OpenAI has revealed that its GPT-6 and Astra models designed a functional physical device in less than a week, handling everything from the enclosure and component selection to the printed circuit board, firmware, and software. The prototype, a small gadget for talking to Codex, was built by two team members as a side project, and an external manufacturer produced a small batch of prototypes in under a month. The experiment was summarized by OpenAI as “designing a prototype device to talk to Codex in less than a week.”
- Rapid design: GPT-6 set requirements and guided the process, while Astra handled the bulk of engineering, including enclosure, PCB layout, UI, firmware, and Codex integration.
- Manufacturing hurdles: Getting the prototype ready for production required multiple review rounds and engineering work on assembly requirements, exposing AI’s limits with physical tolerances.
- Benchmark gap: On OpenAI’s internal BenchCAD 2.0 test, a newer internal model averaged 0.875, while GPT-6 Astra scored 0.594 at its highest reasoning level.
What the Models Actually Did
According to OpenAI, GPT-6 defined the device’s requirements and steered the design process, while Astra took on the heavy lifting: the enclosure, component selection, PCB routing, user interface, firmware, and integration with Codex. The result was a prototype that connects users to Codex through a dedicated physical interface.
The experiment’s goal was to test whether the models could handle a complete hardware development process without human editing of their work. OpenAI does not claim that goal was fully achieved. It is the same frontier between promise and reality seen elsewhere in the industry, where ambitious automation claims often meet physical-world constraints.
Where the Limits Emerge
The real friction appeared in manufacturing. According to reports, making the prototype production-ready required several rounds of review and engineering work on assembly requirements, and OpenAI acknowledged those shortcomings. On an internal test bench called BenchCAD 2.0, which measures 3D component reconstruction and identification of connections on a board, a more recent internal model averaged 0.875, while GPT-6 Astra reached only 0.594 at its maximum reasoning level.
One metric underscores OpenAI’s appetite for this approach: the company reports a 23-fold increase in its hardware teams’ use of its models between February and October 2026, measured by token consumption. This is not an isolated experiment. The company is working with Chip Ganassi Racing to convert 3D scans into editable CAD models and with Mazda on aerodynamics and crash safety. According to a principal engineer at Mazda, AI helps identify which features affect aerodynamics and provides more room to explore shapes. It is the same democratization of assisted design seen in the rise of “vibe coding,” now applied to hardware.
Separating the Headline From the Reality
It is worth separating the headline from the substance. “Designing a device in less than a week” sounds as if AI now handles hardware engineering alone, and that is not what OpenAI itself describes. What AI accelerated was the design; manufacturing still required human engineers to deal with tolerances, assembly, and production—precisely where the physical world defies simulations. The real leap is not that a model can draw an enclosure, but that it can compress weeks of iteration into days. That is already significant and explains the 23-fold increase in internal use.
However, the narrative of full autonomy is, for now, more marketing than reality: between a prototype that powers on and a product manufactured at scale lies a gulf that AI cannot yet cross alone. The experiment shows genuine progress in design acceleration, but it also confirms that physical production remains a human-led discipline, at least for the near term.