Direct answer: Choose Genie 3 when the output needs to react to input in real time; choose a video model when you need a fixed clip to publish.

What Is the Core Difference?

A video model generates a film. It runs once, produces a clip, and that clip is the same forever. A world model generates a simulation. It runs continuously and the next frame depends on what you do.

That single distinction drives every downstream difference — interactivity, reproducibility, export, and cost model.

Feature Comparison

Notice the access row. A video model you can use today. Genie 3 you likely cannot. For most teams that settles it before any technical factor matters.

Which Use Cases Map to Which?

Pick a video model for: ads, social clips, product demos, storyboards, anything where the deliverable is a file someone watches. You need a fixed frame count and a downloadable format.

Pick a world model for: training embodied agents, testing navigation policies, and rapid vibe-checking a game environment before an artist builds it. The point is the interaction, not the footage.

If your goal is a looping background for a landing page, that is video. If your goal is an agent that learns to walk through a room, that is a world model.

How Stable Is a Generated World?

Not very, and this surprises people. Genie 3 holds a scene together for a few minutes, but consistency fades as you move. Objects you walked past may not be there on return. In a video clip this problem cannot occur, because there is no revisiting — the camera only goes forward.

So if your use case requires a coherent map the user explores freely, current world models are not ready. If it requires a linear shot, they were never the right tool.

What About Cost and Access?

Video generation is metered by clips or credits, and you can budget per output. World models are research software: cost is unclear because access is restricted, and there is no per-seat pricing to plan around. For a team that needs predictable costs, that alone rules Genie 3 out today.

Where This Is Heading

The split will not last. Expect world models to gain export and lengthen their stable sessions, and expect video models to add more interactive controls. For now, treat them as separate decisions and do not let a slick demo push you toward the wrong category.

Quick Summary

• Video models produce fixed clips; world models produce interactive simulations.
• Genie 3 is research-preview only; video tools are broadly available.
• Choose video for publishable deliverables, Genie 3 for agent training and prototyping.
• World consistency degrades within a few minutes of movement.
• Predictable per-clip budgeting favors video models today.

Direct answer: Genie 3 generates a real-time, walkable 3D scene from a text prompt; access is limited through Google’s research preview, so you apply rather than sign up freely.

What Is a World Model?

A world model predicts what happens next in an environment. Feed it a prompt plus your movement input, and it generates the next frame consistent with both. Unlike a video generator, the scene responds to you — walk toward a door and it stays a door.

Genie 3 keeps the world stable for a few minutes rather than seconds, which is the core improvement over earlier world models.

How Do You Get Access?

1. Genie 3 is not a public product; it is a research preview.
2. Access is granted to selected researchers, studios, and partners.
3. Google has not opened a general free tier.
4. Early collaborators include game studios testing it for rapid prototyping.

So the honest answer is: you probably cannot try it today. That is a real drawback, not a footnote. Watch DeepMind’s announcements for widened access.

What Can Genie 3 Actually Generate?

Prompt-driven environments with interactive physics. Examples shown publicly include driving scenes, robot navigation setups, and stylized outdoor terrain. You can also prompt for environmental changes mid-session — add rain, shift from day to night — and the world adjusts.

Output resolution is 720p at about 24 fps. That is watchable but not photoreal, and it is far below what a native game engine renders.

How Does It Compare to a Game Engine?

The comparison misses the point though. Genie 3 is for exploring ideas, not shipping them. You cannot export a mesh or hand a scene to an artist.

What Are the Known Limits?

Three that matter. Sessions are short — consistency degrades after a few minutes of movement. There is no asset export, so generated worlds die with the session. And interactivity is limited to movement and prompt-driven environmental changes; you cannot script objects or add game logic.

Memory of the scene is also imperfect. Walk far enough and details behind you may not survive a return trip.

Who Is It For — and Who Is It Not For

Aimed at AI researchers studying embodied agents, and at game studios prototyping a vibe before committing artists. Not for indie developers hoping to ship a level, and not for anyone needing deterministic, reproducible content. If you need a level you can edit, engine + AI asset tools is still the path.

Quick Summary

• Genie 3 is a DeepMind world model that builds interactive 3D scenes from text.
• It runs at 720p and roughly 24 fps in real time.
• Access is research-preview only; there is no public free tier.
• Sessions last minutes, and there is no scene export.
• Built for research and prototyping, not shipping game content.