Genie 4 access: who can use it, what it does, and why there's no API yet

Genie 4 runs inside Project Genie for eligible Google AI Ultra subscribers, with a Pro waitlist. What changed from Genie 3, session limits, no public API, and how to prompt a world.

A snow globe holding a small landscape with a path, pines and a cottage, where snow turns to rain on one side and an orange weathervane swings, illustrating a world that changes while you explore it.

Genie 4 is Google DeepMind’s interactive world model. You can use it today only inside Project Genie, a Google Labs web app at labs.google/fx/projectgenie, and only if you are an eligible Google AI Ultra subscriber aged 18 or over. Google AI Pro subscribers can join a waitlist. As of November 19, 2026, Google documents no public developer API for Genie 4: it is not in the Gemini API model list, changelog or pricing page, and it is not available through Ofox.

If you were hoping to call it from code, the honest answer is “not yet, unless you are an early-access partner”. If you have Ultra, you can open it in a browser now, but Google describes sessions of around a minute of continuous interaction. Genie 3 is still in the model picker and supports longer runs.

How we wrote this: Ofox publishes this guide. It is documentation-based. We did not run Genie 4 or Project Genie, and nothing below is a measured result. Every fact comes from Google’s own pages, checked on November 19, 2026: the Genie 4 availability post (November 18), DeepMind’s Genie 4 research post (October 27), the Genie model page, Google’s Genie prompt guide, the Project Genie help article and the Google AI plans page. Where those pages disagree, we say so.

What a world model is, and why Genie is not a video generator

Google’s January Project Genie post defines a world model as a system that simulates how an environment changes and how actions affect it. The practical difference from a video model is control during generation.

A video model such as Gemini Omni 1.2 Flash or Veo 3.1 takes a prompt and returns a finished clip. You can’t steer the clip while it renders. Genie generates the next frames as you move. You walk with the keyboard, turn the camera, and the model produces what lies ahead in real time. With Genie 4 you can also type a new instruction mid-session (“fog rolls in”, “a cabinet door opens”) and the world responds.

That changes what it’s good for. A video model gives you a deliverable file. A world model gives you an explorable session, and you can only save a recording of it. Project Genie lets you download a video of your playthrough, but Google describes no way to export the world itself as a 3D scene or game asset.

Timeline: how Genie reached consumers

DateWhat happenedSource
August 2025Genie 3 previewed as a research model: real-time worlds at 720p and 24 fps, consistent for a few minutesDeepMind, Genie 3
January 29, 2026Project Genie launched for Google AI Ultra subscribers in the U.S. (18+), powered by Genie 3, Nano Banana Pro and Gemini; generations limited to 60 seconds; promptable events not yet includedGoogle blog, Project Genie
October 27, 2026Genie 4 announced as a research preview for trusted testersDeepMind, Genie 4
October 28, 2026Sundar Pichai mentions Genie 4 in Alphabet’s Q3 earnings remarksQ3 2026 CEO remarks
November 18, 2026Genie 4 rolls out in Project Genie to eligible Ultra subscribers, alongside Genie 3 updates, the Project Genie agent and a community gallery; Pro members can join a waitlistGoogle blog, Genie 4 available

For the wider release picture across Gemini 4, Omni and Decision 1, see our tracker of confirmed and rumoured Gemini 4 models.

What Genie 4 adds over Genie 3, claim by claim

Google’s announcement mixes model capabilities with Project Genie app features. They are listed separately here, because the app features are not things “Genie 4 the model” does on its own.

Model capabilities (Genie 4)

ClaimWhat Google actually saysStatus
Mid-session promptingType a new prompt while navigating to change weather, add a character or redirect the action. The Oct 27 research post adds that Gemini adapts your prompt to the surrounding scene and steers the world between your inputs.Confirmed in both the Nov 18 post and the Oct 27 research post. Not exclusive to Genie 4 any more: the same Nov 18 update added real-time promptable events to Genie 3 in Project Genie.
Native, adaptive audioDeepMind calls Genie 4 its first real-time world model to generate audio natively: ambient sound for each setting, and interactive dialogue in which characters react to your inputs.Confirmed. Google lists real-time audio accuracy as an open limitation.
Realism of physics and peopleBetter photorealism, simulated aerodynamics and collisions, and characters that move naturally and can be steered. DeepMind calls people the biggest improvement.Google’s claim; no public benchmark is given.
Gemini in the loopGemini Omni anchors appearance and behaviour. In DeepMind’s complex-event demos, Gemini 3.8 Flash observes the scene, splits a goal into steps and checks each step.Confirmed in the Oct 27 research post.

