Google Genie 4
Ask an AI to create a rainy street in a futuristic city. An image model gives you one picture. A video model like Veo gives you a clip that plays the same way every time. A world model like Genie 4 tries to let you walk down that street, turn the corner, tell it to stop the rain, and see what the city does next.
That is the whole idea. Since mid-November 2026 it is also something people outside Google can try, though access is still narrow and not cheap.
Status (checked November 20, 2026): rolling out to eligible Google AI Ultra subscribers in Project Genie since November 18, with a waitlist for Google AI Pro members. Generally available on the paid Gemini API as genie-4 since November 19, at about $0.10 per second of streamed video. Google still describes Genie 4 as a model in the frontier research space, with known limits on consistency and physics.
What Is Google Genie 4?
Genie 4 is a world model from Google DeepMind, announced on October 27, 2026 as a research preview for trusted testers. A world model generates an environment and predicts how it should respond to actions. You give it a starting prompt or image, and it produces a scene you can move through in real time, with the view, objects, characters, and sound generated as you go.
It matters because it is a different kind of output from the AI images and videos most people know. Instead of a finished artifact, you get something that reacts. Google frames simulation as a step toward more general AI: a model that can predict how the world responds is useful for training robots and AI agents, not only for entertainment.
The first testers came from creative industries, journalism, education, gaming, and engineering. The November release opened it to Ultra subscribers and API developers, but it is still better suited to experiments than to production work.
How Genie 4 Works
From what Google has published, using Genie 4 looks like this:
- Start a world from a text prompt or an image.
- Move through it with standard keyboard controls, including moving the camera.
- Change it while it runs by typing an instruction, such as changing the weather, moving the camera, or adding a person or object. Google says Gemini adapts each prompt to the surrounding scene and keeps the world going between your inputs.
- Watch it respond. Google says Genie 4 simulates physical effects such as collisions and aerodynamics, and characters that move naturally and can be steered.
- Hear it. Sound is generated alongside the visuals (more below).
Google’s API documentation now gives concrete limits. Genie 4 streams 16:9 video at 1280x720 and 24 frames per second with 48 kHz stereo audio, and a session lasts up to 120 seconds. Google adds that visual coherence and object permanence are strongest in the first 45 to 90 seconds. You can seed a world with text, one 16:9 reference image, or both, steer with WASD or arrow keys, and send text events mid-session. Google serves it from the United States, Western Europe, and Japan; farther away, latency delays steering and can lower the frame rate. Google still has not published the model’s architecture.
Interactive Worlds vs AI Video
The difference is easiest to see side by side.
| AI video (for example Veo) | Interactive world model (Genie 4) | |
|---|---|---|
| What you get | A finished clip | A scene generated as you explore it |
| Your role | Write the prompt, then watch | Steer it continuously with keyboard and text |
| Same prompt twice | Similar clips | Different results depending on what you do |
| Good for | Footage you can edit and publish | Exploring an idea, testing a space, simulation |
| Maturity | Commercially available | Early release: Ultra app and paid API |
A video model decides the whole shot up front. A world model has to keep deciding, frame by frame, in response to you, and it has to remember what it already showed you. That is much harder, which is why consistency is still the main weakness (see limitations below).
Genie 4 is not a game engine. There are no levels, saved state, game rules, or a way to export the world into Unity or Unreal. It generates the experience of a place, not the software behind one.
Native Audio
Genie 4 is the first of Google’s real-time world models to generate audio natively. According to Google:
- Ambient sound and music are generated with each world, on the fly.
- Interactive dialogue: characters can converse and react to your inputs or to events in the scene in real time.
What Google has not shown in detail is how accurate the audio is: whether sounds always match on-screen actions, or how dialogue holds up over a long session. Google lists real-time audio accuracy as an open research area.
World Simulation: Consistency, Physics, and Characters
Consistency. When you look away from something and back, it should still be there. Google says Genie 4 improves on this, and also says plainly that “visual details and object permanence can still shift or drift over time.”
Physics. Google describes simulated aerodynamics and collisions. This is generated behavior that looks plausible, not a physics engine computing real values, so do not use it for anything that depends on physically correct results.
Characters. Characters move, react, and speak, and can be steered in real time. Complex interactions between several objects or characters are one of the areas Google says still need work.
