GPT-6 Astra got a robot arm and taught itself to paint

OpenAI pitched GPT-6 Astra as a model that can do "anything you can do on a computer." So someone gave it a body. In a clip going around X, Astra drives a robot arm holding a paintbrush, watches its own strokes through a camera, and paints the Golden Gate Bridge on a real canvas. It worked out the arm's controls itself, and each attempt came out cleaner than the last.

The painting isn't the interesting part. The closed feedback loop is: act, look, adjust, repeat. That's the same screenshot-driven loop behind Astra's computer-use skills, which OpenAI made the headline feature of the 3 Sep launch. It had already been shown painting in Krita from screenshots. Swap the screen for a camera and the mouse for an arm, and the loop still holds.

Some caution before anyone calls this a robotics breakthrough. This is a user demo, not an OpenAI release. It comes with no paper, success rate, control latency or failure count. Skeptics point out that the gap between "describes how to hold a brush" and "reliably does it" is where most robotics hype dies. Astra is already contentious for another reason: its opaque recurrence reasoning is harder to audit through chain-of-thought monitoring.

Why it matters: if general-purpose agents can learn a new body just by watching themselves, robotics stops needing a custom model for every machine.