Michael JohnsonCan AI be tasked with designing an entire device, rather than just a drawing—from its enclosure and circuit board to its firmware? On November 3, 2026, OpenAI described an experiment in which the Astra model, together with GPT‑6, helped design a device for communicating with Codex. The key takeaway for engineers is that AI is useful here for connecting several design stages and speeding up iterations; moving from a digital model to production-ready hardware still requires checking design and manufacturing details.
In the project as described, the system started with functional requirements. Astra then developed the device’s appearance, enclosure, and internal structures, selected electronic components, designed the circuit board and its component placement and connections, and prepared the interface, firmware, and software integrations. This is an interconnected workflow: changing the enclosure’s dimensions can affect board placement, while the components and their connections constrain the design options.
What changed in the model’s work
A conventional image generator answers the question, “What could this look like?” The scenario OpenAI described aims for a tighter integration of representations: requirements, mechanical geometry, electrical circuitry, and software are handled as part of one task. What matters to engineers is the transition between them—for example, ensuring that the board physically fits inside the enclosure and that the selected components support the device’s intended function.
OpenAI said that two employees built a test device in less than a week as a side project, and a contract manufacturer produced a small batch of prototypes in less than a month. Those timelines apply to this specific internal experiment and prototype. They show that it is possible to move quickly from concept to physical sample, but they do not provide a general estimate of design speed for other devices or teams.
The example is useful because AI contributed to more than the creation of a single artifact. When a model helps coordinate the enclosure, board, and software, engineers can spot interdependencies earlier—ones that would typically emerge when results are handed off between different tools and specialists. This kind of integration is especially important for small teams, where one person coordinates several disciplines.
Why a digital design is not yet production-ready
A model may propose a design that looks consistent on screen but fails to account for real manufacturing constraints. In this experiment, OpenAI specifically noted that Astra accounted inconsistently for the manufacturer’s constraints, tolerances, and assembly requirements. As a result, preparing the device for production required several rounds of review and revision.
The reason for this gap is practical. A digital design specifies the shape and relative positions of parts; manufacturing adds concrete conditions, including acceptable dimensional deviations, processing methods, and assembly requirements. If an enclosure leaves too little room or a design fails to account for how a component is installed, the file may be technically complete, but the product could be difficult or impossible to make using the specified process.
This points to a practical division of responsibility: the model helps prepare and coordinate options, while specialists check them against the manufacturer’s equipment, materials, and processes. For a team, that means judging the result not by how appealing the 3D model looks, but by how many significant engineering checks and corrections remain before a working sample can be made.
Where real-world applications are taking shape
OpenAI’s internal hardware teams use Codex and ChatGPT to create and modify CAD models, prepare manufacturing drawings, configure simulations, and analyze designs. The company said that since February, model usage by its hardware teams, measured in tokens, had increased 23-fold. This metric reflects model consumption within those teams, not reduced design time or costs.
In collaboration with Chip Ganassi Racing, the team is exploring how to turn 3D scans of existing parts into editable CAD models and manufacturing drawings. This scenario is valuable when the original design files are unavailable: the scan provides a starting point for reconstructing the geometry, while the partner’s engineers check the accuracy and manufacturability of the drawings.
In joint work with Mazda, the models are being applied to vehicle design: exploring front-end shapes, simulating airflow, and comparing options based on performance, design, and cost. OpenAI describes this as a research collaboration whose results the company plans to report later. The idea is to explore more shape options and connect aerodynamic simulation with design decisions more quickly; the published materials do not yet quantify improvements to a finished vehicle.
How to evaluate these systems in an engineering team
OpenAI cites BenchCAD 2.0 as one evaluation of models on tasks combining programming and engineering knowledge. The test assesses reconstruction of 3D parts and assemblies, as well as recognition of connections on circuit boards. In this evaluation, a new internal model scored an average of 0.875, while GPT‑6 Astra in maximum-reasoning mode scored 0.594.
For practical selection, this result is a useful signal of progress on a specific class of tasks: building geometry and recognizing relationships. It does not measure the full development cycle, which includes material selection, thermal and mechanical testing, certification, and manufacturability validation. Teams should therefore test models on their own typical tasks and record which output files engineering tools can actually accept and which still need manual rework.
A practical pilot can focus on one recurring stage, such as converting part scans into CAD, preparing drawings, or creating an initial simulation configuration. Teams should define acceptance criteria in advance, assign an engineer to check the geometry and constraints, and log the corrections needed before handing work to the manufacturer. This will show where the model reduces routine workload and where its suggestions create additional review cycles.
The key distinction is between design and validation
OpenAI’s project shows that AI can be applied to connected stages of physical design, not just text, code, or a standalone image. The benefits emerge when a system helps quickly coordinate requirements and options across mechanical engineering, electronics, and software. But manufacturability remains a separate engineering task, determined by tolerances, assembly, and the capabilities of a specific process.
For hardware developers, this offers a clear starting point: evaluate a limited, verifiable stage and measure not the number of models created, but the amount of manual correction needed before engineering approval. This approach makes it possible to benefit from rapid option generation while maintaining oversight where a digital design becomes a physical product.
