OpenAI: GPT-6 Astra Designed a Gadget for Codex, Prototypes Assembled in Less Than a Month

OpenAI shared how its models assist engineers in designing, modeling, and manufacturing physical products. This includes mechanical engineering, electrical engineering, and end-to-end system design.

To test GPT-6 Astra in multimodal tasks and computer management, OpenAI attempted to create a gadget without direct human intervention. The model formulated functional requirements and developed a device for voice communication with Codex: an appearance inspired by collaboration with Work Louder, games, and toys; the body and mounts; selection of components and a printed circuit board to fit the body dimensions; interface, firmware, and integration with Codex.

What the Assembly Showed

Two employees developed the device with Astra in less than a week alongside their main work, and a contract manufacturer released a batch of prototypes in less than a month. Preparation for production required several rounds of checks and refinements: Astra did not always consider production constraints, tolerances, and details needed for assembly. According to OpenAI, these gaps indicate what future models need to learn.

Tokens and Benchmark

Hardware and infrastructure teams use Codex and ChatGPT for CAD models, drawings, simulations, and design verification. Since February, model consumption in terms of tokens has increased 23 times, and the median daily use of Codex by October reached 212 million tokens per employee.

"Codex has radically accelerated the path from idea to finished part. We iterate faster in CAD, turn 3D models into production-ready 2D drawings, and set up simulations and finite element calculations with much less manual work," stated OpenAI.

On the BenchCAD 2.0 benchmark, the new internal model scored an average of 0.875 points compared to 0.594 for GPT-6 Astra at the highest level of reasoning.

Partners: Racing and Automotive Industry

Chip Ganassi Racing is testing the conversion of 3D scans of parts into editable CAD models and 2D drawings – useful when original files are lost. "The model proved useful for transitioning from raw scan data to a solid model. I work with Codex directly in CAD. My comments are immediately applied by the model, manual edits can be made directly, and the process has become faster and requires less involvement," said Matthew Ortman, R&D engineer at Chip Ganassi Racing.

With Mazda, OpenAI is working on new car designs using simulations of airflow and structural behavior. "'Does it really have to look like this?' our designers often ask when we find a shape that improves streamlining. By exploring and calculating dozens of options, Astra in Codex helps us understand which elements are crucial for aerodynamics and where designers have freedom. Therefore, the value of Astra and collaboration with OpenAI goes beyond task automation and shortening development cycles: it helps us combine Mazda's signature design with aerodynamic performance without sacrificing either," stated Mazda.

OpenAI promised to showcase the results of its work with Mazda in the coming months. The company is seeking partners among engineers and software developers to help identify valuable use cases and influence the training of future models.