OpenAI said Tuesday that GPT-6 Astra designed a small working device from requirements through hardware and software without direct human edits.
The big picture: OpenAI is pushing its models from coding into physical engineering: CAD, printed circuit boards, simulation, manufacturing drawings and cross-discipline design reviews. The company is also starting a research collaboration with Mazda on aerodynamics and crash safety.
- Two staff members designed the test device with Astra in under a week as a side project, and a contract manufacturer built a small batch of prototypes in under a month.
- Model usage across OpenAI's hardware teams has grown 23-fold since February, measured in tokens.
Zoom in: Astra selected components, laid out the circuit board, designed the enclosure and internal structures, created the interface and generated firmware and software that connected the device to Codex.
Yes, but: Manufacturing still required several rounds of human review. OpenAI said Astra sometimes missed production limits, tolerances and assembly details, the kinds of errors that separate a plausible design from a reliable product.
By the numbers: A newer internal model scored 0.875 on BenchCAD 2.0, which tests reconstruction of 3D parts and circuit-board connections, compared with 0.594 for GPT-6 Astra at its highest reasoning setting.
What's next: OpenAI and Mazda will test nose shapes, airflow simulations and tradeoffs among aerodynamics, styling and cost. OpenAI says it will publish more results in the coming months and is recruiting additional engineering collaborators.
Why it matters: AI that can move from an idea to manufacturable hardware compresses a much larger part of product development than code generation alone. The remaining tolerance and review failures show where human engineers still hold the line.
