Can you try Genie 4 yet?
Google has published a rollout order, not a date. Pick the group that fits you best and the region you are in, and the checker will tell you where you stand in that order. It does not grant access. It only reads back what Google has announced.
Pick a group and a region to start.
Project Genie has so far been open to Google AI Ultra subscribers aged 18 and over in the US. Google has not said when Genie 4 reaches other countries, so an Ultra subscription outside the US is not a promise of access yet.
The short version
| Item | Genie 4 |
|---|---|
| What it is | A real time world model. You describe a place, or give it an image, and it builds a scene you move through with keyboard controls, frame by frame, instead of rendering a finished clip. |
| Announced | October 27, 2026, by Google DeepMind |
| Status | Research preview for trusted testers. Next, Google says, are Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio. No date is given. |
| What is new | Native audio, people who move and talk back, live text edits mid scene, and Gemini models steering multi step events. Scene coherence rises from about one minute to about two minutes. |
| What it is not | Not a public product, not an API you can call today, not open weights, and not a game engine. Google has published no price, no API rate, and no weights. |
| Watermark | All video and audio Genie 4 generates carries SynthID, Google's imperceptible watermark, across visual and audio output. |
What a world model does differently
A video model answers a prompt with a clip. A world model answers a prompt with a place. Genie 4 starts from your text or image and then predicts, frame by frame, what you would see next as you walk, turn, and act. That is why the demo footage looks like gameplay: the person at the keyboard is steering, not watching.
Four changes carry most of the Genie 4 announcement, and DeepMind lists them in that order.
Steer the scene while it runs
You can type a change mid simulation, for example new weather, a different camera angle, or a new person or object, and Gemini rewrites your request to fit what is already on screen. Between your inputs, Gemini keeps nudging the scene forward so it stays coherent instead of freezing between edits. This is the change Project Genie could not do at its launch, when a world was fixed once it started.
People and physics move first
DeepMind calls the people the most dramatic leap. Characters move naturally, react to their surroundings, and can be steered while they are moving. The physics work covers aerodynamics and collisions, so a thrown object and a walking character behave like parts of the same world instead of separate animations.
Sound comes from the world, not over it
Genie 4 is DeepMind's first real time world model to generate audio natively. Wind in trees, waves at a beach, music, and spoken dialogue are generated at the same time as the scene. Characters converse and react to your inputs or to events in the world, in real time, instead of reading lines recorded for a fixed script.
Gemini runs the plan behind the scene
Two Gemini pieces sit inside the loop. Gemini Omni anchors how the scene looks and behaves, so it stays grounded in real world appearance. Gemini 3.8 Flash acts as a guiding agent: it watches incoming frames, breaks a goal into steps, checks each step worked, and changes the plan when you intervene. DeepMind has also paired Genie 4 with Gemini 4 Argon for robotics work, for example breaking down a game of checkers into steps that Genie 4 renders with Apptronik's Apollo humanoid playing it out.
Genie 4 vs Genie 3
Genie 3 was previewed in August 2025 and has powered Project Genie for Google AI Ultra subscribers in the US since January 2026. The table below only uses figures DeepMind or its coverage states for both versions. Where a figure is not published, the cell says n/a rather than guessing.
| Feature | Genie 3 | Genie 4 |
|---|---|---|
| Preview | August 2025 | October 27, 2026 |
| Scene length | About 1 minute | About 2 minutes |
| Audio | Silent scenes | Native audio: ambient sound, music, and interactive dialogue |
| Live edits | No mid scene edits at Project Genie launch | Yes: weather, camera, people, and objects by typed prompt |
| People | Basic characters | Characters that move naturally, react, and can be steered live |
| Steering model | n/a | Gemini Omni plus Gemini 3.8 Flash as guiding agent |
| Public price | n/a | n/a |
Scene length for Genie 4 is the figure TechCrunch reports from the announcement (up to about two minutes, double Genie 3). DeepMind's own post says only that Genie 4 holds coherent environments over short sessions.
What testers are actually doing with it
DeepMind's tester cohorts span film, journalism, education, gaming, and engineering. The uses it names are concrete, and they are all early stage work, not finished products.
- Film previsualization: a director steps into a live simulation of a scene and runs several takes in real time, changing camera angles, lighting, dialogue, and action on the fly.
- Game prototypes: a developer goes from an idea to an explorable prototype in seconds, testing visual styles and mechanics by typing, before writing engine code.
- Archival photos as places: photojournalists turn a single 2D archival photograph into a grounded simulation of that environment, or visualize a short event like a flash flood unfolding.
- Robot training worlds: enterprise partners generate zero to one scenarios for testing how AI systems behave in simulation before they meet the real world, including the Apollo checkers experiment with Gemini 4 Argon.
What ties these together is speed of iteration, not final quality. A world you can change in seconds is useful for trying an idea, even when it cannot yet hold together for a full scene.
What Genie 4 cannot do yet
DeepMind is unusually plain about the limits, and this page keeps them plain too. None of these is a reason to dismiss the model, but each one is a reason not to plan a product on it yet.
- Drift: visual details and object permanence can shift or drift over time, especially as a simulation runs longer, after multiple complex prompt edits, or while navigating through the scene.
- Audio accuracy: real time audio accuracy is still unresolved, so a character's words and a scene's sounds are not yet reliable.
