Google DeepMind has opened Genie 4, its newest world model, to users of Project Genie. The rollout started on November 18, 2026, for eligible Google AI Ultra subscribers. Genie 4 generates interactive worlds that users can walk through, change with text prompts and, for the first time, hear.
The update also brings a creative agent, a community gallery and an upgraded Genie 3. Google AI Pro subscribers can join a waitlist. DeepMind first showed Genie 4 on October 27, when it went only to trusted testers.
What Genie 4 adds to Project Genie
The new model is a world model, a system that generates a 3D scene frame by frame as the user moves through it. DeepMind calls it its "most realistic and interactive world model yet". It builds on Gemini models: Gemini Omni anchors how the world looks and behaves, and other Gemini models help keep the simulation consistent.
The main change is editing on the fly. While exploring, users can type a new prompt to shift the weather, add a character or redirect the action. The world changes without starting over.
Genie 4 also generates sound in real time. Google says the audio adapts as the scene unfolds, from background ambience to character dialogue. Earlier Genie models produced silent worlds. Genie 3 got an upgrade too.
What changes in Genie 3
The older model stays in Project Genie's model picker for users who want longer runs. Google lists three updates:
- Longer sessions: up to 3 minutes per generation, up from 60 seconds, plus chaining several worlds into one journey
- Panorama grounding: new worlds can start from a 360-degree panorama image
- Live events: text prompts can change the world while the user explores
The split is clear. Genie 4 has the best visuals, physics and audio, while Genie 3 runs longer. DeepMind's own page says Genie 4 currently supports "around a minute" of continuous interaction. The new agent is meant to make building these worlds easier.
How the Project Genie agent and gallery work
The Project Genie agent turns a conversation into a world. Users describe what they want. The agent then uses Google's other models to design the look of the world and compose a score that fits the mood.
Gemini can also join a session as a companion. It watches the screen, suggests where to go next and can trigger events in the world on its own.
The new gallery mixes curated and community worlds. Users can explore them, remix them with the agent and publish their own. Google's examples include:
- Living Memory, which turns personal photos into explorable spaces
- Dino Field Guide, with prehistoric habitats built from 360-degree panoramas
- Villa Giallo, a room-by-room mystery in a gothic manor, made with creative technologist BLVCKL!GHT
- The Life of Ants, made with filmmaker Anand Gandhi, which shows a colony from an ant's point of view
All video and audio from Project Genie carries Google's SynthID watermark. DeepMind is also open about what the model still gets wrong.
Genie 4 limits and what DeepMind still works on
DeepMind calls Genie 4 "an early, evolving technology". Its model page lists several known problems:
- Drift: details and objects can shift over time, especially in long runs or after several complex edits
- Short sessions: about a minute of continuous interaction
- Other agents: complex interactions between users and several characters are still hard
- Real places: the model can't recreate real locations exactly
- Text: clear text usually appears only when it is written into the prompt
- Audio: real-time sound accuracy is still being improved
Google's blog adds camera latency to the list. Genie 3 ran at 720p and 24 frames per second, and Google has not published resolution or frame rate figures for Genie 4. The model matters beyond games, though.
Why Genie 4 matters beyond games
DeepMind pitches Genie 4 for filmmaking, learning, spatial design and training AI agents such as SIMA. Its model page shows an early robotics test. Gemini 4 Argon writes a step-by-step plan for a task, and Genie 4 renders the steps as the simulation runs. DeepMind calls this "an infinite playground" for robot training and a step toward AGI.
The Genie line already has outside users. Waymo built its Waymo World Model on Genie 3 to simulate rare road situations for its robotaxis. When Project Genie first launched in January, shares of several video game companies fell on fears that AI could replace parts of game development.
Replies to DeepMind's post were mostly upbeat, with users praising Google's run of launches. Many asked instead about Gemini 4 Neon, a model fans expect next. Others asked how long a Genie 4 world holds together after several edits. That is the key question for a model limited to about a minute.
Is Genie 4 worth trying
For Google AI Ultra subscribers, yes. Live editing and native audio are a real step from Genie 3's silent, fixed worlds. The agent and gallery also make it easy to start without long prompts. It costs nothing extra on top of the subscription.
For everyone else, Genie 4 is still a research preview behind the most expensive Google plan, with a waitlist for Pro users. Sessions last about a minute, and worlds drift after edits. It is a glimpse of interactive AI worlds, and the real test will be how far DeepMind can stretch session length and consistency.
Sources
- Genie 4 is now available, Google – November 18, 2026
- Genie 4, Google DeepMind – November 2026
- Google DeepMind post on X – November 18, 2026