OpenAI says its GPT-6 Astra model designed a physical gadget, from the casing to the circuit board to the firmware, with no direct human edits, and the company admits the result was not ready for a factory. The account comes from a research post OpenAI published on 3 November 2026 about its work on mechanical and electrical engineering.
The test product is a small desk device for talking to Codex, OpenAI’s coding agent, part of what the company calls its TamaGPT experiments. OpenAI says Astra defined the requirements, shaped the look (drawing on a recent Work Louder collaboration and on games and toys), designed the enclosure and the structures that hold parts in place, chose the electronic components, laid out the board, and wrote the interface and the code that links the gadget to Codex.
According to OpenAI, two employees finished the design with Astra inside a week, working on it part-time. A contract manufacturer then produced a handful of prototypes within a month. Speed is the headline claim, but the same post says preparing the design for mass production took multiple review passes. Astra did not always account for what the factory could build, including tolerances and assembly details. OpenAI says those gaps show what future models need to fix.
Inside the company, the hardware and infrastructure teams, covering robotics and compute infrastructure, rely on Codex and ChatGPT for CAD models, manufacturing drawings, simulation setup, and design review. OpenAI reports that their token consumption has grown 23-fold since February. Tokens count volume, not results, and the post gives no figure for parts shipped or hours saved. Grace Lam, a Forward Deployed Engineer at OpenAI, is quoted saying “Codex has dramatically accelerated how we move from an idea to an engineered part.”
The capability numbers come from a benchmark. BenchCAD 2.0 checks whether a model can rebuild 3D parts and assemblies and spot the connections on a circuit board. OpenAI says a newer internal model averaged 0.875, while Astra, set to think as hard as it can, scored 0.594. The test and both models are OpenAI’s, and the post does not name the newer one. OpenAI itself notes that design is only one slice of the engineering lifecycle.
Outside testing is thinner. OpenAI is widening its work with Chip Ganassi Racing, where engineers are trying to convert 3D scans of real parts into editable CAD files and 2D drawings, handy when the original files no longer exist. The team’s engineers check the output for accuracy and manufacturability. R&D Engineer Matthew Ortman says “I have been working with Codex directly in CAD,” and describes a process that is faster and more hands-off for him.
A new research tie-up with Mazda pairs Astra with the carmaker’s airflow specialists and its crash-safety team. They are running airflow and structural simulations, and Astra is modeling new nose shapes, simulating the air moving over them, and refining the best candidates to cut drag. Principal Engineer Akihiro Nakata says designers often ask “Does it have to look like that?” whenever engineers land on a shape that helps airflow. OpenAI promises results in the coming months and gives no drag figures yet.
For now, the evidence is one gadget, one in-house benchmark, and two partners who have not published numbers. Anyone building hardware should watch the gap Astra left open, because a model that learns a factory’s tolerances is the difference between a fast prototype and a shippable product.
Reported by OpenAI in its research post “Accelerating engineering for the physical world”, published 3 November 2026.
