OpenAI says two employees used AI to design a hardware prototype in under a week

Two OpenAI employees used GPT-6 Astra to design a physical device for talking to its Codex coding agent in under a week, the company disclosed on November 3. The side project covered electronics, mechanical design and software, but reaching a manufactured prototype still required human review and revisions—a significant boundary on the experiment’s automation.

One device, several engineering disciplines

OpenAI framed the project as a test of whether Astra could create a device without direct human edits. That was the experiment’s goal, not its final outcome. The company used GPT-6 to define functional requirements and develop the enclosure, printed circuit board and software connections needed to let users speak to Codex.

The work went beyond generating a visual concept. Astra had to coordinate the device’s outer form with the components inside it, then supply the software that made the physical interface useful. OpenAI described four areas of work:

• Appearance: Astra developed the look using references from games, toys and OpenAI’s recent Work Louder collaboration.
• Mechanical design: it designed the enclosure and the internal structures that hold components in place.
• Electronics: it selected components and designed the printed circuit board, including component placement and electrical connections, to fit inside the enclosure.
• Software: it designed the user interface and generated firmware—the software running on the device—and integrations connecting it to Codex.

The physical build followed a different timetable from the design sprint. A contract manufacturer produced a small batch of prototypes in under a month. Preparing the device for that step took several rounds of review and iteration, rather than a straight handoff from the model to the factory.

Astra did not consistently account for the manufacturer’s production limits, tolerances, or details required for assembly.

OpenAI, Accelerating engineering for the physical world

Better design scores are not a factory test

Alongside the prototype, OpenAI published engineering benchmark results for a newer internal model. It reported an average score of 0.875 on BenchCAD 2.0, compared with 0.594 for GPT-6 Astra at its highest reasoning setting. Those are company-reported evaluation results, separate from the account of building the Codex device.

BenchCAD 2.0 evaluates whether models can create three-dimensional parts and assemblies from engineering drawings or rendered images. It also tests whether they can recover electrical connections from images of circuit boards. Its creator’s October 12 introduction describes six task types, separating individual-part modeling, assembly work and circuit reconstruction rather than treating engineering as one undifferentiated skill.

The benchmark’s Core set contains 100 held-out cases, shared only under an evaluation-only agreement to limit contamination. Answers are scored automatically against references, without an AI judge. Scores measure how closely outputs match those references; they are not a count of finished products successfully built.

That distinction matters for this demonstration. BenchCAD’s scoring covers geometry, placement of parts and circuit connectivity, not manufacturability or physical performance. A higher score therefore does not establish that a model has solved the production problems OpenAI encountered. OpenAI itself describes the engineering lifecycle as extending beyond design.

More use inside OpenAI, measured in tokens

The prototype sits alongside existing internal engineering work. OpenAI says its hardware and infrastructure teams use Codex and ChatGPT to create and revise computer-aided design models, prepare manufacturing drawings, configure simulations and review designs. Those workflows span robotics and computing infrastructure, rather than only small experimental devices.

Usage across its hardware teams has grown 23-fold since February, measured in tokens—the units of text models process. That figure measures consumption, not a corresponding gain in productivity. OpenAI engineer Grace Lam described faster design revisions, conversion of 3D designs into production-ready 2D drawings, and less manual work setting up simulations. Her account describes benefits inside OpenAI, rather than an independently measured time saving.