Genie 4 launched two hours ago. Google DeepMind's new world model keeps a scene consistent for up to two minutes. Twice Genie 3. Everything else in the launch is prettier. That number is the product.

Genie 3 (Aug 2025) → Genie 4 (today):

→ Picture: 720p at 24 fps → photoreal, with aerodynamics and collisions → People: nobody talks back → characters you steer, who move, react and speak → Sound: silent → native ambient, music and live dialogue → Director: your keyboard → Gemini plans the steps and checks every frame → Memory: about one minute → up to two minutes (TechCrunch, today's briefing) → Access: limited preview → trusted testers now, then AI Ultra at $100+ a month and API partners.

DeepMind's own limitations paragraph: "visual details and object permanence can still shift or drift over time—especially as a simulation runs longer."

I run a one-person company on AI agents. I do not build worlds. But every launch gets the same first question: how long does it stay consistent? For a world model it is these two minutes.

Rule 1: Find the memory number before the price. It sets the task size: a robot episode, a previz take, a two-minute loop.

Rule 2: Read the limitations paragraph as the spec. DeepMind wrote it. The demo reel did not.

Rule 3: Budget the resets. Anything longer than the memory is a chain of short runs with a handoff. Same as agents.

The demo got photoreal. The clock got one minute longer.

Which number do you check first on a launch: the price or the limit?