Google DeepMind announced Genie 4 on October 27, 2026, adding native audio and natural-language changes to its experimental model for generating interactive environments. Research-preview access has begun for trusted testers, with wider availability still to come. Google positions the system as a tool for exploring creative ideas and simulations rather than a finished production platform. (Google DeepMind)
The Verge reports experiences lasting about two minutes, compared with about one minute for Genie 3. That comparison describes a reported capability, not a guaranteed session length or documented hard limit. The supplied official announcement does not specify the two-minute duration. (The Verge)
Who can access Genie 4—and when
Google’s product page says the next access phase will begin with Google AI Ultra subscribers through Project Genie and early-access partners using the Gemini API and Google AI Studio. It describes that expansion as coming soon, without a calendar date. These are planned access routes, not confirmation that every subscriber or developer can use Genie 4 now. (Google DeepMind)
Google has also opened a Project Genie waitlist. For builders, the distinction between a research preview and unrestricted API availability matters: the announcement identifies selected partners as the initial developer audience. It does not establish a general release on which teams can base an immediate product launch. (Google DeepMind)
Sound and scene changes during an experience
Google says Genie 4 generates music, ambient sound and interactive dialogue alongside its worlds. Characters can respond to user inputs or events. Natural-language instructions can also change weather, move the camera or introduce people and objects while an experience is running. (Google DeepMind)
The potential creative benefit is the ability to revise an unfolding scene rather than assess only a fixed result. A filmmaker could explore how a change in atmosphere affects a sequence, while a game designer could investigate a different encounter. Those are possible applications of the announced controls, not evidence that the model delivers repeatable results in either workflow. (Google DeepMind)
Google describes Gemini 3.8 Flash as a guiding agent: it observes frames, breaks goals into steps and adjusts its plan as the scene or user instructions change. This is the company’s account of how it seeks to keep complex events on track, not an independently measured guarantee of continuity. (Google DeepMind)
Realism claims remain distinct from reliability
Google claims improvements in human movement and physical scenarios such as collisions and aerodynamics, as reflected in Ars Technica’s announcement coverage. Those claims concern the model’s portrayal of events. They should not be read as proof that its simulations accurately predict real-world physical outcomes. (Ars Technica)
Independent hands-on evidence is limited in the supplied reporting. The Verge’s Jay Peters explicitly says he had not tried Genie 4. His article therefore does not provide a hands-on comparison with its predecessor, even though it describes the announced changes and presents company examples. (The Verge)
Google acknowledges that visual details and object permanence can drift—meaning objects may not remain consistent as an experience progresses. Longer runs, complex edits and navigation can make that problem more noticeable. Audio accuracy and clear text rendering also remain limitations. (Google DeepMind)
For creative evaluation, those caveats suggest a more useful question than whether a single scene looks convincing: can the system preserve the elements a project needs while a user changes something else? Longer experiences would offer more room to explore an idea, but duration alone would not resolve continuity problems. This is an implication of the stated limitations, not a test result. (Google DeepMind)
A broader push toward usable generated worlds
A related industry development illustrates why access and integration deserve attention alongside demonstrations. World Labs announced its public World API on January 21, 2026, bringing its Marble model’s world-generation capabilities into a programmable interface. The company says it accepts inputs including text, images and video and returns navigable 3D environments. (World Labs)
World Labs says those environments can be displayed on the web, exported into other tools or incorporated into interactive systems. This is not a like-for-like performance comparison with Genie 4: generating an explorable 3D environment and maintaining an evolving audiovisual simulation are different product propositions. The comparison instead highlights two practical considerations for developers—what the output is and how it can enter an existing workflow. (World Labs)
Google says trusted testers are exploring film pre-visualization, creative ideation and interactive game prototypes. These remain exploratory uses, not established production outcomes. For prospective users, the next meaningful evidence will be whether broader access allows those ideas to survive repeated edits and longer interactions—not simply whether selected demonstrations look more realistic. (Google DeepMind)
