Google has unveiled Genie 4, the latest version of its AI system that builds interactive, explorable environments from scratch rather than just generating a static image or video. The update doubles the length of experiences it can create and adds real-time editing and sound. For now, it's a research demo, not a product you can buy.
Genie belongs to a category researchers call world models: AI systems trained to simulate physical spaces and the rules that govern them, rather than just describe or illustrate them. Type in a prompt and the model generates a navigable scene, complete with a sense of gravity, object permanence, and cause and effect. The previous version, Genie 3, could sustain roughly a minute of coherent interaction before the simulation would drift or fall apart. Genie 4 pushes that to about two minutes.
The new version also lets users interrupt a generated experience mid-stream and change it with a plain-language instruction, add a storm, remove a wall, swap the lighting, without restarting the simulation. Google says Genie 4 can now generate matching audio as well, and that it runs on the company's Gemini 3.8 Flash model to handle some of the underlying reasoning. Google has positioned the technology as a tool for training other AI systems, particularly robots and game-playing agents, inside simulated environments before they're deployed in the real world.
None of this is available as a consumer or business product today. Google has released it as a research preview, the same path it took with Genie 3 last year, with no stated timeline for broader access, pricing, or an API.
World models have become one of the more active fronts in AI research over the past year, alongside the chatbot and coding-assistant races most business owners have already noticed. NVIDIA has pushed its own Cosmos models for robotics and autonomous vehicle training. Startups including World Labs and Decart have released tools that generate walkable 3D scenes from a single photo. The common thread is training data: realistic simulated environments are far cheaper to generate at scale than filming or hand-building them, which matters enormously for companies training robots, self-driving systems, or game AI.
For most small businesses, Genie 4 isn't a tool to add to a workflow this week. It's a research system aimed at Google's own AI training pipelines and, eventually, at game studios, robotics companies, and simulation-heavy industries like logistics or industrial training. A retail shop or accounting firm won't be using it to make marketing content any time soon.
The businesses that should pay attention now are those adjacent to the categories Google is targeting: game developers, VR and AR studios, and companies building or buying robotics or automation systems that rely on simulated training environments. If world models mature the way text and image generation did, cheaper simulation could eventually lower the cost of building training scenarios for warehouse robots, driving simulators, or employee VR training modules. That's a multi-year horizon, not a next-quarter one.
There's also a cost question worth flagging early: research previews like this one often become metered API products later, following the same arc as Google's and OpenAI's video and image generation tools, free or cheap access during the demo phase, usage-based pricing once it ships widely.
Watch for three things: whether Google opens any developer access or waitlist for Genie 4, whether competitors like NVIDIA or OpenAI respond with their own world-model upgrades in the coming months, and whether Google starts referencing Genie inside its robotics or Gemini product lines rather than keeping it as a standalone research project.
For now, Genie 4 is a signal of where simulation technology is heading rather than a tool to adopt. Small business owners in gaming, training, or robotics-adjacent fields should note the trend; everyone else can safely wait for a product announcement before changing anything.