Google DeepMind begins Genie 4 preview with native audio

Google DeepMind began a research preview of Genie 4 with trusted testers on Oct. 27. It says the world model can generate and steer interactive simulations with native audio in real time.

The model is not generally available. Google said Project Genie access for Google AI Ultra subscribers is coming soon, along with access for early-access partners through the Gemini API and Google AI Studio. It did not provide a release date, Genie-specific pricing or API terms.

A world model generates a simulated environment and predicts how it changes as a user acts. Google describes Genie 4 as its first real-time world model to produce audio natively alongside the scene, including ambient sound, music and interactive dialogue. The company says users can navigate with keyboard and camera controls and change a running simulation with natural-language prompts.

Google’s announcement shows recordings of its own real-time interactions, not independent hands-on tests. The material provides no public benchmark table, evaluation method or numerical performance results establishing Google’s claims about realism, physics, human movement, audio or causal consistency.

Google said Gemini Omni helps ground simulations in real-world appearance and behavior, while other Gemini models support continuity. Gemini can also adapt a user’s prompt to the current scene and monitor the simulation between inputs. The company said Gemini 3.8 Flash can serve as an in-loop guide that watches frames, breaks a goal into steps and checks progress.

Google also disclosed material limits. It said Genie 4 can preserve a coherent environment during short sessions, but visual details and object permanence can drift during longer runs, after several complex prompt edits or as a user moves through a scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency remain active research areas.

The broader world-model category includes NVIDIA’s Cosmos tools for world-model developers. For now, Genie 4 access remains limited to Google’s current testers and the subscriber and partner routes it says are coming soon.

The company said it has paired Genie 4 with Gemini 4 Argon in robotics-oriented experiments, including a demonstration of Apptronik’s Apollo humanoid playing checkers. Google presented that work as an experiment, not a deployed robotics capability.

Google said it conducted internal safety evaluations and added guardrails before limiting the initial preview to trusted partners, but it did not publish independent safety results in the announcement. It also said all video and audio generated by Genie 4 will carry SynthID, its imperceptible watermarking technology.