The model can generate 3D models using TypeScript and Three.js code. In one instance, a camera was decomposed into 122 component groups and 1,877 independent parts over three hours. Another example involved a Tesla vehicle, which was broken down into 334 parts based on official data.
Astra demonstrates advanced computer use capabilities. It successfully recreated a portrait in Canva by controlling the browser interface for about an hour. It also played 'River Flows in You' on a virtual piano by coordinating mouse clicks and keyboard inputs, and subsequently edited the demonstration video.
In robotic tests, Astra achieved a 95% success rate in placing blocks into a bowl across 20 attempts, compared to 40% for Fable 5.1. The model used 16% of the tokens and 43% of the cost of Fable 5.1 in a zero-shot test.
