence comes from Genie and how much from Gemini's steering.
Analysis: A planner model that watches a generator and corrects it is the same pattern agent builders use when a supervisor checks a worker. For anyone thinking of training or testing agents in generated worlds, the gap is evidence. A simulation that looks right is not the same as one whose physics is right. Before using a world model as a test environment, measure how far its dynamics drift from reality on the specific interactions your agent depends on. With drift admitted beyond short sessions, short, scripted scenarios are a safer use than long open-ended training runs.
Sources and evidence
Google DeepMind's Genie 4 (research preview, October 27) generates real-time interactive worlds with native audio and mid-simulation text steering, and uses Gemini models to plan and monitor scene coherence, an early hybrid world model; DeepMind admits drift and object-permanence failures over longer sessions, and there is no technical report or benchmark yet.