OpenAI models design full physical devices, Mazda research tie-up announced

The latest:

A test device was designed end to end by an AI model — requirements, enclosure, printed circuit board, component selection, firmware and user interface — with two OpenAI staff completing the work in under a week, the company said on November 3, 2026. A contract manufacturer produced a small prototype batch in under a month. Several review rounds were still needed before the device was manufacturable.

Details:

• The benchmark: On BenchCAD 2.0, which measures reconstruction of 3D parts and assemblies and the identification of connections in printed circuit boards, a newer internal model averaged 0.875, according to OpenAI. GPT-6 Astra scored 0.594 at its highest reasoning setting, a gap of roughly 47% on the same test.
• The device: OpenAI said it tested GPT-6 Astra by having it design a device for talking to Codex without direct human edits. The model defined the requirements itself, then handled the enclosure, the PCB, component choices, firmware and the interface layer.
• The limits: Preparing the device for manufacturing took multiple rounds of review, OpenAI said, because the model did not consistently account for the contract manufacturer’s production limits, tolerances and assembly details. Human review remained the condition for moving from a design file to a built unit.
• Internal usage: OpenAI’s hardware and infrastructure teams are using Codex and ChatGPT to create and edit CAD models, prepare manufacturing drawings and set up simulations. The company said those teams’ model usage, measured in tokens, has grown 23-fold since February.
• Motorsport work: The company expanded its collaboration with Chip Ganassi Racing, converting 3D scans of existing parts into editable CAD models and 2D manufacturing drawings. The step targets reverse-engineering workflows, where physical components already exist but usable digital files do not.
• The Mazda deal: OpenAI announced a new research collaboration with Mazda to help aerodynamics and crash-safety engineers develop and evaluate vehicle designs. The company did not disclose the duration of the collaboration, its financial terms, or which vehicle programs it covers.
• The scope: The announced advances span mechanical engineering, electrical engineering and end-to-end systems design, according to OpenAI. The company framed the work as assisting engineers in designing, simulating and building physical products rather than replacing the engineering function.
• The invitation: OpenAI called on engineers, engineering firms and software vendors to work with it on the effort. No application deadline, selection criteria or number of partner slots was announced alongside the call.

Background:

BenchCAD 2.0 is an internal-facing engineering evaluation measuring how well a model rebuilds three-dimensional parts and assemblies and maps connections inside printed circuit boards — tasks closer to factory-floor output than to text generation.

Between the lines:

The two headline numbers point in different directions. The 0.875 benchmark score and the one-week design cycle argue that capability is arriving fast; the multiple review rounds over production limits, tolerances and assembly details mark where it stops. The 23-fold token growth since February suggests OpenAI is testing the tooling on its own hardware teams first, before the Mazda and Ganassi work reaches outside production lines.

What’s next

Watch for results from the Mazda aerodynamics and crash-safety collaboration, any published BenchCAD 2.0 figures from outside parties, and whether the Chip Ganassi Racing scan-to-CAD workflow reaches parts that are actually raced.