On 27 October Google DeepMind announced Genie 4, the new version of its world model: a system that builds an explorable three-dimensional environment in real time from a text description. The main novelty is audio, generated together with the images. For now, though, Genie 4 is a research preview, open only to a small group of testers.
What a world model is
Genie doesn’t produce a ready-made video. It calculates the scene frame by frame and reacts to what the user does: you move around with keyboard and mouse, while text is used to change what happens. This is the idea behind Project Genie, the web app Google launched in January for AI Ultra subscribers in the United States. Genie 3, unveiled in August 2025, ran at 720p and 24 frames per second, with worlds that stayed coherent for a few minutes.
What changes with Genie 4
According to DeepMind it is the first real-time world model to generate audio natively: the sound of wind in the trees, music, applause, dialogue between characters. There is no soundtrack added afterwards; the sound is born with the scene.
The second point is control. While the simulation is running you can ask to move the camera, change the weather, or bring in new characters or objects. Behind it is the Gemini family: Gemini Omni provides the foundation that makes the simulation resemble reality, in appearance and behaviour, while other Gemini models adapt the request to the scene in progress. For multi-step events, Gemini 3.8 Flash watches the new frames, checks whether each step succeeded and corrects the plan if something changes.
DeepMind also claims more believable physics (collisions, aerodynamics, water) and people who move naturally and react to their surroundings.
The link with robotics
Google also presents it as a testing ground for robots and agents. In the demo, Gemini 4 Argon breaks a game of draughts down into steps and Genie 4 simulates Apptronik’s humanoid robot Apollo carrying them out. No real robot is involved: it is a test in simulation, meant to find out whether a generated environment can replace costly or risky trials.
Stated limitations
DeepMind calls Genie 4 an “early, experimental” model and lists the limits clearly:
- continuous interaction lasts a few minutes, not long sessions;
- visual details and object permanence can drift in longer simulations or after many edits;
- real places are not reproduced with perfect accuracy;
- legible text often appears only if it is already in the description of the world;
- the accuracy of real-time audio is still being researched.
All generated videos and audio carry the invisible SynthID watermark. Google speaks of safeguards in line with its safety policies and of internal evaluations before release.
Availability and pricing
The preview is open to selected testers; there is a waiting list. Google promises it “soon”, first for AI Ultra subscribers inside Project Genie, then for partners with early access to the Gemini API and AI Studio. No date, price or list of countries has been announced, and Google has not stated Genie 4’s resolution or frame rate. Whether and when it will reach Italy, therefore, can’t be said today.
What we think
Native audio is the most concrete step: a generated world that speaks and makes sound makes the simulation feel like a place rather than a silent clip. A heavy limit remains, though, the few minutes of coherence: as long as objects “drift”, serious uses such as robot training remain promises to be verified, and it is no accident that Google talks about research.
For Italian readers the interest is more indirect. There is nothing to try right now: Project Genie requires an Ultra subscription and, at its January launch, was restricted to the United States. It is worth following the next stage, its arrival in the APIs: that is where developers and game studios will be able to gauge how useful Genie 4 is beyond the demo. The presence of SynthID is also a step in the right direction, because ever more realistic worlds make it more important to recognise what is generated; we discussed this in relation to the SynthID detector.
Written by the SpazioiTech newsroom with the help of AI tools, from facts verified across at least two independent outlets, and checked before publication. Spotted a mistake? Tell us.
Translated from the Italian original. Read in Italian
