OpenAI says its models are speeding up hardware design. The prototype still needed a person at the factory

OpenAI published the account on 3 November. The claim is that its models now help engineers design, simulate and build physical products, and that the company has early evidence of that in industry as well as inside its own hardware teams. Inside, it says Codex and ChatGPT are used for computer-aided design, manufacturing drawings, simulation and design review, on robotics and on computing infrastructure. The one hard internal number is usage. Token use across the hardware teams, measured by the company, is up 23-fold since February. A multiple of tokens is not a multiple of finished parts.

A drawing is not a tolerance

The worked example is a side project called TamaGPT, a test device meant to talk to Codex. OpenAI says two people on the team used GPT-6 Astra to design it in under a week, as a side project, covering the look, the mechanical design, the electrical design and the software. A contract manufacturer then built a small batch of prototypes in under a month. The next sentence is the one that decides whether the week counts. Getting the device ready for manufacturing still took several rounds of review. Astra, the company says, did not consistently account for the manufacturer’s production limits, the tolerances, or the details assembly needs. Those gaps, it says, are a list of what later models have to do better. A prototype that exists because a person corrected the model is not a machine that went from a prompt to a loading dock.

The company also published a benchmark. BenchCAD 2.0, it says, asks whether a model can reconstruct 3D parts and assemblies and can identify connections on a circuit board. A newer internal model scored 0.875 on average. GPT-6 Astra, at its highest reasoning setting, scored 0.594. A score on a reconstruction test is not a part that has been machined.

Two workshops, neither of them a production line

The outside tests are named, and they are still tests. Chip Ganassi Racing is expanding a collaboration in which the model turns 3D scans of existing parts into editable designs and into 2D drawings, for the case where the original files are missing. Matthew Ortman, an R&D engineer there, said the model has been useful for going from raw scan data to a solid body, and that his own feedback and his own manual tweaks are part of the loop. The engineers still review accuracy and whether a shop can make the part. A faster path from a scan to a file is not a car that has been built from the file.

Mazda is a new research collaboration. OpenAI’s forward-deployed and research teams are working with Mazda’s aerodynamics and crash-safety engineers on simulations of airflow and structure, with styling and cost still in the problem. Astra is being used to try nose shapes that cut drag. Akihiro Nakata, a principal engineer, said designers often ask whether a shape that helps the air has to look like that, and that trying dozens of variations shows which features matter and where the designers still have room. The company says more progress and results will come in the following months. A month of shapes on a screen is not a vehicle that has been stamped.

The internal quotes are the same shape. Grace Lam, a forward-deployed engineer, said Codex has sped the path from an idea to a part, including the drawings and the setup of a stress analysis. Qiong Wang, on the technical staff, said the tools are used for electrical design and for design review, and that they have the potential to grow into specialist teams that could take a board from a concept to manufacturing files. Potential is not the team. Jackie Lee, on the infrastructure staff, said building-design work is starting to be automated, and that architects and engineers still check the result. Starting is the right verb. None of the three said a production line has been handed to the model.