OpenAI says GPT-6 Astra designed a Codex hardware device, but manufacturing still took review rounds

OpenAI has published a post on how its models help with physical engineering: CAD, circuit boards, simulation and manufacturing drawings. The headline example is a small device for talking to Codex that OpenAI says GPT-6 Astra designed.

The device test

OpenAI says it wanted to see whether it could “create a device without any direct human edits.” By its account, Astra:

• developed the look of the device, drawing on OpenAI’s Work Louder collaboration and references from games and toys,
• designed the enclosure and the internal structures holding the parts,
• selected the electrical components and designed the printed circuit board to fit the enclosure,
• designed the interface and generated the firmware and software integrations that connect the device to Codex.

Two team members did it as a side project in under a week, and a contract manufacturer built a small batch of prototypes in under a month, OpenAI says.

The post is clear that “no direct human edits” did not mean no human loop. “Getting the device ready for manufacturing still required several rounds of review and iteration,” OpenAI writes. “Astra did not consistently account for the manufacturer’s production limits, tolerances, or details required for assembly.”

Internal use and a benchmark

OpenAI says model usage across its hardware teams, measured in tokens, has grown 23-fold since February, for work such as revising CAD models, preparing manufacturing drawings and setting up simulations.

It also reports a benchmark, BenchCAD 2.0, which “tests whether models can reconstruct 3D parts and assemblies and identify connections in circuit boards.” A newer internal model averaged 0.875, against 0.594 for GPT-6 Astra at its highest reasoning setting. OpenAI does not name the newer model or say when it ships, and the post does not link the benchmark.

Outside collaborators

OpenAI says it is expanding work with Chip Ganassi Racing on turning 3D scans of existing parts into editable CAD models and 2D drawings, with the team’s engineers reviewing output for accuracy and manufacturability. It is also announcing a research collaboration with Mazda’s aerodynamics and crash-safety engineers, using airflow and structural simulations to explore vehicle designs. OpenAI says it will share Mazda results “in the coming months,” and it is asking engineering firms and software vendors to apply as collaborators.

Why this matters

OpenAI’s own account puts the gap where builders will feel it: the model can draft enclosure, board and firmware, but it missed production limits and tolerances, so plan for human design review before anything goes to a manufacturer. The 0.875 BenchCAD 2.0 score is for a model you cannot use yet; the shipping comparison is Astra’s 0.594. Do not line that up against the 95.9% for Astra on BenchCAD with tools in Moonshot’s Kimi K3.1 Pro figures, which is not labeled 2.0 and was run with tools, so the two are not comparable. Like the self-published DrivingBench Corolla test, this is a single-party result until others can run it.