HYPE OR FACT? · CHECKED 4 NOVEMBER 2026
OpenAI: two engineers had GPT-6 Astra design a working Codex device in under a week
OUR VERDICT
Partly factOur original verdict: “Fact or hype: partly fact. It's a real, detailed case study, but every number is OpenAI's own, the stronger internal model isn't available, and 'no direct human edits' stopped at the prototype: getting it ready for manufacturing took several rounds of human review.”
THE CLAIM
“Two OpenAI team members used GPT-6 Astra to design a Codex voice device, enclosure, circuit board and firmware, in under a week as a side project, without any direct human edits; OpenAI's internal model is far ahead of Astra on hardware and CAD work.”
What’s actually true
- OpenAI's post (3 Nov) says two team members used GPT-6 Astra to write the requirements and design the enclosure, the printed circuit board and the firmware for a small device for talking to Codex, as a side project in under a week, 'without any direct human edits'.
- A contract manufacturer built prototypes in under a month. OpenAI says manufacturing readiness still needed several rounds of human review, because Astra didn't consistently account for tolerances and production limits.
- OpenAI says its hardware teams' token use has grown 23-fold since February, and names Chip Ganassi Racing (scan-to-CAD) and Mazda (aerodynamics and crash work) as collaborators.
- The chart that went viral shows an unreleased internal model scoring 0.875 on OpenAI's BenchCAD 2.0 against 0.594 for GPT-6 Astra. That's OpenAI's own benchmark and an unreleased model.
- No independent test of the device or the benchmark has been published.
Sources
Cost
GPT-6 Astra is available in ChatGPT and Codex on paid plans; the stronger internal model isn't available. A real product still needs engineering review and manufacturing costs.
USE IT IF
Product and hardware teams who want help with requirements, CAD clean-up, circuit and firmware drafts and design reviews.
SKIP IT IF
Anyone expecting a factory-ready design with no engineer checking tolerances and manufacturability.
How we knowChecked on 4 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