DeepMind's Genie 4 World Model Goes Live for AI Ultra With Native Audio and Remixing
On this page 7 sections
TL;DR
Genie 4, Google DeepMind's model for simulating interactive worlds in real time, left research preview on 18 November and is live now for eligible Google AI Ultra subscribers in Project Genie. You prompt a 3D world into existence, walk around it with keyboard controls, and retype it from the inside: shift the weather, drop in a character, redirect the action. Worlds now carry native, adaptive audio for the first time, a conversational agent helps build them, and everything you make can go into a community gallery where other people can remix it.
What went live
Project Genie is the browser app, and its model picker now holds both Genie 3 and Genie 4. Genie 4 gives the highest-fidelity visuals, physics, and native audio, and you steer it with text while you navigate. Genie 3 stays in the picker for longer runs and panorama-seeded spaces, and it picked up three upgrades in the same update: sessions of up to three minutes per generation, up from 60 seconds; world chaining, which links separate worlds into one journey; and seeding from 360-degree panorama photos.
The new Project Genie agent takes creation past the single prompt. Describe your vision in conversation and it orchestrates Google's multimodal models to build the initial look plus a background score matched to the mood. Gemini can also ride along in the session, watching the screen, suggesting where to go next, and triggering in-world events as the journey unfolds.
The gallery ships with curated worlds: Living Memory turns personal photos into explorable spaces, Dino Field Guide wanders prehistoric habitats grounded in panoramas, and Rock Paper Scissors pits themed objects against each other with Gemini refereeing in real time. Two launch collaborations show the range. Villa Giallo, built with creative technologist Will aka BLVCKL!GHT, is a room-by-room mystery in a gothic manor. The Life of Ants, made with filmmaker Anand Gandhi, navigates an ant colony from an ant's eye view. Published worlds can be remixed by anyone, right down to the world structure, soundtrack, and rules.
Under the hood
Genie 4 is natively multimodal: music, ambient noise, and dialogue generate alongside the visuals, which DeepMind says is a first for its real-time world models. The model leans on the Gemini family for its grounding. Gemini Omni anchors simulations in real-world appearance and behavior, and other Gemini models supply the reasoning that holds a simulation together across time.
Gemini 3.8 Flash runs as a guide in the loop. As Genie renders, it watches the scene, breaks goals into steps, reviews incoming frames, and checks that each step landed before planning the next; when you intervene, it re-plans on the spot. Picture a continuity editor who never blinks, nudging the set dresser while the film keeps rolling.
The robotics tie-in is explicit. DeepMind is also pairing Gemini 4 Argon with Genie 4: Argon decomposes a task into steps, feeds them to Genie as text captions mid-render, and the simulation adapts toward finishing the task. Their demo has Apptronik's Apollo humanoid playing checkers inside a generated world, part of the bridge to Gemini Robotics and DeepMind's stated interest in training environments for robots and agents.
What the testers did with it
Before this rollout, DeepMind gave cohorts of trusted testers access across creative industries, journalism, education, gaming, and engineering. The uses it published are more concrete than the usual launch demo reel:
- Film crews step into a live simulation of a scene and run multiple takes, changing camera angle, lighting, dialogue, and action as it plays. Pre-viz you can physically walk through.
- Photojournalists turn a single archival photograph into a grounded simulation in seconds, then test how events like a flash flood or a wildfire would unfold in that space.
- Game developers go from an idea to an explorable prototype in seconds, testing visual styles, materials, and mechanics by prompt before writing a line of engine code.
- Enterprise partners are evaluating it as a rapid generator of zero-to-one scenarios.
The catches
DeepMind's own limitations section is unusually blunt. Visual details and object permanence drift over longer sessions, after multiple complex edits, and while navigating the scene. Complex physical interactions, real-time audio precision, legible text, spatial accuracy, and long-term consistency are all listed as open problems. On Hacker News, where the announcement drew 190 points and 143 comments, one commenter noticed the demo footage avoids panning away from an object and back, which is exactly the move that breaks permanence.
Export is the other gap. Requests for FBX, USD, OBJ, or Gaussian-splat exports ran through the thread, and none is announced. Your world lives in the gallery or nowhere.
Access is messy too. The offer is "eligible Google AI Ultra subscribers," and eligibility is hard to pin down: access complaints dominate the HN thread, from people unable to work out whether their account, country, or workspace qualifies. Google AI Pro members can join a waitlist. API access through the Gemini API and AI Studio is early access partners only for now. Everything generated carries SynthID watermarking, and inputs and outputs pass safety screenings.
Why a builder should care
- Game concept to explorable prototype, with no engine code, is now a subscription feature. Styles, materials, and mechanics get tested by prompt before you commit to an engine.
- Film pre-viz becomes walk-in. Directors can stand inside takes and adjust camera, lighting, and dialogue live instead of describing shots over storyboards.
- The gallery is a distribution channel. Publishing a world with its structure, soundtrack, and rules editable by remakers is the UGC loop Roblox spent a decade building, except the worlds are prompts.
- Conversational NPCs arrive early. Gemini already watches sessions, suggests moves, and triggers events, so dialogue-driven world logic needs no custom stack.
- Treat permanence as unstable. Build short experiences, chain them the way Genie 3 now does, and keep anything that must persist in your own code.
- Watch the Gemini API and AI Studio partner track. When worlds become an API surface, agents get a rehearsal space. DeepMind's overview points the same way.
Key Takeaways
- Genie 4 went live on 18 November 2026 for eligible Google AI Ultra subscribers in Project Genie, 22 days after its 27 October research preview; Google AI Pro members can join a waitlist.
- You can steer a world with text mid-session, and audio now generates natively with the visuals, ambient sound and dialogue included, a first for DeepMind's real-time world models.
- Gemini 3.8 Flash acts as an agent in the loop, observing frames, planning steps, and verifying progress; a Gemini 4 Argon pairing already simulates Apptronik's Apollo humanoid playing checkers.
- Genie 3 stays in the model picker with sessions of up to 3 minutes per generation (up from 60 seconds), world chaining, and 360-degree panorama grounding.
- Trusted testers used Genie 4 for film pre-viz with live takes, turning archival photos into simulations, game prototyping before engine code, and enterprise scenario generation.
- Admitted limits: object permanence drift, complex physical interactions, audio precision, legible text, and spatial accuracy. All output is SynthID watermarked, and no export format is announced.
Sources: Google announcement, DeepMind Genie 4 research overview, Project Genie, The Decoder, Hacker News thread