Project Genie features (not the model)

FeatureWhat it doesStatus
Project Genie agentYou describe a world or story in conversation; the agent orchestrates Google’s multimodal models to create starting visuals and a background score. The help article calls the chat step the “Experience Builder”.New on Nov 18.
Gemini during explorationGemini can watch the screen, offer suggestion chips for where to go next, or trigger in-world events.New on Nov 18.
Community gallery and remixBrowse curated and community worlds, remix their environment, characters and audio, and publish your own.Remixing existed since January; publishing to a community gallery is new.

So if your question is “what can I only do with Genie 4?”, the honest list is short: native audio including dialogue, and Google’s highest-fidelity visuals and physics. Live promptable events now work with either model in Project Genie, and Google presents the agent and gallery as Project Genie features rather than Genie 4 features; it does not say whether the agent is limited to one model.

How to get access

Who qualifies

According to the Project Genie help article and Google’s posts:

  • Google AI Ultra subscribers can use it where Google AI Ultra is available. The plans page says Project Genie is offered in more than 140 countries and links a country list.
  • Google AI Pro subscribers can join a waitlist. Google grants access gradually and emails you when it is your turn.
  • Age 18+ is required.
  • Sign in with a personal Google Account, as the help article instructs. The January launch post said Project Genie was not available to Google AI Ultra for Business users. The November pages neither repeat nor lift that exclusion, so Genie 4 access for Ultra for Business (Workspace) accounts is unconfirmed.
  • Google AI Plus is not listed as eligible.

One inconsistency to check before you pay. We re-fetched the plans page on November 19, 2026, and it says three different things. The comparison table’s Project Genie row (“Access to interactive world model”) shows a tick for both the Ultra 5x and Ultra 20x tiers and a cross for Plus and Pro. The Ultra benefits list says “Only available with a Google AI Ultra plan.” Footnote 8, which both of those point to, says “Only available with a Google AI Ultra 20x plan.” Google’s November 18 post and the help article say “eligible Google AI Ultra subscribers” without naming a tier. If you are buying Ultra mainly for Genie, open the plans page for your country and confirm the tier before subscribing.

Usage limits and cost

The help article says you do not spend AI credits to generate worlds while Project Genie is an early-access prototype. Daily usage limits apply and “may vary by plan and demand”. During heavy load you may be placed in a queue. Google publishes no fixed number of sessions per day.

Step by step

  1. Go to labs.google/fx/projectgenie and sign in with the Google Account that holds your Ultra (or waitlist-approved Pro) membership.
  2. If you have no access, you land on the Google Labs page. You can still browse short previews in the gallery. Pro subscribers click Join waitlist.
  3. To build, click Create with Genie, describe your experience in the prompt box or upload an image, select a model (Genie 4 or Genie 3), and click Create.
  4. Refine the experience in the Experience Builder chat, then test it before you share.
  5. Explore with WASD to move, Spacebar to jump or ascend, and the arrow keys to turn the camera. Many experiences also accept typed prompts that trigger events. Some use custom controls.
  6. When the session ends, you can replay it (results will differ), rewatch it in Visit History, download a video, share a link, or remix it.

Session length, resolution and model choice

Genie 4Genie 3 (in Project Genie, after Nov 18)
Continuous interaction”Around a minute” per Google’s model-page limitationsUp to 3 minutes per generation (up from 60 seconds), and worlds can be chained into longer journeys
Seeding from a 360° panoramaNot mentionedYes, new on Nov 18
Live promptable eventsYesYes, new on Nov 18
Native audio and dialogueYesNot claimed
Published resolution / frame rateNot published720p at 24 fps was the research spec in August 2025; Google does not say this is the Project Genie stream

The help article adds that how long you can explore “depends on the model and experience” and that stream quality may drop under server load. Treat every number above as Google’s stated ceiling, not a guarantee for your session.

