OpenAI revealed that two team members used the GPT-6 Astra model to design a complete hardware device as a side project in under a week. The prototype is a push-to-talk device with a small screen to interact with the Codex coding agent, internally nicknamed the TamaGPT experiment.
What happened?
According to OpenAI, Astra handled the entire process: enclosure design, component selection, printed circuit board (PCB) layout, firmware generation, and user interface. A contract manufacturer produced a small batch of prototypes in under a month, though several rounds of human review were needed because the model did not always account for production limits, tolerances, and assembly details.
The company also mentioned an unreleased internal model that scored 0.875 on the BenchCAD 2.0 benchmark, versus 0.594 for Astra at its highest reasoning level. Token usage across hardware teams grew 23 times since February.
Why does this matter?
The experiment demonstrates the advance of multimodal models in physical engineering workflows, integrating mechanical, electrical, and software design that normally requires separate specialists. Although still dependent on human review for manufacturing, it drastically reduces prototyping time.
OpenAI is also collaborating with Mazda on aerodynamics and vehicle safety simulations, and with the Chip Ganassi Racing team to convert 3D scans into editable CAD models.
What changes in practice?
For engineers and designers, tools like Astra can accelerate hardware development cycles, from consumer devices to infrastructure. However, the need for human iteration highlights that AI is still an assistant, not a complete replacement.
The case reinforces the trend of generative AI entering the physical world, beyond software.
By Grok
