Google ships Genie 4 world model with real-time audio in Project Genie

Google ships Genie 4 world model with real-time audio in Project Genie

Google has announced a new Project Genie experience, available from today to eligible Google AI Ultra subscribers. Project Genie was introduced earlier this year as an experimental research prototype for generating and exploring interactive worlds. Google says early testers wanted deeper control over what happens inside those worlds and easier ways to build and share them. The update delivers four things: first-time access to Genie 4, upgrades to Genie 3, a Project Genie agent that helps users build their vision, and a community gallery to discover, explore and remix worlds. Everything runs in the browser, and Project Genie pairs Google's world models with its Gemini models.

Genie 4 is described as Google's most realistic and interactive world model yet. Users can now type a new prompt while navigating a world and edit the environment in real time: shift the weather, drop in a character, or redirect the action. Google also says that for the first time Genie 4 pairs its visuals with real-time native audio that adapts as the scene unfolds, from ambient soundscapes to character dialogue.

Genie 3 stays in Project Genie's model picker for people who want longer runs or panorama-seeded spaces, while Genie 4 offers the highest-fidelity visuals, physics and native audio. Genie 3 gets three updates. Extended sessions let users explore for up to 3 minutes per generation, up from 60 seconds, and world chaining links multiple worlds into longer journeys. 360-degree panorama image grounding lets users seed new worlds from panoramas. Real-time in-world events let users change the environment with natural-language prompts while exploring.

The new Project Genie agent goes beyond a single text or image prompt. A user describes a vision in natural language, and the agent orchestrates Google's suite of multimodal models to produce initial visuals that set the world's look, plus custom background scores tailored to the mood. Gemini models can also be brought into a session: Gemini observes what is on screen, offers context-aware suggestion chips for where to go next, or triggers in-world events automatically.

The new gallery holds curated worlds and community creations. Examples named by Google include "Living Memory", which turns personal photos into explorable spaces; "Dino Field Guide", set in prehistoric habitats grounded in 360-degree panoramas; and "Rock Paper Scissors", where themed objects face off on a desk with Gemini as real-time referee. Two collaborations with creators who had early access are also featured. Creative technologist Will aka BLVCKL!GHT worked with Google on "Villa Giallo", a room-by-room mystery in a gothic manor. Filmmaker Anand Gandhi worked on "The Life of Ants", where players navigate a colony from an ant's perspective, drawing on Genie 4's photorealism. Users can also remix gallery experiences with the agent, changing the world structure, soundtrack or rules, and publish their own creations to the gallery.

On safety, Google says all video and audio generated in Project Genie carries SynthID watermarking and that all model inputs and outputs go through safety screenings. It also says Genie 3 and Genie 4 remain world models in the frontier research space, and that it is still working on camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence. Google AI Pro members can join a waitlist on the same page.

Key facts

  • Genie 4, which Google calls its most realistic and interactive world model yet, reaches eligible Google AI Ultra subscribers through Project Genie starting today; Google AI Pro members can join a waitlist.
  • Genie 4 lets users prompt mid-session to edit the world in real time, and Google says it is the first to pair visuals with real-time native audio, from ambient sound to character dialogue.
  • Genie 3 stays in the model picker with extended sessions of up to 3 minutes per generation (up from 60 seconds), world chaining, 360-degree panorama seeding and real-time in-world events.
  • A new Project Genie agent builds worlds from a natural-language description, with initial visuals and custom background scores, and Gemini can suggest next steps or trigger events.
  • A community gallery lets users explore, remix and publish worlds; all output carries SynthID watermarking, and Google lists camera latency, physical interactions, audio precision and scene permanence as open problems.

Why it matters

World models that generate explorable, interactive environments are a frontier research area, and this release moves Genie from a research prototype toward something people can build in and share. Google presents Genie 4 as its most realistic and interactive world model yet. The headline additions are mid-session prompting, which lets a user change a world while moving through it, and real-time native audio, which Google says Genie 4 offers for the first time. Pairing a world model with the Gemini models and an agent also shows Google treating world creation as a conversation rather than a single prompt.

Who it affects

Eligible Google AI Ultra subscribers get the new experience today. Google AI Pro members can only join the waitlist. Creators and hobbyists who want to make interactive scenes in the browser are the intended audience, and the gallery adds a social layer of remixing and publishing. Google named a creative technologist (Will aka BLVCKL!GHT) and a filmmaker (Anand Gandhi) as early collaborators.

How to use it

Open Project Genie in the browser; Google says the Ultra experience starts today and the Pro waitlist is on the same page. Pick Genie 4 for the highest-fidelity visuals, physics and audio, or Genie 3 in the model picker for longer runs (up to 3 minutes per generation) or worlds seeded from 360-degree panoramas. Describe a world to the Project Genie agent, or start from a prompt, then type new prompts as you explore to change the scene. You can also remix gallery worlds and publish your own. The source gives no price for either subscription tier, and it does not say which countries are eligible.

How solid is it

This is a first-party announcement on Google's own blog, so every claim is Google's own description; the post quotes no outside testers or reviewers. It gives no benchmarks, no comparison numbers between Genie 4 and Genie 3, and no technical specifications such as resolution, frame rate or latency. The only concrete figure is the Genie 3 session limit of up to 3 minutes, up from 60 seconds. The post does not state how long Genie 4 sessions run. It points to a separate Genie 4 model research overview for limitations.

Risks and caveats

Google itself calls Genie 3 and Genie 4 world models in the frontier research space and says it is still working on camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence. Access is limited to eligible Ultra subscribers, with Pro members on a waitlist and no timeline given for wider access. On safety, Google says all video and audio carries SynthID watermarking and all model inputs and outputs go through safety screenings. Because users can publish worlds to a community gallery, that is the layer where those safeguards apply.

“Genie 3 and Genie 4 remain world models in the frontier research space.”

— Google, Genie 4 announcement on the Google blog