Google DeepMind Releases Genie 4 World Model With Audio, Real-Time Steering, and Realistic Physics

Google DeepMind announced Genie 4 on Tuesday, a new world model that generates interactive real-time simulations with native audio, realistic physics, and human motion, while accepting natural language prompts mid-simulation to steer the scene. The model is available now as a research preview for trusted testers, with a wider rollout planned for Google AI Ultra subscribers and early API partners in the coming weeks [5][4].

What Genie 4 Does

Genie 4 is natively multimodal: it generates synchronized audio—ambient sound, music, and interactive dialogue—alongside video. Users navigate with keyboard controls and type prompts to change weather, move the camera, or add people and objects while the simulation runs. Google says the model maintains a consistent scene for up to two minutes, double the limit of Genie 3 [5][6].

The system integrates Gemini models at two levels. Gemini Omni anchors simulations in real-world appearance and behavior, while Gemini 3.8 Flash acts as a guiding agent that breaks down goals into steps, reviews each generated frame, and adapts the plan when the user intervenes or the world shifts [4][7]. In a research experiment tied to Gemini Robotics, a Gemini 4 Argon model decomposed tasks into steps that Genie 4 simulated in real time, demonstrating a possible pipeline for testing robot scenarios before real-world deployment [5].

Demonstrated Capabilities

The DeepMind blog shows several extended examples: a saxophonist on a floating platform with fog-draped lake ambience; a stream of pink satin fabric billowing through a meadow; a first-person hot-air balloon ride over Antarctica; a watermelon on a stone plinth with slow-motion droplet physics; a maple seed falling into grass; a crowded 1980s Japanese train station; a fully explorable stained-glass world; a first-person dragon flight at 140 km/h with wingbeat audio; a Hong Kong motorcycle scene shot on 35mm film; a UR5 robot dispensing liquid with electric sparks; and a first-person walk through a porcelain-filled antique salon [4].

Google characterizes these as recordings of real-time interactions. They are, by selection, the company's own chosen demonstrations. Google has released only its own selected demonstrations; no independent benchmark results are cited in the announcement.

Limitations Acknowledged

DeepMind engineers acknowledged during briefings that details and objects can drift the longer a simulation runs, and that audio accuracy and in-world text legibility remain unsolved [5]. The Verge reports the company warns that "details and object permanence can shift over time" [6]. Ars Technica notes Genie 4 worlds "still only remain consistent for a couple of minutes before they forget how things started" [7].

Availability and Pricing

Access begins with trusted testers. In the coming weeks, Genie 4 will roll out to Google AI Ultra subscribers (starting at $100 per month) via Project Genie, and to early access partners of the Gemini API and Google AI Studio. Developer pricing for the API has not been announced. Google says it may expand to AI Pro and AI Plus plans eventually, but gave no timeline [7].

Competitive Context

The release follows Runway's GWM Worlds 2, a playable world model with audio generation, by a couple of months [5]. The world-model space is consolidating: AMD is acquiring Fei-Fei Li's World Labs, Amazon is in talks to buy Decart, and Rivian acquired Third Dimension for self-driving simulation [5]. Google also offers Veo, a separate video generation model that produces high-definition clips with synchronized audio from text, image, or video prompts [5].

What Remains Unknown

• Independent tests of physics fidelity, audio sync, and steering precision have not been reported in the available coverage.
• Google's two-minute consistency claim is based on its own demonstrations; independent verification has not been reported.
• The impact of Gemini 3.8 Flash on causal coherence compared to open-loop generation remains to be independently evaluated.
• Developer API pricing and broader consumer availability dates are unspecified.

Genie 4 represents a visible advance in real-time, steerable world simulation with native audio. Its practical utility for robotics training, film pre-visualization, or game prototyping will depend on how well the demonstrated capabilities hold up under sustained, unscripted use.