What Genie is
Genie takes a prompt, text or an image, and produces an environment you can walk through with a keyboard. It does not build a 3D scene and render it. It predicts the next frame from the previous frames and your input, many times a second, so the world exists only as long as the model keeps generating it. That is the definition of a world model in its most literal form: given a state and an action, predict the next state.
The version history is mostly a story about how long the model can remember what it already drew:
- Genie 1 (early 2024) made 2D platformer-style worlds at about one frame per second. A research result, not something you could play.
- Genie 2 (December 2024) added 3D, but scenes held together for only 10 to 20 seconds.
- Genie 3 (August 2025) was the first that ran in real time: 720p at 24 frames per second, consistent for a few minutes, with visual memory reaching back about a minute.
- Project Genie (29 January 2026) put Genie 3 in front of paying users: Google AI Ultra subscribers in the US, 18 and over, with sessions capped at 60 seconds. Video-game stocks including Unity and Roblox fell sharply the next day.
- Genie 4 (27 October 2026), covered below.
There is also a derivative worth knowing about. In February 2026 Waymo announced the Waymo World Model, built on Genie 3, to simulate rare driving situations for its robotaxis and output lidar as well as camera views. That is the first time a Genie model became infrastructure for somebody else’s product.
What changed in Genie 4
Google DeepMind’s announcement and TechCrunch’s coverage list four changes:
- Native audio. Music, ambient sound and dialogue are generated with the frames. Google calls it its first real-time world model to do this.
- Steering in plain language. While the world runs you can type "make it rain" or add a person, and the scene adapts. Keyboard controls still handle movement and camera.
- Longer consistency. Up to about two minutes, double Genie 3.
- Gemini as director. A Gemini model watches the scene, breaks a goal into steps and checks whether each one worked, while Genie draws the frames. In one demonstration a Gemini Robotics model planned a task for Apptronik’s Apollo humanoid and Genie 4 simulated it in real time.
What Google did not publish matters as much. There is no resolution, no frame rate, no benchmark and no comparison with Genie 3 in the official post. Google lists its own limits: detail and object permanence drift in longer sessions and after several edits, and complex physical interactions, audio accuracy, legible text and long-term consistency are still open problems. Access is a research preview for trusted testers, with Google AI Ultra subscribers (in Project Genie) and Gemini API and AI Studio early-access partners next.
Four bets that share one name
The phrase "world model" now covers at least four kinds of system. They are not competing for the same customer, so ranking them on one axis produces nonsense.
- Interactive generators produce pixels you can steer in real time. Genie and Runway’s GWM Worlds are here. The customer today is a game designer, a film previz team or a curious subscriber.
- Persistent 3D scene generators produce a world once and hand you the asset. World Labs’ Marble exports Gaussian splats and meshes that load into a game engine. Nothing drifts, because nothing is regenerated.
- Predictive planners never draw a frame. Meta’s V-JEPA 2 and the approach Yann LeCun took to AMI Labs predict in an abstract representation, which is cheaper and is what a robot needs to choose an action.
- Training-data engines generate physically plausible scenarios to train robot and vehicle policies. Nvidia Cosmos and the Waymo World Model are here.
Genie straddles the first and fourth categories, which is why Google keeps showing both a game demo and a robot demo. Most competitors have picked one.
