Two People Build a Hardware Prototype in a Week; OpenAI Admits Mass Production Is Still Held Up by Tolerances

On November 3, OpenAI published a set of hardware engineering progress updates. ChatGPT and Codex no longer only write code; they are beginning to participate in mechanical design, electrical engineering, and simulation. In the promotional video, a hand holds a purple device, the screen displays a cartoon character, and the image reads "Hardware is the new software." OpenAI says two employees used GPT-6 Astra to make a device prototype that can converse with Codex, in less than a week, and did it as a side project.

This purple device was not just an exterior drawing. According to OpenAI's description, Astra designed the enclosure, internal structure, circuit boards, component selection, and connections, and also did the interface, firmware, and software integration. A contract manufacturer produced a small batch of prototypes within a month. The problems are also stated directly: to reach mass production, multiple rounds of review and iteration are still needed, and Astra did not consistently account for manufacturers' production constraints, tolerances, and assembly details. These gaps were included in the list of things future models need to improve.

OpenAI's internal hardware team is also using Codex and ChatGPT to modify CAD, produce manufacturing drawings, set up simulations, and conduct design reviews. OpenAI says that from February to now, the hardware team's model usage, measured in tokens, has grown 23 times. In external collaborations, Chip Ganassi Racing's R&D team is trying to convert 3D scans of existing parts into editable CAD and 2D manufacturing drawings, which is useful when original design files are missing. Mazda is using the models for aerodynamics and crash-safety simulation to explore new vehicle styling. Mazda's Akihiro Nakata mentioned that designers often ask, "Does it have to look like that?" Simulating dozens of variations can help them determine which features really affect airflow.

OpenAI also released a test set, BenchCAD 2.0, to test models on reconstructing 3D parts and assemblies and identifying circuit board connections. An updated internal model scored an average of 0.875, while GPT-6 Astra scored 0.594 at the highest reasoning setting. OpenAI itself says the engineering lifecycle is more than design, and these scores are only early evidence.

The reply section is closer to real-world deployment than the promotional video. Some users asked what Codex actually produces in mechanical design: CAD scripts, parametric models, or reviewing files that engineers have already made. Others asked whether it can directly read CAD files, or can only go through scripts such as CadQuery and OpenSCAD. A user doing mechanical engineering asked about tolerance stacks. Someone also left a comment: iteration speed is easy; sign-off is slow. At the end, OpenAI is recruiting engineers, engineering firms, and software partners to participate in testing, with the link https://openai.com/index/accelerating-engineering-for-the-physical-world/. The demo shortens design iteration, but molds, suppliers, reviews, and sign-off have not gotten faster along with it. Hardware is not like changing code: changing one line does not require reopening a mold.