Google's Genie 4 Doubles Interactive Worlds to Two Minutes and Adds Native Audio

Google announced Genie 4 on Monday, the newest generation of the world model family built by Google DeepMind. Researchers Raia Hadsell, Diego Rivas and Hang Qi describe it as a leap in realism and real-time interactive simulation. Unlike video generators that render a clip from a prompt, Genie 4 runs a persistent, explorable environment that keeps responding while the user acts inside it.

The release lists four headline upgrades. Worlds look more realistic and obey a stronger physical model, covering scenarios like aerodynamics and collisions. Users can steer a running simulation with plain language, changing the weather, moving the camera or introducing a new person or object, and the system adapts the instruction to the surrounding scene. Genie 4 is also the first real-time world model from Google to generate audio natively, from ambient wind and waves to interactive dialogue in which characters converse and react on the fly. And human characters now move naturally, responding organically to their environment.

Under the hood, the model leans on the Gemini family. Gemini Omni anchors the simulation in real-world appearance and behavior, while Gemini 3.8 Flash works as a guiding agent in the loop: it continuously observes incoming frames, breaks the user goal into steps, verifies that each step succeeded and re-plans when the world or the user intervenes. Google is also experimenting with Gemini 4 Argon and Genie 4 alongside Gemini Robotics, demonstrated with Apptronik's Apollo humanoid playing checkers as Argon decomposes the task into steps for the simulation to render.

The hard numbers are modest but doubled: interactive sessions now run about two minutes, up from roughly one minute for Genie 3. Google is candid that this remains the frontier. As simulations run longer, details drift, objects may not stay where the user left them, complex physical interactions and readable text remain unsolved, and long-horizon consistency is still the core challenge for the entire category.

Availability is staged. Genie 4 launches today as a research preview for trusted testers, with a waitlist open for Project Genie. It will reach Google AI Ultra subscribers soon, followed by early-access partners of the Gemini API and Google AI Studio. No parameter count, benchmark scores or pricing have been disclosed, so for now the realism claims rest entirely on demonstrations produced by Google itself.

The competitive backdrop is crowded. Nvidia has open-sourced Cosmos3-H-Dreams, a surgical world model trained on real operating room footage, VAST shipped EdenLive, a 5B real-time world model running above 24 FPS at 720p, and Odyssey drove a car, ran a humanoid and played GTA V from a single prompt. DeepMind frames world simulation as a crucial aspect of building AGI: safe, infinitely variable environments for training robots and agents before anything touches the real world.

What it means: real-time steering and native audio move world models from watchable video to playable scenes, which is closer to what filmmakers, game studios and robotics labs would actually pay for. But the two-minute ceiling and Google's own admission of drift show that long-horizon consistency, not raw fidelity, is the wall between demo and product.