Genie vs OpenAI, Meta, Nvidia, World Labs and Anthropic
| Organisation | Model | What kind of world model | Status (Oct 2026) | Access |
|---|---|---|---|---|
| Google DeepMind | Genie 4 (preview); Genie 3 in Project Genie | Interactive generator: playable pixels from a prompt | Genie 4 announced 27 Oct 2026, trusted testers; Genie 3 live for AI Ultra (US) since 29 Jan 2026 | Closed; consumer subscription, Gemini API early access next |
| OpenAI | Sora / Sora 2 (discontinued) | Video generator, pitched in 2024 as a "world simulator" | Shutdown announced 24 Mar 2026; app closed Apr 2026; API end Sep 2026. Team moved to world simulation for robotics | No current public world model |
| Anthropic | None; Model Hardware Standard | Not a world model: an interface for agents to operate real lab hardware | Research preview, late Aug 2026, with HHMI | Open-source release planned |
| Meta / AMI Labs | V-JEPA 2 (Meta); AMI Labs (LeCun) | Predictive model in an abstract embedding space, for planning rather than pixels | V-JEPA 2 released Jun 2025. LeCun left to found AMI Labs (Dec 2025, Paris), $1.03B seed, no product expected for years | V-JEPA 2 open; AMI none yet |
| World Labs (AMD) | Marble; Atlas | Persistent 3D scene generator with exportable assets | Marble Nov 2025; Atlas omni model Sep 2026. AMD agreed to buy the company for $8.2B in stock (late Sep 2026) | Commercial product |
| Nvidia | Cosmos 3 (Super, Nano; Edge coming) | World foundation model for robot and AV training data | Launched at Computex, Jun 2026 | Open weights on Hugging Face |
| Runway | GWM Worlds 2 | Interactive generator, closest rival to Genie | Research preview, Sep 2026: 720p, 24 fps, synchronized audio, no fixed session limit | Business enquiries only |
OpenAI: from "world simulator" to shutdown
OpenAI used the language first. Its February 2024 Sora report was titled Video generation models as world simulators, and Sora 2 followed in September 2025 with a social app. On 24 March 2026 OpenAI announced it was discontinuing Sora; the app closed in April and the API was scheduled to end in September. Reports linked the decision to compute cost, and an OpenAI spokesperson told CNN that "the Sora research team continues to focus on world simulation research to advance robotics." As of this writing OpenAI has no public world model, so "Genie vs Sora" is a historical comparison: Sora produced a clip you watched, Genie produces a world you move through.
Anthropic: not in this race
Anthropic has not released a video or world-simulation model, and I could find no public research programme for one. Its physical-world move in 2026 was the Model Hardware Standard, previewed in late August with HHMI: a way to describe a microscope or a robotic arm so that a Claude agent can operate it safely. That is the opposite end of the problem. Genie imagines a world so an agent can practise; the hardware standard lets an agent act on the real one. If you are comparing vendors for a robotics or lab automation stack, the two can sit in the same architecture without competing.
Meta and AMI Labs: predict, do not draw
Meta’s answer is V-JEPA 2, released openly in June 2025. It is trained on video but predicts in an abstract representation rather than generating pixels, and Meta showed it planning robot pick-and-place actions in environments it had not seen. The argument behind it, made most loudly by Yann LeCun, is that generating every pixel is wasted effort when what an agent needs is the consequence of an action. LeCun left Meta and founded AMI Labs in Paris in December 2025 to pursue that bet; it raised a $1.03 billion seed round in March 2026 and does not expect a sellable product for years. If Genie is right, AMI’s approach is too abstract to be useful. If AMI is right, Genie is spending most of its compute on rendering.
Nvidia: the open option for robotics teams
Cosmos 3, launched at Computex in June 2026, is the one most robotics companies can actually download. It ships open weights in Super and Nano sizes (Edge is coming), handles text, images, video, sound and actions, and is built to generate training scenarios and serve as a backbone for robot policies. Its launch partners include Skild AI, Agile Robots, Doosan, LG and Samsung. Nvidia’s incentive is plain: every synthetic training hour runs on its hardware. Genie is the better demo; Cosmos is the one you can put in your own pipeline today.
World Labs: a world you keep
Fei-Fei Li’s World Labs launched Marble in November 2025 and the Atlas omni model in September 2026. Marble’s worlds are persistent and exportable, which solves Genie’s drift problem by not regenerating the scene at all, at the cost of interactivity. In late September 2026 AMD agreed to buy World Labs for $8.2 billion in stock, expected to close by the end of the year. That makes it the second chipmaker, after Nvidia, to own a world-model strategy outright.
Runway: Genie’s closest like-for-like rival
If you want the same category as Genie, it is Runway. GWM Worlds 2, shown about two months before Genie 4, streams a playable world at 720p and 24 frames per second with synchronized audio and, Runway says, no fixed session limit. It is a research preview with no self-serve access. Decart is the other real-time player; Amazon is reportedly in talks to buy it.
How to read the demos
Every one of these launches came with video, and none came with the number that would settle the question. For an interactive generator, the useful measures are how long a scene stays consistent and how quickly it responds to input; Google publishes the first loosely and the second not at all. For a robotics world model, the measure is transfer: train a policy inside the model, run it on hardware, and report how far it degrades. I made this argument in the world models explainer and the Genie 4 launch does not change it. A convincing two-minute walk through a generated village shows that the model has learned what plausible video looks like. It does not show that it learned physics.
