Google DeepMind on November 18 began rolling out Genie 4, its interactive world model, to eligible Google AI Ultra subscribers in Project Genie, with a waitlist for Google AI Pro members.

Genie 4 lets users change a generated world by typing prompts while they move through it, and it generates audio, including character dialogue, alongside the visuals. Every capability and safety claim below comes from DeepMind's own pages, which cite no independent testing.

The November 18 blog post1 announces the consumer rollout. DeepMind's separate technical overview2 carries no date. It describes Genie 4 as a research preview for trusted testers and says the model is "coming soon" to Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. The overview therefore predates the rollout, and the blog post does not say when the Gemini API and AI Studio access will begin.

According to the blog post, users can shift weather, add characters or redirect action mid-session, and hear ambient sound that adapts as the scene changes. Project Genie runs in a browser, and the Project Genie page3 says it is powered by Genie 4. Access is limited to eligible Ultra subscribers.

The technical overview says Gemini Omni anchors simulations in real-world appearance and that Gemini 3.8 Flash acts as a guiding agent, observing the scene and adjusting its plan as the user intervenes. DeepMind says this produces worlds with "reliable causality."

The update also changes Genie 3, which stays in the model picker. Per the blog post, sessions can run up to 3 minutes per generation, up from 60 seconds, and multiple worlds can be linked. Worlds can also be seeded from 360-degree panorama images. DeepMind's demonstration says panorama grounding "keeps Petra's canyon and facade consistent as you look around."

A new Project Genie agent builds worlds from a conversation instead of a single prompt, and generates opening visuals and a background score. A community gallery lets users publish and remix worlds. DeepMind says "Villa Giallo" and "The Life of Ants," made with outside creators who had early access, use Genie 4's photorealism. They are DeepMind-curated showcases, and the company's pages offer no independent evaluation of the worlds.

Limitations DeepMind lists

The technical overview calls Genie 4 "an early, experimental model." It says visual details and object permanence can shift or drift, especially in longer runs, after multiple complex prompt edits or while navigating the scene. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are described as areas of ongoing research.

On safety, the blog post says all video and audio in Project Genie carries SynthID watermarking and that all model inputs and outputs go through safety screening. The overview says internal safety evaluations took place "before launching this preview," meaning the trusted-tester release rather than the November 18 rollout. No independent assessment of these measures is cited.

The use cases come from different parts of the overview. DeepMind itself frames world simulation as useful for developing training environments for robots and agents. Prototyping uses for filmmakers and game developers are described as things trusted testers are trying. DeepMind names camera latency, long-term scene permanence and hybrid world models, which combine code and language-model reasoning with pixel-level simulation, as continuing work.

The announcement does not state price, regional availability or what makes an Ultra subscriber "eligible."