Google DeepMind's Genie 4 holds AI-made worlds for 2 minutes

Genie 4 sits in a category Google calls world models: AIs that build a navigable world on the spot from words or images, rather than playing back pre-made video. The jump from Genie 3 to Genie 4 is mostly about continuity. TechCrunch reports the new model can keep a scene from breaking apart for up to 2 minutes, double the previous generation, and it adds synchronized audio while improving how objects move and how human figures look. Users can also alter the generated world while it runs, which points to interactive rather than passive output.

The rollout is deliberately narrow. Ars Technica reports that access will first go to developers and to subscribers of Google's paid Google AI Ultra plan before widening further. That gap between announcement and general availability is the practical detail for anyone outside those groups: the capability is real, but the audience for now is limited.

The timing matters because the same day's news carried a counterweight. A Cohere report raised questions about a different open model, and Bank of America warned customers against handing banking passwords to AI assistants. Genie 4 shows the build side of AI advancing, while those items show the trust side still being worked out — a pairing that suggests capability and verification are moving on separate tracks.