OpenAI's hardware team used GPT-6 Astra to design a hardware device that lets users talk to the Codex coding agent from scratch in under a week as a side project. According to OpenAI, the model handled every step of the development process, including enclosure design, PCB layout with component selection and trace routing, firmware, and the user interface. A contract manufacturer subsequently produced a small batch of prototypes in under less than a month, although several revision rounds proved necessary because Astra did not consistently account for manufacturing constraints, tolerances, and assembly details.
In addition to this project, OpenAI highlighted significant performance improvements across its models. A newer internal model achieved a score of 0.875 on the BenchCAD 2.0 benchmark, whereas GPT-6 Astra reached 0.594 at its highest reasoning level, with the benchmark testing the ability to reconstruct 3D parts and assemblies and identify PCB connections. OpenAI noted that its hardware teams routinely utilize Codex and ChatGPT for computer-aided design models, manufacturing drawings, and simulations spanning robotics and compute infrastructure, resulting in token usage multiplying by 23 times since February.
This development highlights the expanding role of advanced artificial intelligence models in physical engineering and hardware prototyping workflows.