Fable 5.1 builds a walkable Three.js model of San Francisco's Union Square

I like that Fable 5.1 generated an interactive Union Square and then filmed a tour inside it. You can walk from Powell to Stockton, read storefront signs, cross an intersection, see cable cars, and enter Apple and Nintendo.

Using OpenStreetMap geometry and USGS 3DEP elevation data is how I would have approached it. If Fable writes the conversion code into a usable Three.js model, that removes a lot of setup work.

"World model" was not what I expected here. The term has too many meanings now, though I think the bar is whether it can simulate a world.

I wanted a 2 to 3 minute YouTube demo, especially to see whether the pedestrians and cars are on rails. The repository has a longer one-minute video, and the author linked the relevant logic.

The demo needs time, cost, reliability, failure-rate, and babysitting details. It looks impressive, but those details decide whether it is practical.

This was a one-shot result with a very long prompt that explicitly directed subagents and a self-QA loop. The author says it took about two hours, used about 8 million tokens, and cost about $33 through the API.

I can see this enabling open-world or AR games similar to Pokemon Go or Ingress. The author says they are trying open-world RPGs with real economies, and views this as a hard multimodal coding benchmark.

I still question the point of AI demos that mainly make people say "an AI made this." I have not seen agent-generated software that clears the bar of being independently useful.

A world at this scale would usually take thousands of hours of manual modeling, placement, and quality control. Prompt-to-CAD agents let hobbyists turn more ideas into working artifacts rather than spending a weekend in Fusion 360.

Code-generated worlds avoid some mesh-generation problems because Three.js or Blender scripts can build geometry from primitives, CSG, and parametric construction. Texturing is still weak, though procedural materials cover part of the gap.

I have used this kind of modeling for an RTS. Opus 5 is about as good and cheaper in my experience, but these systems produce high-poly simple geometry; I get better game assets by making low-poly silhouettes and baking detailed textures with Meshy.

There is a fidelity threshold where Opus fails for foliage L-systems, physical mob animation, or plausible detailed interiors. Fable takes longer but can move across levels of abstraction, and prompting it to mesh, instance, and build an art pipeline can improve efficiency.