📅 November 3, 2026 • OpenAI / The Decoder
OpenAI says GPT-6 Astra designed a working hardware device from scratch in under a week. In a post titled “Accelerating engineering for the physical world”, the company describes an internal side project: two employees set out to build a small device for communicating with Codex, and the model carried the entire engineering path. Per OpenAI, Astra produced the industrial design, the enclosure, component selection, the printed circuit board layout, the firmware and the user interface. A contract manufacturer then turned the released files into a batch of working prototypes in under a month, so the project moved from idea to physical hardware in weeks rather than months. The device itself remains an internal experiment — but the prototype batch was actually built.
Model usage inside OpenAI's own hardware teams is scaling fast. The company reports a 23x increase since February, measured in tokens. On BenchCAD 2.0, a CAD engineering benchmark, a new internal OpenAI model scores 0.875 versus 0.594 for GPT-6 Astra at maximum reasoning. The gap is presented as the remaining distance on full engineering workflows — the places where a design must survive contact with real manufacturing constraints rather than a clean specification. Both numbers point in the same direction — engineering is becoming a model-native workflow inside OpenAI.
OpenAI is explicit about the limits of the result. Per the company, Astra did not always account for tolerances and assembly requirements, so the project required rounds of human review before the design went to the manufacturer. The outcome is framed as a working prototype pipeline with engineers in the loop — not an autonomous factory, and not a claim that the model replaces manufacturing engineers. The point of the case study, per OpenAI, is the speed of the loop, not the absence of people.
The physical-world push now has named partners. OpenAI announces an expanded collaboration with Chip Ganassi Racing, converting 3D scans of race car components into editable CAD, and a new research collaboration with Mazda, applying the models to aerodynamics and crash simulations. Per OpenAI, both are test collaborations on real engineering problems, with partner engineers reviewing model outputs for accuracy, rather than product integrations.
IMPACT: when one model can carry a product from concept to PCB, firmware and a manufactured batch, hardware iteration compresses from months toward weeks — with human review still mandatory at the tolerance and assembly stages where the model is weakest.
Sources: openai.com, the-decoder.com
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
Top comments (0)