Google DeepMind has presented Genie 4an artificial intelligence capable of creating virtual worlds that the user can navigate and modify while they are generated. The new version promises more convincing physics and characters that move and speak naturally, although it still has difficulty maintaining consistency of what is happening beyond two minutesaccording to Ars Technica.
A world model is an AI that simulates an environment and allows you to navigate through it with the keyboard and modify it using written instructionssuch as asking for rain, adding objects, making a character appear, or modifying the camera. In Genie, simply describe the scenario for the system to generate each frame in real time, without having to previously design all its elements or program each interaction as happens in a video game.
Genie 1which generated controllable two-dimensional environments, was presented in February 2024. Genie 2 made the leap to 3D in December of that year and Genie 3announced in August 2025, allows you to explore more realistic scenarios in real time, at 720p and 24 frames per second.
None of these models have been widely available. Genie 3 can be accessed at Project Geniean experimental website that combines that model with Nano Banana Pro and Gemini. Requires being at least 18 years old and a subscription to Google AI Ultrawith explorations limited to 60 seconds.
Although Genie has been linked to the creation of video games, Google considers this technology a step towards artificial general intelligencecapable of tackling very diverse tasks. These worlds could provide countless scenarios to train robots and AI agentslearn the consequences of their actions and test their capabilities before transferring them to the physical environment.
This is Genie 4, our most realistic and interactive world model yet. đ Explore dynamic environments in real time to see how the scene responds, sounds, and how the people in it move, react, and speak.
â Google DeepMind (@GoogleDeepMind) October 27, 2026
Compared to Genie 3, the new version improves, according to DeepMind, simulating collisions and aerodynamics, as well as human movement and reactions. But the most notable novelty is that now has native audiogenerated together with the images and integrated into the environment, such as the wind through the trees or the waves. Also incorporates human speech and interactive dialoguesso that characters can converse and react to events in the simulation.
The system incorporates capabilities of Gemini Omni to improve realism and uses Gemini 3.8 Flash as supervisor. He observes the scene, divides the objectives into steps, checks the frames and adapts the plan when the situation changes; Google also uses Gemini 4 Argon to plan tasks that Genie represents, like a robot playing checkers.
The loss of coherence or consistency It means that an object may change appearance or a place may not be reconstructed the same when returning to a location in the simulation. Ars Technica places that continuity in some two minutesalthough Google has not mentioned a specific limit but does mention deviations that can increase with duration, complex modifications and movement around the stage.
Genie 4 is reserved for now to selected testers and Coming soon to Ultra subscribers through Project Geniefor which there is a waiting list. Select partners will also be accessing it via the Gemini API and AI Studio, with no specific date or developer fees announced.
What it won’t be is cheap. In Spain, Google AI Ultra offers modalities 99.99 and 219.99 euros monthlybut it is not known how much use the Genie 4 can be given in each of them.