Genie 4 adds sound and talking characters to Google’s steerable world model

Google DeepMind has introduced Genie 4, a world model that generates an interactive environment in real time and, for the first time in the Genie line, produces its own sound. In its announcement on X, Google called it its “most realistic and interactive world model yet”. It is a research preview for trusted testers only.

Users steer the scene with natural language: they can change the weather, move the camera or add new elements while the world keeps running. According to Google, Gemini acts as the “brain” of the simulation, turning requests into actions, planning complex events and tracking what happens on screen.

Gemini steers the scene and the soundscape

DeepMind’s blog post describes a division of labor. Gemini Omni anchors the simulation in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a scene coherent between inputs. Navigation uses standard keyboard controls, with text prompts layered on top.

DeepMind says Gemini 3.8 Flash can also act as a guiding agent inside the world. It breaks a goal into steps, checks frames to see whether each step is done, and adapts the plan. The post adds that Gemini 4 Argon was used in experiments pairing Genie 4 with Gemini Robotics, including a demo in which Apptronik’s Apollo humanoid plays checkers.

Audio is the headline addition. DeepMind calls Genie 4 its “first real-time world model to generate audio natively”, covering ambient noise, music and interactive dialogue. Characters can talk with the user and react to events as they happen. The company also says the model simulates physics such as aerodynamics and collisions, and that characters move and respond to steering.

Trusted testers first, Ultra subscribers next

For now, only a limited group of trusted testers has access. Google says Genie 4 is coming soon to Google AI Ultra subscribers in Project Genie, and to early access partners in the Gemini API and Google AI Studio. DeepMind has opened a waitlist for Project Genie.

The sources give no release date, price, supported regions, resolution, frame rate or maximum session length. Those are the details that decide whether developers can build on it, and they are missing.

DeepMind lists tester use cases including pre-visualization, game prototyping, enterprise scenario generation and turning a single archival photo into a simulation. It also names training environments for robots and agents, and describes simulating the world as a crucial step toward AGI. These are the company’s own framing, not independent findings.

On safety, DeepMind says Genie 4 has strict guardrails consistent with its policies for other models, went through internal evaluations, and that all generated video and audio carries SynthID watermarking.

Genie 3 had no sound and a short memory

The clearest way to read Genie 4 is against its predecessor. When DeepMind introduced Genie 3 in August 2025, it described worlds navigable at 720p and 24 frames per second. They stayed consistent for a few minutes, with visual memory reaching back about one minute. Access was a limited preview for a small cohort of academics and creators, and the page did not mention audio.

Genie 4 keeps the same cautious rollout but changes what the model is. Sound and dialogue turn a silent, navigable video into something closer to a scene with people in it, and Gemini’s role as planner suggests DeepMind wants the world to be directed, not just explored. What the sources do not establish is whether the technical limits moved. DeepMind’s Genie 4 page does not compare the two models, and it still concedes that visual details and object permanence can shift or drift over time in longer sessions or after several edits. It lists complex physical interactions, real-time audio accuracy, legible text and long-term consistency as open research areas.

The release also fits a pattern of Google tying its world and robotics work to the Gemini family. Earlier today the company opened Gemini Robotics ER 2 in preview on Google Cloud, and it has been showing what Gemini 4 Argon can do in practice. The checkers demo links those threads: a simulated world as a place to test agents before they act in physical ones.

What readers can and cannot check yet

Everything about Genie 4’s realism, responsiveness and audio quality currently comes from Google’s own demos and posts. No outside group has published tests, and the preview is closed. Anyone hoping to try it needs to be a selected tester, join the Project Genie waitlist, or wait for an Ultra or API rollout that has no date.