Breaking

GPT-6 Astra Generates Two Runtime Trains in Three.js With No 3D Model Files

September 5, 2026 at 23:46 EDT

Foundation Models
Software Dev & Coding
Media Generation

OpenAI's GPT-6 Astra generated two trains entirely at runtime from TypeScript in Three.js, with no 3D model files. Wheel motion and explode-and-reassemble animations were code-driven, and the scene ran smoothly in a web browser.

Sep 5, 2026 · GPT-6 Astra

Trains from code, not model files

GPT-6 Astra generated two locomotives at runtime in Three.js from TypeScript. Wheel motion and explode-and-reassemble animations were entirely code-driven, with no 3D model files.

How the trains were built

TypeScript specs

Dimensions and profiles

→

Geometry functions

Shapes drawn in code

→

Runtime trains

Two locomotives, no files

→

Code-driven motion

Wheels, explode, reassemble

Trains generated at runtime

3D model files used

3,295

Editable objects in a related Blender train

95.9%

Astra score on BenchCAD

BenchCAD geometric overlap

Higher is better. Column height is scaled to each published score.

95.9%

GPT-6 Astra

84.3%

Fable 5.1

83.3%

GPT-5.6 Sol

Astra leads BenchCAD by 11.6 points over Fable 5.1 and 12.6 points over GPT-5.6 Sol.

What developers like

Living worlds in code

No imported 3D model files
Smooth playback in the browser
Geometry stays editable in TypeScript

What still breaks

A human pass is still needed

Rounded forms can still need hand tweaks
Photo reconstructions can drift from source details
Fable 5.1 still leads some front-end design work

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.