ResearchOpenAI

OpenAI: ChatGPT and Codex help design physical products

Original title: Accelerating engineering for the physical world

Plain language

What it is: OpenAI shared how its AI models, including ChatGPT and Codex, help engineers design, simulate, and build physical products. The models support tasks in mechanical engineering, electrical engineering, and system design.
How it works: The models help by preparing drawings, setting up simulations, and coordinating design changes. For example, they can help design physical enclosures, select electrical components, and generate software to connect devices.
Where it is available: These capabilities are currently used by OpenAI's internal teams and a small group of industry collaborators, such as Mazda and Chip Ganassi Racing. OpenAI is looking for more engineering firms and software vendors to collaborate with.
Example: Mazda uses the Astra model to simulate airflow and model new nose shapes to reduce vehicle drag.

What it is: OpenAI announced progress in training its frontier models, including GPT-6 and Astra, to assist with physical engineering workflows like CAD, PCB design, and simulation setup. They also introduced BenchCAD 2.0 to measure model performance in reconstructing 3D parts and identifying circuit board connections.
How it works: Models process multimodal inputs and perform computer-use workflows to generate CAD models, convert 3D designs to 2D production drawings, and write firmware. Engineers can interact with Codex directly inside CAD tools to apply feedback and automate repetitive modeling tasks.
Where it is available: The tools are used internally by OpenAI's hardware teams and externally by select partners in private collaborations. Developers and engineering firms can apply to collaborate and integrate these models with their software tools.
Example: Chip Ganassi Racing uses Codex to convert raw 3D scan data of physical parts into editable CAD models and 2D manufacturing drawings.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

Accelerating engineering for the physical world