OpenAI says two of its hardware engineers used GPT-6 Astra to design a working device for talking to its Codex coding agent in under a week. The company described the side project in a November 3 post, "Accelerating engineering for the physical world," saying Astra handled the enclosure design, component selection, circuit board layout, firmware and user interface.
The device is a push-to-talk puck with a small display, built so OpenAI researcher Justin Lin can check on his coding agents without opening a laptop. OpenAI refers to it internally as the "TamaGPT device experiment."
The design sprint did not produce a factory-ready product. OpenAI said Astra did not consistently account for the manufacturer's production limits, tolerances or assembly details, and preparing the device for production took several rounds of engineering review and revision.
Key facts
- 01Two OpenAI hardware engineers used GPT-6 Astra to design a device for talking to Codex in under a week, as a side project.
- 02Astra produced the enclosure, component selection, PCB layout, firmware and user interface for the push-to-talk puck.
- 03A contract manufacturer built a small prototype batch in under a month after several rounds of engineering review.
- 04On BenchCAD 2.0, an unreleased OpenAI internal model scored 0.875 versus 0.594 for GPT-6 Astra at its highest reasoning setting.
- 05Token usage across OpenAI's hardware teams has grown 23x since February, the company said.
After that review, a contract manufacturer produced a small batch of prototypes in under a month.
“Obviously, took a few rounds with the factory, but this is what came back.”— Justin Lin, OpenAI researcher
OpenAI paired the anecdote with benchmark numbers. On BenchCAD 2.0, which tests whether models can reconstruct 3D parts and assemblies and recover connections from circuit board images, an unreleased OpenAI internal model scored 0.875. GPT-6 Astra at its highest reasoning setting scored 0.594. OpenAI has not shared details about the internal model.
BenchCAD 2.0 was introduced on October 12 and has six task types. Its Core set contains 100 held-out cases scored automatically against reference answers, without an AI judge. The benchmark measures geometry, part placement and circuit connectivity, not manufacturability or physical performance.
OpenAI said its hardware and infrastructure teams now use Codex and ChatGPT for CAD models, manufacturing drawings, simulation setup and design review across robotics and compute infrastructure. Token usage on those teams has grown 23x since February. OpenAI engineer Grace Lam described faster design revisions, conversion of 3D designs into production-ready 2D drawings and less manual simulation setup.
OpenAI is also testing the work with outside partners. Chip Ganassi Racing is using it to turn 3D scans into editable CAD and 2D drawings, and Mazda has joined as a research partner with its aerodynamics and crash-safety engineers.
The company is inviting other engineers and hardware firms to apply to a computer-use research collaboration program. Separately, OpenAI has said it is building its own consumer device, positioned between an AI speaker and a smartphone.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
