From design to the factory floor, AI is confronted with the constraints of physical engineering.

A device designed with GPT-6 Astra, reconstructed auto parts, and simulations at Mazda illustrate AI advances in engineering.

A printed enclosure, an electronic board, a screen, and software capable of communicating with Codex. Two OpenAI collaborators managed to design this prototype in less than a week with GPT-6 Astra, without directly modifying the files generated by the model. Manufacturing turned out to be more complicated: several verification cycles were needed to adapt the plans to industrial constraints. This experiment illustrates both the progress of artificial intelligence in the design of physical objects and the distance that still separates a digital project from a manufacturable product.

In a publication devoted to engineering for the physical world, the company details several experiments combining ChatGPT and Codex with computer-aided design (CAD), simulation, and electronics development software. From internal applications to collaborations with Chip Ganassi Racing and Mazda, the work focuses on coordinating different disciplines, from part geometry to mechanical properties and production constraints. A prototype designed in one week, manufactured in one month

The experimental device, nicknamed TamaGPT, was conceived as a physical interface for talking to Codex. Its aesthetics borrow from game consoles and electronic toys, with references drawn from the collaboration between OpenAI and Work Louder. The exercise consisted in entrusting the model with several steps usually split among industrial designers, electronics engineers, and developers.

Based on the functional requirements, GPT-6 Astra developed the outer shell and the internal structures intended to hold the components in place. It selected the electronic parts, organized their layout on a printed circuit board, and defined the necessary connections. The work also included the user interface, the firmware, and the software integrations ensuring communication with Codex.

A contract manufacturer then produced a small series of prototypes in less than a month. The move to manufacturing nevertheless revealed several shortcomings in the generated files. Certain limitations specific to industrial processes, dimensional tolerances, and assembly details had not been properly anticipated, requiring new iterations with the engineers.

The experiment demonstrates the model’s ability to coordinate tasks spanning multiple specialties, without establishing that it can independently design and industrialize a device on its own. The distinction between the absence of direct human modification during design and the absence of supervision over the entire project remains essential. Usage multiplied by 23 in hardware teams

The experiment is part of a growing use of AI within OpenAI’s hardware and infrastructure teams. Since February 2026, the volume of tokens consumed by these teams has increased 23-fold, according to the company’s internal data.

The applications involve creating and modifying CAD models, converting 3D geometries into manufacturing plans, preparing mechanical simulations, and verifying electronic designs. In infrastructure projects, the tools are also used to compare specifications, identify inconsistencies between documents, and organize the necessary corrections.

This increase measures a rise in model usage, not an equivalent increase in productivity. No consolidated figure for hours of work saved, manufacturing costs, or error reduction accompanies the data shared.

The company is also tracking its models’ capabilities with BenchCAD 2.0, an internal evaluation that includes reconstructing 3D parts and assemblies, as well as identifying connections on printed circuit boards. A newer internal model achieves an average score of 0.875, compared with 0.594 for GPT-6 Astra at its highest reasoning level.

The gap represents 0.281 point on the scale used. However, the full protocol and reproduction conditions for BenchCAD 2.0 are not provided in the publication, which prevents directly inferring a performance gain in a real industrial environment. The internal model evaluated is not identified, and no availability date is announced. Rebuilding automotive parts without their original plans

The work carried out with Chip Ganassi Racing addresses a different problem: digitally reconstructing an existing part when the design files are no longer available. The American motorsport team is experimenting with procedures capable of turning 3D scans into editable CAD models, then into technical drawings usable for manufacturing.

A scan provides a representation of an object’s surface, but not necessarily the information needed to modify or produce it. Engineers generally have to reconstruct the volumes, functional dimensions, and geometric relationships between the different parts. Codex helps with this conversion, with the option of applying corrections directly in the CAD environment.

Matthew Ortman, an R&D engineer at Chip Ganassi Racing, says this workflow reduces the manual handling needed to move from raw data to a usable solid model. The geometries and drawings are still reviewed by engineers, especially to verify their accuracy and compatibility with manufacturing processes.

This use therefore does not involve freely generating automotive parts from a text prompt. It relies on existing physical objects, geometric measurements, and corrections made within a defined technical framework. At Mazda, aerodynamics talks with design

A new research collaboration with Mazda extends these experiments to automotive design. The Japanese manufacturer’s aerodynamics and structural safety teams are working with OpenAI researchers and engineers to explore vehicle modifications based on digital simulations.

One exercise consists in changing the shape of the front end of a car to reduce air resistance. GPT-6 Astra, used in Codex, helps build geometric variants, prepare flow simulations, and compare the results. The work also takes aesthetic and economic constraints into account, in order to examine the trade-offs between performance and design.

For Akihiro Nakata, chief engineer