A project called fable51-worlds shows autonomous Claude Fable 5.1 agents researching, modeling, and quality-checking browser-native 3D reconstructions of real places — Union Square in San Francisco and Higashiyama in Kyoto — then shipping them as plain Three.js apps. It is code-generated geometry, not a video or image model.
2026-09-03 · by Moe Ameen
PhiloLabs published fable51-worlds, an open-source project (MIT-licensed) that uses autonomous Claude Fable 5.1 agent swarms to create explorable, browser-native reconstructions of real places and ship them as plain Three.js applications. The tagline is literal: "worlds via code, from fable 5.1." Rather than a model that renders pixels, this is a coding agent that writes the geometry, the scene assembly, and the runtime — every building, storefront, sign, tree, and traffic light is generated from open data and public reference imagery by code.
The pipeline runs in four stages. Research: parallel agents gather OpenStreetMap geometry (ODbL), USGS 3DEP elevation, transit specs, and storefront data with source attribution. Asset generation: Blender scripts (bpy) emit optimized GLB kits — façade modules, street furniture, vehicles, vegetation. Runtime assembly: a pure Three.js app combines terrain, streets, façades, props, crowds, and traffic from JSON specs. Quality assurance: Playwright-driven camera matching screenshots fixed viewpoints and compares them against photographs taken from the same spots, and independent reviewer roles (architect, geographer, technical artist) file reports.
Two worlds ship in the repo. Union Square, San Francisco: 453 building footprints, 75 hand-authored façades, 129 named storefronts, 220 pedestrians, 109 vehicles, working traffic lights and cable cars, day/sunset/night modes, two explorable interiors (Apple and Nintendo), and 34 camera-matched validation points. Higashiyama, Kyoto: 266 buildings, 471 shopfronts, a 2.3 km continuously walkable route, a cel-shaded anime look, and a self-contained single-HTML deployment option. According to the creator, a single world ran in roughly two hours for about 8 million tokens and around $33 in API cost — and the agent even filmed its own walkthrough tour video inside the finished scene.
Keep the framing straight. This is a demonstration of Fable 5.1's agentic coding applied to spatial reconstruction, built by PhiloLabs — not a new Anthropic product, and not a generative video or image model in the Sora or Veo sense. Early reactions praised the achievement while noting the honest limits: topology can be messy and texturing is hard, and it is a research-grade showcase, not a consumer app you sign up for. Treat the specifics above as a snapshot of the repo on the authoring date and check the source for current figures.
The part of this a creator can actually post is the walkthrough video — the tour the agent filmed inside its own generated world. A Three.js app does not go to TikTok, but that fly-through is real footage, and footage is where [Kompozy](/) starts. Screen-record the world tour (or export the agent's walkthrough), drop it into Kompozy as a source, and one clip becomes a run of finished pieces held to a single [Persona Brief](/glossary/persona-brief) so they read as one creator: vertical [clipped shorts](/glossary/clipped-short) cut from the long fly-through, [Persona Shorts](/glossary/persona-shorts) where your talking-head avatar narrates over the world, quote graphics pulled from the best on-screen moments, a blog article on how it was built, and an email newsletter. [Autopilot](/glossary/autopilot) then schedules and publishes the set across the eight social platforms plus blog and email behind a per-post review.
That review gate matters here for one honest reason: AI-built worlds are the kind of thing an audience wants disclosed. Kompozy's per-post step is where you add your own AI-disclosure line before anything ships. Fable 5.1 World Modeling builds the spectacle; Kompozy turns the two minutes of footage it hands you into a week of [repurposed](/glossary/content-repurposing), captioned, on-brand posts and gets them out the door.
It refers to PhiloLabs' fable51-worlds project, which uses autonomous Claude Fable 5.1 coding agents to research, model, and quality-check explorable 3D reconstructions of real places, then ship them as plain Three.js browser apps. It generates the scene as code from open data — it is not a generative video or image model.
No. Fable 5.1 is Anthropic's text-and-reasoning model (released September 1, 2026); fable51-worlds is a third-party project by PhiloLabs that applies Fable 5.1's agentic coding to 3D reconstruction. It is a demonstration of the model's coding ability, not a separate Anthropic world-generation product.
The creator reports that a single world — for example Union Square, San Francisco — ran in roughly two hours for about 8 million tokens and around $33 in API cost. Treat that as a self-reported snapshot; your cost depends on the model rates and the scope of the scene.
Not directly — the output is a Three.js web app, not a video file. The shareable asset is the walkthrough video the agent films inside the world. You can screen-record or export that fly-through and then use a tool like Kompozy to clip, caption, and publish it across platforms.