- Legible text: rendering text inside a world, like a sign or a label, is not yet reliable.
- Spatial accuracy: spatial accuracy and long term consistency are active research areas, not solved problems.
- Real places: Genie 4 cannot simulate real world locations with perfect accuracy, so a generated street is not evidence about a real street.
- Session length: beyond the reported about two minutes of coherent scene, Google gives no session cap, resolution, or frame rate figure.
What Google has not published
For a launch this visible, the unpublished list is long. Treat any page that fills these cells with numbers as guessing, unless it cites a newer Google source than the October 27 announcement.
| Item | Status on October 28, 2026 |
|---|---|
| Public release date | n/a: "coming soon" to Ultra subscribers and API partners, no date |
| Price or API rate | n/a: no Genie 4 price published |
| Countries beyond the US | n/a: Project Genie is US only for Ultra subscribers so far, and no Genie 4 rollout map is published |
| Resolution and frame rate | n/a: not in the announcement |
| Benchmarks | n/a: no benchmark table in the announcement |
| Open weights | n/a: no weights announced |
| Maximum session length | n/a: only the reported about two minutes of coherent scene |
Genie 4 is not Veo, and not a game engine
Three nearby products get confused with Genie 4, and the differences matter for what you would actually use each one for.
Veo is Google's video generation model. It renders a high definition clip with synchronized audio from a prompt. You watch the result. Genie 4 renders a world you enter and steer. Veo is for a finished clip; Genie 4 is for an interactive scene that has no finished form until you play it.
A game engine runs hand built 3D assets and code you wrote. Genie 4 builds its world from a prompt, frame by frame, with no assets and no code. That makes it fast to start and hard to control precisely, which is the opposite trade from an engine.
Runway's GWM Worlds 2 is the closest rival TechCrunch names: a playable world model that responds to inputs and includes audio, released a couple of months before Genie 4. The field around it is consolidating fast, with World Labs being acquired by AMD and Amazon reported in talks to buy Decart, but Genie 4's own access, price, and limits are still the ones in the table above, not a rival's.
How to get in line
There is one official route for consumers today, and it is a waitlist, not access. DeepMind links to the Project Genie waitlist from its announcement. Joining it does not promise a date, and Google has not said whether Ultra subscribers outside the US will be included in the first wave.
- Join the Project Genie waitlist (Google Form linked from the DeepMind announcement)
- Read the DeepMind announcement (the primary source for this page)
Do not pay anyone who claims to sell Genie 4 access, an API key, or a download. Google has announced no such product, and any seller is trading on a research preview that only trusted testers hold.
Frequently asked questions
What is Genie 4?
Genie 4 is Google DeepMind's world model, announced October 27, 2026. It generates an interactive scene in real time from a text prompt or an image, with native audio, and lets you change the scene by typing while it runs. It is a world you move through, not a video clip you watch.
Can I try Genie 4 today?
Almost certainly not yet, unless you are a trusted tester. Genie 4 is a research preview for a small group of testers. Google says Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio are next, with no date given. A waitlist is open in Project Genie.
How much does Genie 4 cost?
Google has not published a Genie 4 price, API rate, or subscription tier for it. Project Genie itself has so far been tied to the Google AI Ultra plan, but Google has not said Genie 4 will cost the same, or anything else, so this page does not guess.
Is Genie 4 a video generator?
No. A video generator renders a clip you watch from start to finish. Genie 4 renders a scene frame by frame as you move through it with keyboard controls, and it keeps responding to what you do and type. That is why DeepMind calls it a world model rather than a video model.
How long does a Genie 4 world last?
TechCrunch reports Genie 4 can hold a consistent scene for up to about two minutes, double the one minute limit of Genie 3. DeepMind itself only says it holds coherent environments over short sessions, and warns that details and object permanence can drift as a session runs longer or after complex edits.
What is SynthID on Genie 4 output?
SynthID is Google's imperceptible watermark for AI generated content. DeepMind says all video and audio generated by Genie 4 will carry it, across visual and audio output. It is meant to help identify content as AI generated, not to prove who made it or what is in it.
Sources
- Google DeepMind, Genie 4: A new generation of world simulation (October 27, 2026): https://deepmind.google/blog/genie-4-a-new-generation-of-world-simulation/ (announcement, four capabilities, tester uses, limits, SynthID, rollout order, waitlist link).
- TechCrunch, Google's Genie 4 world model adds audio, more realistic worlds, and real-time steering (October 27, 2026): https://techcrunch.com/2026/10/27/googles-genie-4-world-model-adds-audio-more-realistic-worlds-and-real-time-steering/ (two minute scene figure, Runway GWM Worlds 2 context, rollout in coming weeks).
- madrobot.blog, Genie 4: DeepMind's world model adds sound and talking people (October 27, 2026): https://madrobot.blog/2026/10/27/google-deepmind-genie-4-world-model-sound-dialogue/ (Genie 3 history, Project Genie US and age limits, no figures for session length, resolution, or frame rate).
Where this page says "reported", the figure comes from coverage, not from DeepMind's post. Where a cell says n/a, Google has not published that item as of October 28, 2026. This page is an independent explainer, not affiliated with Google.
Agents
Machine readable summary for AI assistants: