OpenAI Says GPT-6 Astra Designed a Working Codex Device in Under a Week

OpenAI says two of its employees used GPT-6 Astra to design a physical device for talking to its Codex coding agent in under a week, from the enclosure to the circuit board to the firmware. The company shared the project on Tuesday, November 3, in a blog post and a video on X.

OpenAI says the team did not directly edit Astra’s design outputs. A manufacturer still needed several rounds of revisions before the prototypes worked, which shows how far AI hardware design has come and where it still depends on people.

ChatGPT and Codex are helping engineers build for the physical world.We’re sharing our progress across mechanical design, electrical engineering, and simulation—from helping our hardware teams iterate faster to testing new workflows with industry collaborators. https://t.co/b3rW9zYJGA

— OpenAI (@OpenAI) November 3, 2026

What OpenAI says Astra did

OpenAI says it used GPT-6 to define the device’s requirements and then to design three layers of it:

• Mechanical: the enclosure and the internal structures that hold the parts in place.
• Electrical: the choice of components and the printed circuit board, including placement and connections, sized to fit the enclosure.
• Software: the user interface, the firmware, and the links that connect the device to Codex.

Two team members worked on it as a side project, according to OpenAI. The Decoder reported that a contract manufacturer then produced a small batch of prototypes in under a month.

The researcher’s account

Justin Lin, who works on OpenAI’s research team, fronts the video. He says it was his first time going from a bare board to an assembled PCB, and that he “gave it some plans and it designed the board.” He says it took a few rounds with the factory, and that the team made many parallel designs, including many versions of the outer shell.

Lin says the idea came from not wanting to open his laptop to talk to Codex. He wanted a small “Codex pet” with a display and a push-to-talk button. In the demo he asks it whether it has any active agents, and it answers that it has five.

Where people still stepped in

Astra did not always account for manufacturing limits, tolerances and assembly details, which is why the factory sent the design back several times, The Decoder reported. OpenAI describes the result as a working prototype, not a finished product.

OpenAI also published benchmark numbers. On BenchCAD 2.0, which asks models to rebuild 3D parts and identify connections on circuit boards, GPT-6 Astra at its highest reasoning setting scored 0.594. A newer internal model scored 0.875. OpenAI has not shared details about that model.

The company says its hardware teams’ use of Codex and ChatGPT has grown 23-fold since February. That figure measures tokens used, not output or productivity. All of the numbers come from OpenAI, and the project has not been independently reviewed.

Beyond the desk project

OpenAI says its teams use Codex and ChatGPT for CAD models, manufacturing drawings, simulation setup and design reviews, in areas such as robotics and computing infrastructure. Its post also includes a statement from Mazda about using Astra in Codex to explore dozens of design variations for aerodynamics.

Outside testers have also tried Astra on circuit-board layout. PCB software maker Quilter has collected public tests, and some of those boards were never built, so their results are untested in the real world.

The project follows OpenAI’s push into autonomous agents, which we covered after DevDay. OpenAI is also reportedly building its own consumer hardware device, though it has not said what it will be.