No 3D Models Imported: This Little City Was Built with Three.js Code

After the release of Three.js r187, developer Mohamed used Codex and GPT-6 Astra for a stress test: build a small city without importing any 3D models. The tweet says the result surprised even him. First, watch the video.

The video shows an aerial view. White buildings form blocks, elevated tracks connect the buildings into a ring, and at the center is a fountain plaza with a few tables and chairs. Green trees and lawns are interspersed, and several roads run through the entire area. The overall palette is white, clean, and orderly. The source material claims that the entire small city was built with Three.js, without importing any external 3D models.

A bit of context. Three.js is a common library for doing 3D in the browser, and r187 is the version number. Codex is OpenAI's coding tool, and GPT-6 Astra is the model mentioned in the source material. The tweet does not elaborate on the prompt, generation steps, or performance data, so we can only discuss the result as a result.

Demos like this can easily pull your attention to the visuals. What's really worth looking at is the zero external models. Geometry, materials, layout, and scene organization are all described by code, and the asset list is empty. For developers, this means early prototypes don't have to wait for modeling assets first; they can directly change code to adjust layout, proportions, and materials, and the scene can also go into version control and be reproduced by parameters. It is suitable for validating spatial direction and visual tone, and also for testing capability boundaries.

But a demo video is not the same as a usable product. Between a beautiful city and a city that can actually run, there are interaction, performance budgets, collision, resource management, and long-term maintenance. The further procedural generation goes, the more it has to answer which parts should be parameterized and which should enter the asset pipeline. Watch enough of these and it's easy to think that 3D content production is down to a single prompt. In reality, someone still has to read the code, and someone still has to test the performance.

Original post: https://x.com/OpenAIDevs/status/2116673933990912217