Meta Muse Wonder: the AI demo you can actually play

A picture with rules you can test

Meta describes the kitchen as a playable 3D game built from customizable image inputs. We opened its first level: a compact room, a small chef, ingredient stations, a recipe order and a tutorial. The page exposes movement, pick-up and use controls. That is a more informative object to inspect than a launch-film montage.

The published artifact does not reveal the number of attempts, the human corrections or the full build history. Opening it verifies that this particular example loads; it does not independently reproduce Meta’s creation process. Its usefulness is that the claim has something tangible attached to it: a reader can inspect the scene and try the rules.

Sources: Meta Research [1] · Meta Research [2]

The game is a software result

Wonder accepts text, images, video and audio; its documented output is text. Code, tools and a runtime are the bridge to an interactive scene. Meta separately offers Muse Image for generating and editing pictures. Treating all of this as a new video generator would hide the most useful technical distinction.

In a prototype, a door can have a position, a collision boundary and a condition for opening. Those are decisions that can be inspected and changed. A rendered image can suggest the same door beautifully without providing any of that behavior. For an interactive pitch, the difference is whether a client is approving a look or experiencing a small system.

Sources: Meta [3] · Meta Developers [4]

A million tokens still need the right decision

Active Memory is Meta’s name for retrieving important context across conversations. Active Focus describes sustained work through blockers and changing information. Developer documentation lists a 1,048,576-token context window. Capacity and memory reliability are different questions: storing more material does not establish that the latest approval will win over an obsolete instruction.

Consider a simple art-direction change. The first brief specifies a blue counter; the approved revision makes it brushed steel but keeps the blue pot. Later, someone asks for warmer light. A useful assistant should retrieve the current material decision and leave the pot alone. That is a concrete evaluation scenario, not a result we measured with Wonder.

Sources: Meta Research [1] · Meta Developers [4]

The revealing detail is in the model card

Meta lists visual and spatial reasoning, especially 3D, among known limitations. Its MemoryBench is an internal test of 64K–100K-token conversations. The SWE-Marathon comparison also uses different agent products rather than a common setup. These qualifications matter when translating a leaderboard into expectations for your own project.

A convincing kitchen and imperfect spatial reasoning can coexist. A selected scene may work well while a changed layout exposes a problem. The practical review is therefore local: can the character reach the station, does a moved object remain usable, and does a revision preserve the rest of the scene? Those checks turn visual confidence into something testable.

Sources: Meta [3]

What you can use, and what it costs

As checked on 13 October, Wonder is marked preview in the docs. It is offered through Meta Model API and Muse Code. API pricing is $4 per million input tokens, $0.20 for cached input and $20 for output. Wonder Full is listed for Muse Code’s $50 Power Usage and $200 Maximum Usage monthly plans; these are usage plans, not promises of unlimited work.

Consumer access needs care: Meta’s research blog and Newsroom say Muse rollout is coming soon, subscribers first; Zuckerberg’s launch thread says subscribers can use it today. That wording does not confirm availability for every account or region. Check the actual product before buying a plan for this specific model.

Sources: Meta Developers [5] · Meta Developers [4] · Meta Newsroom [6] · Mark Zuckerberg / X [9]

An interactive pitch is the interesting next step

For a director or designer, the promising use is a small, inspectable prototype: a set layout, a branching presentation or a scene with one meaningful interaction. It could make a discussion more precise because the team can point to a behavior, change it and compare the result. That is a workflow hypothesis worth testing, not a claim that Wonder delivers production-ready film assets.

The kitchen earns attention because a reference becomes something a person can act inside. The next useful measure is how much of that intent survives a change. If the system can carry an approved decision through code, tools and revision, the creative gain will be felt in the work between the demos.