Google DeepMind model guide ·

Genie 4 API and official access

Genie 4 is in research preview. Google has announced early access for selected developer partners. Its public Gemini API catalog does not list a Genie 4 model.

Looking for Google’s Genie 4?

Join Google’s Project Genie waitlist to express interest in trying the model. Access is subject to Google’s rollout.

Open Google’s waitlist

Want to explore an image today?

Turn one scene image into a shareable world preview on Img2World. Generation requires an account and credits.

Try image-to-world

Image-to-world generator

Upload a scene image to create a world you can explore and share with Img2World.

Is the Genie 4 API public?

As of November 10, 2026, Google’s public Gemini API model catalog, changelog, and pricing pages do not document a Genie 4 model ID, request endpoint, or price. DeepMind’s October 27 announcement describes a research preview for trusted testers and a planned rollout to early access partners through the Gemini API and Google AI Studio. That announcement gives teams a channel to watch. Your project still needs explicit access and a documented interface before it can call the model.

If a service advertises an endpoint under the Genie name, ask which model actually runs and how the provider verifies its access. Then ask what the response contains. An MP4 link can be useful for review, while an interactive session needs a way to keep sending actions after the first frame. A world asset has its own requirements for viewing, ownership, and reuse. A screenshot of an AI Studio picker or a code snippet with an assumed model ID answers none of those questions.

Before putting a preview model in a product, ask for the supported inputs, session and concurrency limits, output formats, and the billing unit. Check what happens if a viewer disconnects, if a request times out after acceptance, or if two people open the same link. These details determine whether you can promise a persistent world, safely retry a request, or estimate what a minute of exploration will cost. Until Google publishes a contract or shares one with your project, keep those decisions open.

How to try Genie 4

Start with the official waitlist above if you want Genie 4 specifically. For the broader Project Genie prototype, Google Labs documents access for eligible Google AI Ultra subscribers aged 18 or over. Sign in with the personal Google account attached to the membership. The current help page still refers to Genie 3, so check which version your account actually offers before subscribing or planning a demo around Genie 4.

There are three different things to verify when you see a demo online. First, can your account open Project Genie? Second, does the experience identify the model as Genie 4? Third, does your developer project have programmatic access? A clip can help you judge the model’s behavior, and a consumer session can help you try the controls. For an integration, ask Google for project-level eligibility and developer documentation.

If you just want to see what the model does, start with DeepMind’s own recordings. Look for moments where someone changes direction or gives a new instruction. Those moments show more about interactivity than a smooth forward camera move does. The public examples are a good way to decide what you want to test before you request access.

What does Genie 4 simulate?

Genie 4 generates an environment as you interact with it. DeepMind’s examples show keyboard navigation, camera control, and text instructions that alter a running scene. It also generates sound with the scene, including ambience, music, and dialogue. DeepMind notes that objects, spatial details, and audio can drift, especially during longer sessions or after multiple changes.

To judge whether that output fits your work, pick one action and one thing that should stay stable. Walk around the corner of a building, then turn back and see whether the same doorway is there. Ask for rain in a courtyard and check whether the layout survives the change. If you need a reusable review link, check whether a second viewer sees the same space when they open it later. Each test answers a different question about a world model.

For a story or game pitch, a short scene that reacts to input can help demonstrate the feel of an idea. If your team is evaluating the placement of an entrance or the sightline to a landmark, the ability to revisit a specific generated layout can matter more. If you need measurements, collision meshes, or an editable level, decide on those deliverables first and verify them independently.

Turn a scene image into an explorable world

If your goal is to walk through a concept from a photo or illustration today, use the Img2World generator above. Start with one image you have permission to upload. Choose a scene with a visible route and enough depth to tell near, middle, and far areas apart. Sign in, check the credits shown by the generator, and start the generation. When it finishes, open the world to explore and share it for review.

For a game environment, use a corridor that opens into a larger room. Keep the entrance and a destination visible in the image, then move through the preview and ask whether the destination still draws your attention from inside. For a venue concept, start with a wide view from the entrance. Check what becomes visible as you turn, and send a reviewer the link with a specific viewpoint to inspect. For an illustrated world, pick a street with a distinct landmark. Ask a collaborator whether it still provides a sense of direction when they look away and return.

The most informative first check is to compare the opening view to your source and then inspect one area outside the original frame. A one-image workflow has to infer surfaces and objects the photo never shows. If those areas look confusing, try a wider or better-lit reference before starting another generation. Use the result to discuss atmosphere and layout, and use independently measured material when a decision depends on exact geometry.

For your own integration, first decide what has to survive after the session. If reviewers need to open a generated world later, ask how long links last, whether access can be restricted, and which assets can be saved. If your users need to change a running scene with dialogue, weather edits, or continuous controls, check whether the interface supports those actions after generation starts. These are different requirements, and they can change how you build sharing, previews, and session recovery.

Genie 4 FAQ

Can I call Genie 4 with my existing Gemini API key?

Google’s public Gemini API documentation does not list a Genie 4 model ID. Early access, if offered to your team, depends on the developer project Google authorizes.

How much does the Genie 4 API cost?

The public Gemini API pricing page does not list Genie 4. Google AI Ultra pricing covers a consumer subscription, so it cannot be used to estimate per-session developer costs.

Does a Project Genie invitation include API access?

Google describes Project Genie access and Gemini API early access as separate paths. Confirm what your invitation covers before planning a developer integration.

Can I export an editable 3D scene from Genie 4?

Google’s preview announcement shows interactive simulations and recordings. It does not document a Genie 4 mesh or game-engine export. Check the documented output for your specific workflow before committing to a format.

Where can I try the official Genie 4 preview?

Use the Google waitlist linked at the top of this page. Img2World’s image-to-world generator is available separately for creating a scene preview from your own image.

Where should I check for new Genie 4 developer access?

Use the DeepMind announcement for rollout news and the Gemini API model list and changelog for callable models. For private early access, rely on the documentation and entitlement tied to your developer project.

Official sources