OpenAI says its hardware team used GPT-6 Astra to design a device for talking to the Codex coding agent. The work was completed as a side project in under a week, covering the device’s physical design as well as its software.
The project offers a concrete example of how OpenAI says its models are being applied to hardware work. It also highlights a boundary: designs produced with AI still had to be revised to account for the realities of manufacturing.
One model handled the design workflow
According to OpenAI, Astra contributed across the full design process. That included creating the enclosure, laying out the printed circuit board, selecting components, routing traces, and developing firmware and the user interface.
These tasks connect different parts of product development. The enclosure defines the device’s physical form, while the board layout and component choices determine how its electronics fit together. Firmware and the interface provide the functions people interact with.
OpenAI describes the device as a way for users to talk to Codex. The source does not provide further details about how the prototype works or what its final form looks like.
Manufacturing required several rounds of revision
A contract manufacturer produced a small batch of prototypes in under a month. Getting from the initial design to those prototypes was not a single-step process: several revision rounds were needed.
OpenAI says Astra did not always account for manufacturing constraints, tolerances, and assembly details. Those considerations matter when a design moves from a model or drawing into parts that must be made and fitted together.
The revisions show the difference between producing a design and producing one that can be manufactured reliably. AI assistance covered many design tasks, but the prototype process still exposed issues that needed to be worked through before the hardware could be built.
OpenAI reports a higher benchmark score for an internal model
OpenAI also points to results on BenchCAD 2.0, a benchmark that tests whether models can reconstruct 3D parts and assemblies and identify connections on circuit boards.
A newer “internal model” scored 0.875 on the benchmark. GPT-6 Astra scored 0.594 at its highest reasoning level. OpenAI has not shared details about the internal model, so the source does not explain how it differs from Astra or what produced the performance gap.
The scores describe performance on the benchmark’s specified tasks. They do not, by themselves, explain how well either model handles the manufacturing and assembly issues encountered in the prototype project.
AI tools are part of broader hardware work
OpenAI says its hardware teams use Codex and ChatGPT to create CAD models, manufacturing drawings, and simulations. The company names robotics and compute infrastructure as areas where this work is used.
Token usage for these activities has grown 23x since February, according to OpenAI. That figure indicates increased use of the tools across its hardware work, though the source does not give a starting or ending token total.
OpenAI is also building its own hardware device, reportedly something between an AI speaker and a smartphone. The article does not establish whether the Codex device designed with Astra is connected to that separate effort.
Together, the project and benchmark results point to an expanding role for AI in hardware design at OpenAI. The prototype story is also a reminder that generating designs is only part of the process: manufacturing constraints, tolerances, and assembly can call for further iteration.