# Genie 4 Adds Native Audio and Gemini Reasoning to DeepMind's Interactive World Model: What It Is, and Why It Is Not a Shot Generator - Published: 28 October 2026 - Updated: 28 October 2026 - Author: RenderBob team - Category: News - Canonical: https://renderbob.ai/blog/genie-4-deepmind-world-model-native-audio-filmmakers Google DeepMind announced Genie 4 on 27 October, a research-preview world model that generates explorable environments with synchronised sound. DeepMind publishes no resolution or session length. Google DeepMind announced Genie 4 on 27 October, a research-preview world model that generates explorable environments with synchronised sound. It is early and experimental, and DeepMind publishes no resolution or session length. ## What was announced DeepMind's own post was not reachable for us, so this account comes from two secondary reports, OfficeChai and AI Weekly, both dated 27 October. They say Genie 4 builds an interactive, navigable environment from a text prompt or reference image, just over a year after Genie 3. While a simulation runs, a user can change the weather or camera angle, or add objects and characters, with natural-language prompts, and characters respond in real time. Physics, collisions and human motion are described as more convincing than before. AI Weekly calls it DeepMind's first real-time world model to generate audio natively, covering music, ambient sound and interactive dialogue. It says Gemini Omni grounds scenes in real-world appearance and behaviour, and that Gemini 3.8 Flash acts as a guiding agent that watches the scene as it is generated. Raia Hadsell, Diego Rivas and Hang Qi are credited. ## Access and safeguards It is a research preview for trusted testers. Next come Google AI Ultra subscribers through Project Genie, then early-access partners on the Gemini API and Google AI Studio, with a waitlist and broader access planned later. Google says outputs carry SynthID watermarks and that the model went through internal evaluations to guard against harmful outputs. ## What DeepMind says it cannot do yet Both reports quote the same caution. DeepMind calls it an early, experimental model. Visual details and object permanence can shift or drift over time, especially in longer runs, after several complex edits or while moving around a scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are still under research. No resolution, frame rate, inference cost or maximum session length has been published. ## Where it fits for a filmmaker Our 14 October piece on Odyssey-3 looked at what interactive world models are, and are not yet good for, in production. Genie 4 moves the same direction with sound and better motion, but the same distinction holds. A film needs a shot that stays identical when you cut to it twice. A live, steerable simulation does not promise that, and neither report says whether a session can be exported as a clip. Possible uses are exploring a location or blocking a scene before you generate the shots you will cut. Questions to ask when access opens: - Can you export a run as video, with its audio? - Does the same prompt give the same space twice? - What are the terms for commercial use? - Does the SynthID mark survive export and editing? ## What to watch DeepMind's own post with technical details, published resolution and session length, export options, and when Ultra subscribers and API partners get access. ## RenderBob 3 and this story RenderBob 3 is built for fixed, editable shots rather than live simulation. Each generated clip lands on its own lane, with the model your choice per clip, local or cloud, so you can explore a space elsewhere and still make the final shots as clips you can cut, review and re-roll.