What shipped

Ask Muse Code for an object and it now saves a textured GLB to your project. That is the whole pitch of the 3D tools Meta switched on inside its CLI coding agent on November 20: the same session that writes your game code can also produce the props in it, from a text description or a reference image, with no separate 3D service in between.

The models behind the session, Muse Wonder, Muse Spark and Muse Ember, invoke the 3D tools as built-in calls, next to Muse Image, per AI Compass's summary of Meta's announcement. Output lands in the repository as a textured GLB file.

The allowance is the real spec

Meta sells this as part of the subscription rather than as a metered extra, which makes the plan limits the number worth reading. On the $5-a-month Everyday Usage plan, Meta's plan documentation, as summarized by AI Compass, supports roughly 30 textured models per five-hour window with the faster Nano 3D tier — an estimate drawn from the shared allowance, not a guaranteed allotment. The paid steps above it are multipliers in the same account: High Usage at $15 a month carries five times the Everyday capacity, Power Usage at $50 carries twenty times and adds Muse Wonder Full, and Maximum Usage at $200 carries eighty times. The higher-resolution Pro 3D setting sits on the upper plans.

The five-hour window also means a heavy generation session competes with the coding prompts the same allowance pays for. Generation is not metered separately; it is rationed jointly.

What it cannot do

A textured mesh is not a character. Per AI Compass's account of Meta's documentation, Muse Code has no built-in tool call for skeletal rigging or animation, and Meta's own guidance for animated characters is a prompt workflow that hands the mesh to Mesh2Motion skeletons and Blender's automatic weights outside the agent. Anything that needs to move still goes through a conventional 3D toolchain for that step.

None of the sources reviewed cites a quality benchmark for the meshes themselves: no polygon counts, no texture resolution figures, no comparison against dedicated text-to-3D services. Until measurements exist, the output is best judged by the two sample projects Meta ships, a Dungeon Crawler and a 3D web viewer.

Where this leaves the workflow

The change that matters is placement, not raw capability. Text-to-3D tools have existed as separate subscriptions for a while; putting generation inside the coding agent removes the manual transfer step for the rough version of an asset. The finishing work, rigging, cleanup and animation, stays where it was.

Sources

  1. [1] Meta for Developers, Developer News listing and Muse Code documentation (developers.meta.com / dev.meta.ai, read via search index), November 2026Read source
  2. [2] Meta for Developers, launch post, LinkedIn, November 20, 2026Read source
  3. [3] AI Compass, November 20, 2026 edition (summary of Meta's "Build 3D models with Muse Code" post and Muse Code plans page)Read source