What I would do with this
- Games, film and training content: get on the Project Genie waitlist and test Genie 4 against Runway on your own prompts. Treat output as previsualisation, not shippable assets; for assets, Marble’s exports are the better fit.
- Robotics and autonomy: start with Cosmos 3 because you can run it in your own environment, and ask any vendor pitching a closed world model for sim-to-real transfer numbers on a task like yours.
- Agents that touch hardware: the world model is the practice field, not the control layer. Interfaces like Anthropic’s hardware standard, and the safety gates covered in the physical AI hub, decide whether anything ships.
- Vendor risk: two of the seven efforts in the table changed hands or shut down in 2026. Do not build a pipeline that depends on one closed preview.
Related
- World models: what they are, and the full list of every notable world model
- World Labs Marble and Atlas: persistent 3D worlds, and the AMD deal
- AMI Labs and JEPA: the predict-don’t-draw alternative
- Nvidia Cosmos: the open option for robotics teams
- World model startups: Decart, Odyssey, General Intuition and Runway
- Physical AI: the cluster hub, and what actually gates deployment
- AI limitations: why fluent output is not evidence of a correct model underneath
Frequently asked questions
What is Google Genie?
Genie is Google DeepMind’s family of world models: systems that generate an interactive environment from a text or image prompt and keep generating it, frame by frame, as you move through it. Genie 1 (2024) made 2D platformer-style worlds; Genie 2 (December 2024) added 3D but held a scene for only 10 to 20 seconds; Genie 3 (August 2025) ran in real time at 720p and 24 frames per second; Genie 4 (October 2026) added native audio and natural-language steering.
What is new in Genie 4?
Google DeepMind announced Genie 4 on 27 October 2026 as a research preview. It generates music, ambient sound and dialogue alongside the visuals, lets you change weather, camera or add people and objects by typing while the world runs, and holds a consistent scene for up to about two minutes, which TechCrunch reports is double Genie 3. Gemini acts as a director that plans steps and checks whether each one worked. Google did not publish a resolution, frame rate or benchmark scores.
Can I use Genie 4 or Genie 3 today?
Genie 3 is available through Project Genie, a Google Labs site launched on 29 January 2026 for Google AI Ultra subscribers in the US aged 18 and over, with sessions capped at 60 seconds. Genie 4 is with trusted testers first. Google says it will reach Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio in the coming weeks, and a waitlist is open.
Is Genie 3 better than OpenAI Sora?
They were never the same product. Sora generated fixed video clips from a prompt; Genie generates a world you can move through and change as it renders. The question is also moot: OpenAI announced on 24 March 2026 that it was discontinuing Sora, closed the app in April and scheduled the API to close in September 2026. OpenAI says the Sora team continues world simulation research aimed at robotics, but it has not released a world model.
Does Anthropic have a world model?
Not publicly. Anthropic builds the Claude language models and has not released a video or world-simulation model. Its main physical-world announcement in 2026 was the Model Hardware Standard, previewed in late August, which describes lab instruments and robotic arms so that AI agents can operate them. That is a control interface for real hardware, not a learned simulator.
Which world model is best for robotics?
For training robots, the relevant comparison is not Genie against Sora but Nvidia Cosmos, Meta’s V-JEPA 2 and purpose-built models such as the Waymo World Model, which Waymo built on Genie 3 for driving edge cases. Cosmos 3 ships open weights, V-JEPA 2 is open and was shown planning robot actions zero-shot, and Genie 4 has been demonstrated simulating a humanoid task. None of them has published the number that matters most: how much a policy trained inside the model degrades on real hardware.
Who is leading the world model race in 2026?
It depends which race. Google DeepMind leads on real-time interactive generation and has the widest consumer access via Project Genie. Nvidia leads on open models used by robotics companies. World Labs leads on exportable 3D scenes and is being acquired by AMD for $8.2 billion. Yann LeCun’s AMI Labs has the most capital behind the predictive, non-generative approach, but expects no product for years.
Sources: Google DeepMind (Genie 4 announcement, 27 October 2026), TechCrunch, Ars Technica, The Verge, Bloomberg, OpenAI, CNN, Meta AI, ITPro, AI Business, GamesBeat, AlphaSignal. Specifications are as published by each company; none of the models were independently benchmarked for this page.