OpenAI seeks engineering partners after using Astra to prototype a device

OpenAI described early work applying AI to physical engineering and invited additional partners.

OpenAI asked engineers, engineering firms and software vendors on November 3 to work with it on applying frontier models to physical design, simulation and manufacturing. Its announcement describes an internal demonstration in which GPT-6 Astra helped design a device for talking to Codex, and an expanded collaboration with Chip Ganassi Racing to test ways of turning physical parts into editable design files. The announcement describes experiments and a call for partners. [OpenAI announcement](https://openai.com/index/accelerating-engineering-for-the-physical-world/)

The prototype exposed a manufacturing gap

OpenAI says two team members used Astra on a side project that produced an enclosure design, a printed circuit board, firmware and software integrations. A contract manufacturer built a small batch of prototypes. Before manufacturing, however, the design went through several rounds of review and iteration. OpenAI says the model did not consistently account for the manufacturer's tolerances, production limits and assembly details. [Prototype account](https://openai.com/index/accelerating-engineering-for-the-physical-world/)

That is a useful distinction for engineers evaluating the workflow. A drawing may capture the intended shape and still miss the information a shop needs to make and assemble the part. A trial can measure how much time it takes to produce a useful first pass, then separately measure the revisions required before an engineer will sign it off. A fast first pass alone does not establish an improvement across the production process.

From scans to editable designs

In the racing collaboration, the company says engineers are testing whether a model can turn a 3D scan into an editable CAD model and manufacturing drawings, particularly when the original design files are unavailable. The racing team's engineers review the results for accuracy and manufacturability. That gives the collaboration a specific job to measure and a human review point. [OpenAI's description of the collaboration](https://openai.com/index/accelerating-engineering-for-the-physical-world/)

OpenAI also reports that a newer internal model reached 0.875 on BenchCAD 2.0, versus 0.594 for GPT-6 Astra at its highest reasoning setting. The tasks include reconstructing parts and identifying circuit board connections. The benchmark provides one measure of progress on design tasks, while the prototype shows why manufacturing review remains a separate step. [BenchCAD results and scope](https://openai.com/index/accelerating-engineering-for-the-physical-world/)