māchine · world-models

Google DeepMind opens Genie 4 world model to AI Ultra subscribers

Project Genie now lets paying users build interactive worlds with Genie 4, edit them by typing mid-session, hear generated audio, and share them in a community gallery.

Google DeepMind is rolling out Genie 4 to eligible Google AI Ultra subscribers through Project Genie, its browser-based tool for generating and exploring interactive worlds. Users can type a prompt while they move through a world and the scene changes in real time. For the first time, the visuals come with native audio generated alongside them. Google AI Pro members can join a waitlist.

The release matters because world models have so far been mostly research demonstrations. Genie 4 was announced on 27 October 2026 as a preview for trusted testers. Three weeks later it is part of a paid consumer product, with an agent that helps build worlds and a gallery for publishing them. DeepMind still describes both Genie 3 and Genie 4 as "world models in the frontier research space".

What users can generate

A world model produces an environment frame by frame, responding to the user's movement and input, instead of rendering a fixed video. According to DeepMind's announcement, Genie 4 lets users "shift the weather, drop in a character, or redirect the action" by typing a new prompt during a session. The audio adapts as the scene unfolds, ranging from ambient sound to character dialogue. DeepMind calls Genie 4 its "most realistic and interactive world model yet" and says it has the company's highest-fidelity visuals, physics and audio.

Genie 3 remains in the model picker. DeepMind says users should choose it for longer runs or spaces seeded from a 360-degree panorama. Updates to Genie 3 include longer runs, panorama grounding that keeps a location consistent as the camera turns, and real-time promptable events. DeepMind gives examples: Petra's canyon and facade staying stable as the user looks around, and a dragon summoned on command during a flight.

The new Project Genie agent takes a conversational description and, according to Google, orchestrates its multimodal models to produce the world's initial visuals and a background score matched to the mood. Gemini models can also join a session. They observe the screen, offer suggestions for where to go next, or trigger in-world events automatically.

The gallery and early collaborations

A community gallery collects curated worlds and user creations. Users can explore them, remix them with the agent by changing the world structure, soundtrack or rules, and publish their own. Curated examples include "Living Memory", which turns personal photos into explorable spaces, "Dino Field Guide", built on 360-degree panoramas, and "Rock Paper Scissors", in which Gemini acts as a real-time referee.

DeepMind also worked with creators who had early access. With creative technologist Will, known as BLVCKL!GHT, it built "Villa Giallo", a room-by-room mystery in a gothic manor. With filmmaker Anand Gandhi it built "The Life of Ants", which lets users move through a colony from an ant's perspective.

All video and audio generated in Project Genie carries SynthID watermarking, and Google says all model inputs and outputs go through safety screening.

Beyond entertainment

The consumer release is about creative play, but the October announcement placed Genie 4 in a wider frame. DeepMind said its trusted testers came from creative industries, journalism, education, gaming and engineering. Use cases it named include previsualization, turning archival photos into simulations, prototyping game concepts and scenario generation.

That post also described a robotics experiment. Gemini 4 Argon split a task into steps, which were passed to Genie 4 as text captions while it rendered a simulation of Apptronik's Apollo humanoid robot playing checkers. DeepMind did not publish metrics for the demo. The idea is that a robot's task could be rehearsed in a generated environment before it is tried in the physical world. In the same post, DeepMind described Gemini 3.8 Flash acting as a real-time guiding agent that watches frames, breaks goals into steps and checks whether each step succeeded.

Developer access is not open yet. In October, Google said Genie 4 was "coming soon" to early-access partners of the Gemini API and Google AI Studio. Neither post gives pricing, resolution, frame rate or benchmark results.

Known limits

DeepMind lists the open problems directly. In today's post: camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence. The October post added that "visual details and object permanence can still shift or drift over time", especially in longer runs or after several edits, and that legible text and spatial accuracy remain unsolved. A world that drifts is acceptable for exploration. It is a harder problem for simulation, where the environment has to behave the same way each time.

What it means for work

For game designers, filmmakers and educators, Genie 4 compresses early-stage work: blocking out a scene, testing a mood with sound, or trying a game concept can now start from a sentence and be changed by typing. The gallery's remix model means a prototype can pass between people without anyone touching an engine or asset pipeline. For robotics and engineering teams the Apollo demo points to generated scenario testing, but without published metrics, API access or reliable physics it remains a direction and not a tool. The skilled work that remains is judgement: deciding what a world should contain, noticing where its physics or continuity break, and knowing when a generated simulation is good enough to trust.

Sources

This article was researched, written and illustrated by AI, and published without human editing. Spotted an error? Let us know.