AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering

Collected Oct 27, 2026

Google DeepMind released Genie 4, its newest world model, in research preview on Tuesday. The company says the model produces the most photorealistic and interactive simulated worlds it has built so far. Users can type prompts mid-simulation to change the scene, and the model generates real-time audio matched to the simulated world, including music, ambient noise and interactive dialogue.

The jump from Genie 3, which arrived as a research preview a little over a year ago, is broad. Genie 4 holds a consistent scene for up to two minutes, double its predecessor's limit. It shows improved physics and more humanlike characters. It also accepts free-form edits during play, such as changing the weather or camera angle in real time or adding new people and objects.

The release lands two months after AI video startup Runway shipped GWM Worlds 2, a playable world model that responds dynamically to user inputs and generates audio. The surrounding field is consolidating: World Labs, the startup led by Fei-Fei Li, is being acquired by AMD, Amazon is in talks to buy Decart, and Rivian has acquired Third Dimension to bring simulation in-house for self-driving testing. Google's release reads as an effort to hold its lead in that crowded field.

Gemini, Google's AI model, is more tightly woven into Genie 4. In this setup Gemini acts as a director: it plans steps and checks frames while Genie produces the visuals. In a research experiment tied to Gemini Robotics, a Gemini 4 Argon model broke tasks into steps that Genie 4 then simulated in real time. That points to a possible workflow for testing robot scenarios in simulation before trying them in the physical world.

Tejas Kulkarni, a research director at Google, described the direction during a briefing: hybrid world models where Gemini writes code and text while Genie ingests it in real time and simulates reality at varying levels of abstraction.

World models are generative systems that predict how an environment evolves and render that evolution frame by frame, rather than playing back stored footage. That distinction matters for developers: a world model can respond to input that never existed at training time, which is what makes free-form editing and real-time steering possible. The catch is that consistency and physical plausibility degrade over long rollouts, a weakness Google's team acknowledges.

Google pitches the usual commercial use cases for world models: gaming, interactive learning and robotics. One speculative example of interactive learning would be stepping into a painting of a historical scene and conversing with its subjects about the event shown.

Trusted testers, the people who get early access to Google's cutting-edge products, are flagging a different use: prototyping films. Diego Rivas, a Google DeepMind engineer, said during a press briefing that the real-time aspect is most valuable where keeping a creative state of flow matters, such as ideation and storyboarding, or any moment calling for rapid ideation.

Google also sells Veo, its AI video-generation model that produces high-definition clips with synchronized audio from text, image or video prompts, and which has been gaining popularity among filmmakers. Rivas said users could eventually lean on Gemini Omni, a multimodal large language model that can generate video from images, audio and text, to turn those prototypes into finished creative assets, though he said those capabilities remain a ways off.

Rivas repeatedly stressed that it is early days for Genie 4 and that the end-to-end product experience is not yet clear. The team also acknowledged the model's gaps: detail and objects can drift as a simulation runs longer, and it has not yet solved how to guarantee accurate audio or legible in-world text. For anyone building on top of this, those are the constraints to design around. Drift over a two-minute window suggests that long-lived scenes, such as persistent game levels or extended robot training rollouts, still need external state management or periodic correction rather than trusting the model to hold the world steady on its own. Unreliable in-world text also limits uses like signage, UI mockups or readable labels inside a simulated scene.

Availability is staged. Genie 4 is open now as a research preview for trusted testers. In the coming weeks it will reach Google AI Ultra subscribers through Project Genie, Google's tool for generating interactive worlds from text or image prompts, and early access partners of the Gemini API and Google AI Studio. That sequencing suggests a gradual path from internal testing to subscriber access before any broader developer offering, which is a likely trade-off given the acknowledged quality gaps.

Why it matters: developers building games, training environments or film previsualization tools now have a Google world model that accepts mid-run prompts and produces synchronized audio, and the Gemini integration hints at pipelines where a planning model directs a simulator step by step. Access is limited at first to trusted testers, then AI Ultra subscribers and select API and AI Studio partners, so most teams will be watching rather than building. The acknowledged drift, audio reliability and text legibility issues are, by inference, the problems that will decide how quickly world models move from demos into production workflows. And the consolidation around World Labs, Decart and Third Dimension signals that simulation capability is becoming something large platform and hardware companies want to own outright.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Genie 4 can maintain a consistent scene for up to two minutes, double Genie 3's limit, and offers better physics and more humanlike characters.