OpenAI has demonstrated how its GPT-6 Astra model can contribute to designing physical hardware, completing the design of a working prototype device in less than a week. Two OpenAI team members used the AI model to develop a small device that allows users to communicate with the company’s Codex coding agent through a dedicated physical interface.
The experiment covered multiple engineering disciplines, including mechanical design, electronics, printed circuit board (PCB) development, firmware and software integration. OpenAI says Astra handled the design work without direct human edits, although the prototype still required several rounds of review and adjustments before it could be manufactured.
What Hardware Device Did GPT-6 Astra Design?
The prototype was developed as a side project by two OpenAI team members. The device, described in coverage as a Codex companion or Codex pet, is designed to provide a dedicated physical interface for interacting with Codex.
Instead of opening a laptop and accessing the coding agent through a conventional interface, users can communicate with Codex through the small hardware device.
The prototype includes a physical push-to-talk button and a small display, allowing users to interact with their coding agents.
OpenAI used the experiment to test whether its latest AI models could handle a complete hardware development workflow, from initial requirements to a physical prototype.
According to the company, GPT-6 helped establish the functional requirements, while Astra contributed to the device’s appearance, internal components, electronics and software.
The project was completed in under a week, while a contract manufacturer produced a small batch of prototypes in less than a month.
How GPT-6 Astra Designed the Hardware From Scratch
The experiment involved several stages of hardware development that traditionally require different engineering specialists.
OpenAI says Astra contributed across four major areas.
1. Mechanical Design and Enclosure Development
A hardware device requires more than electronic components. Engineers must also design an enclosure that accommodates the internal hardware while maintaining appropriate dimensions, structural integrity and usability.
Astra developed the device’s external appearance and designed the internal structures required to hold its components in place.
The design drew inspiration from OpenAI’s recent collaboration with Work Louder, as well as references from games and toys.
2. Electrical Design and PCB Development
One of the more technically demanding parts involved designing the printed circuit board.
A PCB connects the electronic components inside a device and provides the electrical pathways required for communication between them.
Astra selected the electrical components and designed the PCB, including component placement and electrical connections.
The model also had to ensure that the board fitted within the previously designed enclosure.
3. Firmware Development
Hardware components require firmware to control their operations.
Astra generated the firmware needed to operate the prototype and coordinate its hardware functions.
Firmware acts as the software layer between the physical components and the applications running on a device.
4. UI/UX and Software Integration
The final stage involved creating the user interface and connecting the device to Codex.
Astra generated the software integrations required to establish communication between the hardware and OpenAI’s coding agent.
This allowed the team to build a dedicated physical interface for interacting with an AI coding assistant.
Why the Prototype Still Needed Human Engineering Reviews
Although OpenAI completed the design work in under a week, the company acknowledged that the process was not entirely ready for direct manufacturing.
The contract manufacturer required several rounds of review and iteration before producing the prototypes.
OpenAI identified specific limitations in Astra’s designs, particularly its ability to account for manufacturing constraints.
These included:
- Production limitations at manufacturing facilities.
- Component and assembly tolerances.
- Details required to assemble the device correctly.
- Practical considerations that emerge when moving from digital designs to physical products.
This distinction is important because designing a device digitally and preparing it for mass production are two different engineering challenges.
A design may look correct in computer-aided design software but still contain problems when components are manufactured and assembled.
OpenAI says these limitations provide useful feedback for improving future models.
OpenAI’s Hardware Teams Are Using AI More Frequently
The Codex companion experiment is part of a broader effort to integrate AI into professional engineering workflows.
OpenAI says its hardware and infrastructure teams already use ChatGPT and Codex for several engineering activities, including:
- Creating and revising computer-aided design (CAD) models.
- Preparing manufacturing drawings.
- Setting up engineering simulations.
- Reviewing mechanical and electrical designs.
- Supporting robotics and computing infrastructure projects.
The company reported that model usage across its hardware teams, measured in tokens, increased 23-fold between February and October 2026.
This suggests that AI tools are becoming more deeply integrated into internal engineering operations.
However, a 23-fold increase in token consumption does not necessarily mean engineering productivity or completed hardware output increased by the same amount.
The company views these developments as early evidence that AI can help engineers test ideas, explore alternatives and accelerate design iterations.
OpenAI Tests AI Engineering Capabilities With Mazda and Chip Ganassi Racing
OpenAI is also expanding its work with external industry partners to understand how AI can contribute to real-world engineering.
Two notable collaborations involve Mazda and Chip Ganassi Racing (CGR).
Collaboration With Mazda
OpenAI has announced a research collaboration with Mazda’s aerodynamics and crash-safety engineers.
The teams are exploring how AI models can help develop and evaluate vehicle designs through engineering simulations.
One area of focus is vehicle aerodynamics.
GPT-6 Astra is being used to explore different vehicle nose shapes, simulate airflow and refine promising designs to investigate ways of reducing aerodynamic drag.
The work also considers structural performance, styling and manufacturing costs.
The objective is to help engineers evaluate more design possibilities while reducing the time required to explore alternatives.
Partnership With Chip Ganassi Racing
OpenAI is expanding its collaboration with Chip Ganassi Racing to investigate AI-assisted engineering and prototyping.
One application involves converting three-dimensional scans of existing physical components into editable CAD models and two-dimensional manufacturing drawings.
This is particularly useful when original engineering design files are unavailable.
The AI-generated designs and drawings are reviewed by racing engineers to check their accuracy and manufacturability.
The collaboration demonstrates how AI could help engineering teams recreate, modify and improve existing hardware components.
OpenAI Measures Progress With the BenchCAD 2.0 Benchmark
Alongside its hardware experiments, OpenAI is evaluating how effectively its models handle engineering-specific tasks.
The company highlighted BenchCAD 2.0, a benchmark designed to test AI capabilities in areas such as reconstructing three-dimensional parts and assemblies and identifying connections on circuit boards.
OpenAI reported the following results:
| Model | BenchCAD 2.0 Average Score |
|---|---|
| GPT-6 Astra | 0.594 |
| Newer internal OpenAI model | 0.875 |
The newer internal model achieved a substantially higher score than GPT-6 Astra at its highest reasoning setting.
However, OpenAI has not disclosed the identity or technical details of this newer model.
The benchmark also covers only selected engineering capabilities. It does not establish whether an AI system can independently manage the entire engineering lifecycle, including manufacturing, assembly, safety validation and quality control.
The Bigger Picture
OpenAI’s hardware experiment highlights a significant expansion in the potential applications of generative AI. AI models are moving beyond writing software, generating images and analysing information towards contributing to the design of physical products.
The ability to coordinate mechanical engineering, electrical design, firmware and software development within one workflow could change how hardware prototypes are developed. Smaller engineering teams may eventually be able to explore more ideas without requiring separate specialists for every early-stage design task.
However, the manufacturing limitations identified by OpenAI demonstrate that human engineering expertise remains essential. Physical products must satisfy real-world requirements involving materials, tolerances, safety and assembly. AI can accelerate the design process, but producing reliable hardware requires much more than generating technically convincing files.
Looking Ahead
OpenAI plans to continue improving its models’ engineering capabilities through internal hardware projects and collaborations with companies such as Mazda and Chip Ganassi Racing. These partnerships will help evaluate how AI performs on practical mechanical, electrical and manufacturing challenges.
The longer-term opportunity is to develop AI systems capable of supporting more of the product development lifecycle, from initial concepts and simulations to engineering drawings and manufacturing preparation. If models become more reliable at handling physical constraints, AI-assisted hardware development could become an important part of industries ranging from automotive engineering to robotics and consumer electronics.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
