Genie 4: World Simulation, Preview Access and Costs

Google announced Genie 4 on October 27, 2026, and began a trusted-tester research preview. Start at the official model page and follow its waitlist for Genie 4 in Project Genie.

Google plans to expand access to Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio.

Genie 4 is the model; Project Genie is a product through which Google plans to expose it. Keep the version visible in the product separate from the version shown in a launch demonstration when comparing features or subscription value.

A session evolves as you navigate and give instructions. You can describe a landscape, introduce a character, change the weather or request an event, then continue exploring the result. Google’s demonstrations use keyboard and mouse controls for navigation and camera movement, with text prompts steering the scene.

Audio is generated alongside the world: environmental sounds and characters’ speech are part of the simulation. Google says Gemini Omni helps anchor appearance and behavior while other Gemini models contribute reasoning for continuity.

For a first session after access is enabled, try a small change you can inspect: “A quiet courtyard with a fountain. Walk toward the fountain. Now make it rain while keeping the courtyard layout.” Watch the location, sound and object continuity before layering several changes into the same scene.

The official launch and model pages checked on October 28 do not list a Genie 4 per-minute, per-session or API rate. Session quotas, overage charges and API billing will depend on the access offer.

Before budgeting a project, obtain the account’s included usage, billable unit, session limits and retry rules. For an API integration, also confirm the model identifier and endpoint supplied to the early-access account.

For a prototype, track accepted interactive minutes, restarts and elapsed operator time. Those measurements help compare an eventual hosted quote with a self-hosted alternative. Track interaction latency separately from output frame rate.

DeepMind describes a few minutes of continuous interaction, with visual details and object permanence able to drift as a session grows or receives complex edits. It also identifies physical interactions, readable text, accurate real-world locations and audio accuracy as active limitations. Use short concept exploration as the starting workload.

Ethan Mollick shared a Genie 4 example on October 27 in which an office scene changes around a DOOM theme and a demon coworker. His first-person Bluesky post describes the changes as happening in real time.

When trying the system, revisit a landmark after turning away, ask a character a follow-up question and repeat a small scene edit. Check whether the result remains useful for the task. A convincing short clip is useful for ideation; a production workflow also needs repeatability, control and suitable export terms.

Choose the Genie 4 preview to explore Google’s integrated world simulation when access is available. Google supplies the hosted system, including its audio and steering behavior. The public launch does not supply downloadable Genie 4 weights or a self-host GPU configuration.

Tencent Hy-Nomas offers a different route: downloadable video-world weights, multi-turn text and camera control, and an official low-memory RTX 4090 configuration. You provision its runtime and optional interactive services. Its published 24 FPS setting is playback; generation speed depends on the complete runtime. Genie 4 includes native audio; Hy-Nomas provides a video-generation stack.

For a fixed video deliverable, compare available video-generation services by accepted clips, duration, resolution and reruns. For an interactive prototype, compare controllability and response time as well as machine or session cost.