Pick Genie 4 when you care about realism, sound or a talking character in a short scene, for example a one-minute walkthrough for a film pre-visualization pitch.

Pick Genie 3 when length matters more than fidelity: a longer flight across chained worlds, or a space seeded from a 360° panorama of a real place you photographed.

Is there a Genie 4 API?

No public developer API is documented as of November 19, 2026. We checked three places:

  • Google’s Gemini API model list, changelog and pricing page: none mentions Genie. There is no model ID, endpoint or price.
  • DeepMind’s October 27 post says Genie 4 is “coming soon to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and Google AI Studio.” The Ultra half arrived on November 18. The developer half is limited to early access partners, which is not general availability, and the post links only a Project Genie waitlist, with no way to apply for API access. The November 18 post does not mention the API at all.
  • Ofox: Genie 4 did not appear in GET https://api.ofox.ai/v1/models when we queried it on November 19, 2026. You cannot call Genie through Ofox, and we will not suggest a workaround.

We did not check Vertex AI’s model garden for this article.

If you need generated motion inside an application today, you need a different kind of model. Video models return a finished clip from a request, with no live steering:

Neither of these is a substitute for an explorable world. They are what you can actually ship in code this week.

How Genie 4 relates to Gemini 4

Genie 4 is not a Gemini 4 text model and does not belong in the Carbon / Neon / Hydrogen / Argon line-up covered in our Gemini 4 model selection guide. You won’t find it in the Gemini app’s model picker or behind generateContent.

The two families are connected in three documented ways:

  1. Inside Genie 4, Gemini Omni anchors simulations in real-world appearance and behaviour, and other Gemini models handle the reasoning that keeps a scene coherent.
  2. In the loop, DeepMind’s demos use Gemini 3.8 Flash as a guiding agent that plans steps and checks frames, and Project Genie lets Gemini watch your session and suggest or trigger events.
  3. In robotics research, DeepMind shows Gemini 4 Argon breaking a task into steps that are fed to Genie 4 as text captions while it renders. The example simulates Apptronik’s Apollo humanoid playing checkers. If Argon is the part you need, our Gemini 4 Argon access guide covers the subscription and API routes.

What Google says it is for

Google names consumer and research uses in roughly equal measure.

Agent and robot training. Pichai’s Q3 remarks called Genie 4 “an important building block for training robots and agents.” The model page points to SIMA, DeepMind’s game-playing agent, learning skills inside generated worlds. It also calls robotics simulation an early application. Keep this in proportion: Google describes research demos and enterprise evaluations, not a product you can buy for robot training.

Creative and professional testing. DeepMind reports that trusted testers used Genie 4 for:

  • film pre-visualization, re-running a scene with different camera angles, lighting and dialogue;
  • turning an archival photograph into an explorable setting, or showing how a flash flood unfolds;
  • game concept prototyping before writing engine code;
  • “zero-to-one” scenario generation for enterprise simulation.

Play. The gallery examples include an ant’s-eye colony (“The Life of Ants”), a room-by-room mystery (“Villa Giallo”), a guess-the-location game and a rock-paper-scissors match refereed by Gemini.

Limitations and safety

Google lists these model limitations for Genie 4:

  • Visual details and object permanence drift over time, especially in long sessions, after several complex prompt edits, or while moving through the scene.
  • Interactions between users, multiple independent agents and shared environments are hard to model accurately.
  • Real-world locations are not simulated accurately.
  • Legible text usually appears only when it is in your input description.
  • Continuous interaction is limited to around a minute.
  • Real-time audio accuracy is still being improved. The Nov 18 post adds camera latency to the open problems.

The help article lists app-level known issues with workarounds:

SymptomGoogle’s suggested fix
Character runs backwardsPress Spacebar to reset the character
You lose control of the characterKeep the subject centred and prominent in uploaded photos; turn real-world scenes into a more game-like version and retry
Controls respond slowlyKnown latency; no user fix
Stream looks low quality or darkens over timeServer load and a known issue; retry later
Generation fails or stops earlyPolicy violation in input or output, daily limit reached, or too many users; retry the next day or later

On safety, Google says all Project Genie video and audio carries SynthID watermarking, and that inputs and outputs pass safety screening. Genie 4 was released first to trusted testers after internal safety evaluations. The help article prohibits harmful content and asks you not to upload images of other people without permission. Repeated violations can cost you access.

