OpenAI's Astra designs a Codex device with no human edits

OpenAI frames this as a shift from chat to physical product work: models preparing drawings, setting up simulations, and coordinating changes across mechanical, electrical and software disciplines. The internal evidence is a usage signal — tokens consumed by OpenAI's hardware teams grew 23-fold since February — plus the Astra experiment, where GPT-6 handled requirements, enclosure, PCB and firmware for a device for talking to Codex.

The Astra test is presented with its limits intact. A contract manufacturer produced a small prototype batch in under a month, but the device needed several rounds of review because the model missed production limits and tolerances. OpenAI says those gaps are informing what future models must improve, which reads as an admission that design fluency and manufacturability are not the same skill.

The external collaborators show where OpenAI wants validation. Chip Ganassi Racing is expanding its work to turn 3D scans of existing parts into editable CAD models and 2D manufacturing drawings when original design files are unavailable, with CGR engineers checking accuracy and manufacturability. Mazda is a new research collaboration, with aerodynamics and crash-safety engineers using airflow and structural simulations to explore vehicle designs; OpenAI says more results will come in the coming months.