Google's Genie 4 Builds Interactive Worlds From a Written Prompt

Google has announced Genie 4, an AI model that generates interactive experiences from text prompts. It will first reach a small group of trusted testers as a research preview. The announcement was reported on 2026-10-27 GMT. The Verge described the model as producing more realistic worlds than its predecessor.

A world model is an AI system that learns how an environment behaves, so it can produce what happens next as you go. Think of a flight simulator that draws its own scenery while you fly, rather than loading a fixed map.

Users can change a running experience with a plain-language instruction. They do not need to start generation over. Genie 4 can also generate audio, and it uses Gemini 3.8 Flash to help with its simulations. Experiences can last about two minutes, up from roughly one minute for Genie 3.

Google suggests three uses: previewing shots for filmmakers, prototyping game ideas, and building training scenarios for robots. Each of these needs the world to stay consistent over time, and Google has said so directly. As a simulation runs longer, details can shift and objects may not persist. In other words, a cup on the table may change into something else or disappear.

Access is staged. Google has opened a waitlist for trusted testers to apply through Project Genie. It says the model will come soon to Google AI Ultra subscribers.

Genie 4 continues work that Google DeepMind began publicly with Genie 2, which it described in December 2024 as a large-scale foundation world model. DeepMind's Genie 3 announcement in August 2025 called it the lab's first world model to allow interaction in real time, with better consistency and realism than Genie 2. In January 2026, Google introduced Project Genie, an experimental research prototype that lets users create, explore and remix their own interactive worlds. Later that year, Google said Genie is a general-purpose world model that can generate varied, interactive environments. Project Genie accepts text and images as input.

In May 2026, Google said it was connecting Project Genie with nearly 20 years of Street View imagery, so generated worlds could be tied to real places. That is the clearest sign so far of where the work is heading. It is moving away from purely invented scenes and toward environments based on mapped geography.

Markets have reacted before. Reuters reported that videogame stocks slid after Google announced an AI model that turns prompts into playable worlds. Genie 4 is pitched for previews and prototypes, not as a consumer game engine. Still, its overlap with interactive entertainment is why that January reaction drew attention.

In my view, the most important figure for professional use is not the two-minute limit but the drift. A film shot or a robot training run needs an object to stay the same object from the start of a sequence to the end. Google's own caveat suggests that keeping a world consistent over long periods is still the hard engineering problem. Real-time editing raises the stakes, because each prompt given mid-run is another chance for the scene to drift.

The broader context here is that world models are judged on two things at once. One is how convincing the visuals and sound are. The other is whether the simulated world stays coherent when someone interacts with it. Genie 4 addresses the first more visibly than the second. Its editing and audio features extend what a single session can do. If preview testers find that the world holds together for the target lengths, the same approach could let studios and robotics teams test scenarios cheaply before spending on physical sets or production work.