How Genie 4 Uses Gemini
Google says Genie 4 builds on the Gemini family in several ways:
- Gemini Omni helps anchor simulations in real-world appearance and behavior.
- Other Gemini models provide reasoning that keeps a simulation coherent over time.
- Gemini interprets your prompts, adapting a typed instruction to what is happening in the scene.
- As a planning agent: in Google’s robotics demonstrations, Gemini 3.8 Flash observed the scene, broke a goal into steps, and checked whether each step succeeded while Genie 4 simulated the world. Google also mentions experiments pairing Genie 4 with Gemini 4 Argon for task planning.
The robotics examples are research demonstrations, including a Universal Robots UR5 arm dispensing liquid in a lab and an Apptronik Apollo humanoid playing checkers. They show where Google is heading, not a feature you can turn on.
Project Genie: The App, the Agent, and the Gallery
Most people will meet Genie 4 through Project Genie, Google’s browser app for creating and exploring worlds, rather than through the API. The November 18 update added four things:
- Genie 4 in the model picker, with mid-session text prompts to change the weather, add a character, or redirect the action, and native audio that adapts as the scene unfolds.
- An updated Genie 3, still selectable when you want longer runs: up to 3 minutes per generation (up from 60 seconds), worlds chained together, worlds seeded from 360-degree panorama images, and prompted in-world events.
- The Project Genie agent, which builds a world through conversation. You describe what you want, and it uses Google’s other models to set the visual style and compose a background score. Gemini can also join a session to watch what is on screen, suggest where to go next, or trigger events.
- A community gallery of curated and user-made worlds that you can explore, remix (changing the structure, soundtrack, or rules), and publish your own creations to.
All video and audio from Project Genie carries a SynthID watermark, and inputs and outputs pass safety screening. Google’s update post does not restate the earlier restriction to US users aged 18 and over, so check the eligibility notes on the Project Genie page.
Building With the API
On the Gemini API, Genie 4 runs on the Interactions API (v1beta), over WebRTC for browsers or WebSockets for servers, with short-lived session tokens for client apps. Each session follows the same lifecycle: create it with a world setup, connect, steer it with keyboard actions and text events, and call complete() to end it. That last step matters for your bill, because charges accrue for every second of video until the session ends. A full two-minute session costs about $12 in video output alone.
For fixed-length clips, multi-turn video editing, or 4K output, Google points developers to Gemini Omni instead. Genie 4 is for live, steerable streams.
Realistic Use Cases
These are the directions Google and its early users point to. None of them is a turnkey commercial workflow yet.
- Game and level concepts: walk through a rough idea of a place before anyone builds it.
- Film previsualization: explore camera angles and the mood of a set.
- Interactive storytelling and education: step inside a scene instead of reading about it.
- Robotics and agent research: generate varied training and test environments, which is the use Google emphasizes most.
- Creative prototyping: quick, disposable experiments with a look or atmosphere.
For finished footage, a video model such as Veo is still the practical choice. For a real game, you still need a game engine.
Availability and Pricing
| Access route | Status (November 20, 2026) |
|---|---|
| Google AI Ultra, via Project Genie | Rolling out to eligible subscribers since November 18 |
| Google AI Pro | Waitlist on the Project Genie page; no date |
| Gemini API and Google AI Studio | Generally available on the paid tier since November 19 (genie-4) |
| API price | $1.50 per 1M input tokens; $9 per 1M text output; $17.50 per 1M video output (about $0.10 per second, audio included) |
| Free API tier, Batch, Flex | Not supported |
For context, Project Genie launched on January 29, 2026 with the earlier Genie 3 model, for Google AI Ultra subscribers in the US aged 18 and over (not Google AI Ultra for Business), with generations limited to 60 seconds. Genie 3 sessions now run up to 3 minutes.
All Genie 4 output is watermarked with SynthID in both video and audio.
Current Limitations
Google lists these as active research areas, and they are the reasons to keep expectations modest:
- Short sessions: up to 120 seconds per API session, with coherence strongest in the first 45 to 90 seconds.
- Long-term consistency and object permanence: details can drift over time, more noticeably the longer a simulation runs.
- Complex physical interactions between multiple objects or characters.
- Real-time audio accuracy.
- Legible text inside scenes, such as signs and labels.