Two privacy details are easy to miss:

  • When you share or publish with remix allowed, others can see the prompts, images and other assets you used, and reuse them.
  • Deleting your Project Genie data also deletes worlds you have published.

Writing world prompts: Google’s guidance, condensed

Everything in this section comes from Google’s Genie prompt guide and the example prompts on DeepMind’s pages. We have not run any of these prompts. The templates are our restructuring of Google’s advice, not tested recipes.

Google frames a Genie 4 prompt as three things: the scene and camera view, the live events and sound, and the starting image.

1. The opening scene prompt

Google’s guidance, in order:

  • Environment. Name the place and be specific about surfaces: a dirt path, an asphalt road, a calm ocean.
  • Style and camera. State the style (photorealistic, painterly, claymation) and whether you explore in first or third person.
  • Foreground and midground. Anchor what’s close (handlebars, a table edge) and what’s just ahead (an archway, a character).
  • Movement and sound. Say what moves and what it sounds like, such as wind in trees or gravel underfoot. Add “No music” if you want only ambient sound.

DeepMind’s published examples follow this pattern. Some are two short blocks labelled “Environment” and “Character”. Others are long cinematic briefs that set film stock, lighting, camera behaviour and a list of things to exclude. Both forms appear in Google’s own showcase, so length is not the issue. Clarity about camera and scene is.

A template built from that guidance (untested):

Environment: [place], [ground surface], [key landmarks near and far], [lighting and time of day].
Style: [photorealistic / painterly / claymation / other], [one or two texture or film cues].
Camera: [first-person, anchored by ___ | third-person following ___, centred with room to move].
Foreground: [what is right in front of the viewer]. Just ahead: [the next thing to walk toward].
Motion and sound: [what moves], [surface sounds], [ambient sound]. [No music.]

2. Live prompts while you explore

Google’s rule here is the most practical one in the guide: prompt one clear change at a time, in short declarative sentences. Its examples are as plain as coastal fog rolling in, opening a cabinet door, or placing a lamp beside a chair. Google lists drift after “multiple complex prompt edits” as a limitation, so stacking five changes into one live prompt works against you.

Coastal fog rolls in from the left.
The cabinet door swings open.
A street musician starts playing an accordion near the fountain.

3. Characters and dialogue

  • Choose first person (anchored by hands, handlebars or a workbench) or third person (character centred with space around them).
  • Describe how movement affects the world, such as boots crunching in snow.
  • For speech, Google says dialogue works best with one character in the scene. Set the vocal tone up front or put a short line in quotes.

4. Starting from an image

  • Upload a clear, well-lit, high-quality image. For third-person scenes, keep the subject centred with room to move.
  • For text-only prompts, Project Genie shows an image preview first. Google advises refining the prompt until the preview looks right before you enter, because the preview sets the look of the world.
  • Google also suggests using the Gemini app to expand or rewrite a description.

A quick pre-flight check

Before you spend a daily generation, check that your prompt does each of these:

  • names the camera mode;
  • describes the ground under the viewer;
  • puts something specific in the foreground;
  • says what should be heard;
  • avoids relying on legible signs or text (unless the text is in the prompt itself).

That last point follows directly from Google’s text-rendering limitation.

Who should care right now

  • You have Ultra and make visual work. Try it for pre-visualization and mood exploration. Plan around one-minute Genie 4 sessions, and download recordings of anything worth keeping.
  • You have Pro. Join the waitlist. There is no published timeline.
  • You are a developer. There is nothing to integrate yet. Watch the Gemini API changelog for a Genie entry. Until a model ID and pricing appear there, any third-party “Genie 4 API” claim deserves suspicion. Use a video model if you need generated motion in production.
  • You work on agents or robotics. The Gemini-in-the-loop and Argon planning demos show where Google is heading. Today there is no programmable access outside Google’s partner arrangements.

We will update this page when Google publishes API access, pricing or new session limits.