Genie 4: world model, access and what Google has shown

What Genie 4 actually generates

Announced on 27 October 2026, Genie 4 is Google DeepMind’s real-time world model. A scene continues to develop as you navigate with a keyboard and mouse or change it with text. Gemini Omni grounds its appearance and behaviour; other Gemini models help maintain continuity. This is a research preview, and we have not tested it.

For a film team, the relevant promise is interactive previs: change a camera position, lighting or action while the scene is running. DeepMind also describes game prototyping, reconstructing an environment from an archival photograph and business simulations. Those are demonstrations and proposed uses, not evidence of production reliability.

Four changes shown by DeepMind

Live direction. A text instruction can introduce an object, change weather or reposition the camera. In the lake demonstration, a saxophonist remains on a floating platform while the viewpoint explores the water. The scene description sets photography, movement and sound together.

Physical behaviour and people. DeepMind highlights aerodynamics, collisions and more natural human movement. A maple seed is a compact physics example. These clips do not establish engineering accuracy: the company still warns that objects and details can drift.

Native sound. Audio is generated with the world, including ambience, music and dialogue. The saxophone scene connects a visible instrument to its sound. DeepMind says the audio in its demonstrations comes from the model; audio accuracy remains an open problem.

Multi-step events. Gemini 3.8 Flash observes the simulation, breaks a goal into steps and checks subsequent frames. A separate robotics example combines Gemini 4 Argon with a simulated Apptronik Apollo humanoid playing draughts. This is a research workflow, not a downloadable game or robot controller.

Genie 4 vs Genie 3

Access, subscription price and API

As checked on 28 October 2026, Google One’s US page advertises Google AI Ultra from $99.99 a month, with a separate $199.99 tier in its comparison. This is a subscription price, not a Genie 4 usage quote. DeepMind has not specified which Ultra tier, country rollout, generation allowance or launch date will apply. Do not subscribe solely on the assumption of immediate Genie 4 access.

The official waitlist is the stated route for interest. Wider access is coming first to Ultra subscribers through Project Genie and early access partners through the Gemini API and Google AI Studio. There is no published Genie 4 API price or general endpoint in the announcement.

Limits worth checking before planning a project

DeepMind describes a few minutes of continuous interaction, drift during longer navigation and after repeated complex edits, and difficulty with several independent agents. Recognisable locations are not reproduced with exact fidelity. Legible lettering usually needs to be in the world description.

Every generated video and audio output is to carry SynthID. The official examples are illustrative and shortened recordings of live interactions. The embedded film below is Google’s demonstration; it is not our test. Neither its editing nor the 720p/24 fps figures for Genie 3 establish Genie 4’s output specifications.

Sources

Official documentation and demonstrations; checked on 28 October 2026.

• Google DeepMind: Genie 4 - 2026-10-28
• Genie: capabilities, limitations - 2026-10-28
• Google: introducing Genie 4 - 2026-10-28
• Project Genie (2026-01-29) - 2026-10-28
• Project Genie: prompt guide - 2026-10-28
• Genie 3 (2025) - 2026-10-28
• Google One - Google AI Ultra - 2026-10-28