- Spatial accuracy.
- Access and cost: Ultra only in the app, a waitlist for Pro, and per-second API billing that adds up quickly.
Google also lists camera latency as an open problem, and earlier noted that characters in Project Genie could be less controllable or respond with noticeable latency. Independent reviews of the November release were still thin when we checked.
Genie 4 vs Genie 3
| Genie 3 | Genie 4 | |
|---|---|---|
| Announced | August 2025 | October 27, 2026 |
| Public access | Project Genie, Google AI Ultra (US, 18+) since January 2026 | Project Genie for eligible Ultra subscribers since Nov 18, 2026; paid Gemini API since Nov 19 |
| Audio | No native audio | Native ambient sound, music, and dialogue |
| Changing a running world | Promptable events were announced but not in the Project Genie prototype | Text prompts to change weather, camera, people, and objects |
| Realism | Google’s first real-time interactive world model | Google cites improved photorealism, movement, and physics |
| Session length | Up to 3 minutes per generation since Nov 2026 (was 60 s) | Up to 120 s per API session |
| Gemini role | Gemini and Nano Banana Pro used for setting up worlds in Project Genie | Gemini Omni grounding, Gemini reasoning, and Gemini planning agents |
Google has not published side-by-side benchmarks between the two.
Related
- World models explained, our guide to where Genie 4 fits in the bigger picture
- Gemini, Google’s AI assistant and model family that Genie 4 builds on
- Veo, Google’s AI video model, for finished clips
- Multimodal AI, the broader idea of models that work across images, sound, and text
- Computer-use agents, another area where models learn to act inside an environment
Continue learning
Explore related guides, tools, workflows, and prompts that help you go deeper into this topic.
See how this tool fits into a workflow
Browse step-by-step AI workflows that use ChatGPT, Claude, Gemini, and other tools.
Frequently Asked Questions
Can I use Genie 4 today?
Yes, if you have an eligible Google AI Ultra subscription: Google began rolling Genie 4 out in Project Genie on November 18, 2026. Google AI Pro members can join a waitlist on the same page, with no date given. Developers can use it on the paid tier of the Gemini API as genie-4, generally available since November 19, 2026.
How much does Genie 4 cost?
In Project Genie it is included with an eligible Google AI Ultra subscription. On the Gemini API it is billed by tokens: $1.50 per million input tokens, $9 per million text output tokens, and $17.50 per million video output tokens. Video is 5,760 tokens per second at 24 frames per second, which works out to about $0.10 per second, or roughly $6 per minute, with audio included. Billing runs until you end the session, and there is no free tier, Batch, or Flex option.
What is the difference between Genie 4, Project Genie, and the Project Genie agent?
Genie 4 is the model that generates the world. Project Genie is Google's web app where Ultra subscribers create, explore, share, and remix worlds, and it lets you choose between Genie 4 and an updated Genie 3. The Project Genie agent is a creative assistant inside that app: you describe a world in conversation and it uses Google's models to set up its look and a background score. The API gives developers the Genie 4 model directly, without the app, agent, or gallery.
Is Genie 4 a video generator?
No. A video generator like Veo produces a finished clip you watch. Genie 4 generates a world frame by frame while you move through it with the keyboard and change it with text prompts, so what you see depends on what you do. It is closer to a playable simulation than to a video.
Does Genie 4 generate sound?
Yes. Google says Genie 4 is its first real-time world model to generate audio natively: ambient sound, music, and interactive dialogue, with characters that can talk and react to your inputs. Google also lists real-time audio accuracy as an area still being researched.
Can I build a game with Genie 4?
Not in the way you would with a game engine. Genie 4 has no game logic, saved state, scoring, or export to an engine, and Google says visual details and object permanence can drift over time. It is useful for exploring ideas and prototyping a look or feel, not for shipping a game.
How does Genie 4 relate to Gemini?
Google says Genie 4 builds on capabilities from the Gemini family. Gemini Omni helps ground simulations in real-world appearance and behavior, other Gemini models help keep scenes coherent, and Gemini adapts your text prompts to the scene. In robotics demos, Gemini 3.8 Flash acted as a planning agent that broke a goal into steps while Genie 4 simulated them. Genie 4 is not a mode inside the Gemini app.
Last updated: