Google DeepMind unveiled its new world model Genie 4 on October 27. Unlike a 3D program, a world model does not pre-render a finished scene: Genie 4 generates image and sound continuously while you control movement and camera with the keyboard. A text prompt or a photo that becomes the first frame of the scene serves as the starting point.
You can't try Genie 4 yet. In its announcement, Google DeepMind writes that initially only selected testing partners will get the model. After that, subscribers of Google AI Ultra will follow in Project Genie, a prototype from Google Labs, as well as developers with early access to the Gemini API and Google AI Studio. Google does not name a date; a waitlist is already running for Project Genie.
What is new compared to Genie 3?
Project Genie launched in January 2026 with the predecessor Genie 3. At that time, Google limited each session to 60 seconds, and events that change a world during exploration were missing from the prototype, as Google wrote in the post on the launch of Project Genie.
Genie 4 delivers these events. Via text input, you can make rain set in midway through exploration, pan the camera, or place a person in the scene. A Gemini model adapts the input to the environment and keeps the flow coherent between your inputs. Sound is also new: music, ambient noise, and dialogue are created simultaneously with the image. Characters react to your inputs and speak in real time. In addition, Google DeepMind promises more photorealistic images, more natural human movements, and more believable physics in collisions.
Who gets Genie 4?
Project Genie is only included in the most expensive subscription tier: on the German subscription page of Google One, Google lists the prototype for both variants of Google AI Ultra, but not for Google AI Pro. For Project Genie, Google names more than 140 countries and a minimum age of 18. Google has not yet said whether Genie 4 will be available in Project Genie in Germany right at launch.
Where are the limits?
Google DeepMind itself describes Genie 4 as an early, experimental model. The worlds remain coherent only over short sessions. The longer a simulation runs and the more often you rebuild it via input, the more details change, and objects do not reliably stay in place. The model does not yet reliably handle legible text, complex physical processes, and exact spatial relationships.
Every video and every audio track from Genie 4 contains SynthID, an invisible watermark from Google by which AI content can be recognized.
What this means for you
In the video and film sector, Genie 4 is above all suited to planning scenes before the shoot. Google DeepMind describes tests in which filmmakers play through a scene several times and change camera angles, lighting, dialogue, and plot live. If you already have an Ultra subscription, it's best to sign up for the Project Genie waitlist right away. Until Genie 4 arrives, you can use Genie 3 to check whether a quickly generated scene is enough as a draft for a client conversation.
Taking out an Ultra subscription just for Genie 4 is not worthwhile for now: a date and release for Germany are still missing. For storyboards with signs, titles, or logos, the model is not suitable yet anyway because of its weakness with text.
If you show material from Genie 4 in a client project, tell your client that the images come from an AI. How to plan and generate entire film scenes with AI tools is shown in the tutorial on AI videos on TutKit.
Google DeepMind unveiled Genie 4 on October 27, 2026.
Sources
