OpenAI details Codex hardware workflows, with manufacturing review still required · 2026-11-04
November 3: Codex use extends into hardware engineering workflows
Reported by OpenAI: Its hardware and infrastructure teams use Codex and ChatGPT to create and revise CAD models, prepare manufacturing drawings, set up simulations and review designs across robotics and compute infrastructure. Model usage by its hardware teams, measured in tokens, has grown 23-fold since February.
The same article describes an experiment using GPT-6 Astra to design a test device, covering its enclosure, PCB, firmware and software integrations for connecting to Codex. OpenAI says two team members designed it in under a week as a side project; a contract manufacturer then built a small prototype batch in under a month. Those are separate design and manufacturing intervals, not evidence of a deliverable product completed within a week.
Our assessment: For hardware engineers and developers building agents for engineering tools, the useful development is the breadth of work across mechanical, electrical and software artifacts. The 23-fold figure measures usage, however, and cannot establish gains in productivity, quality or cost.
Manufacturing constraints remain a boundary; industry work is still being evaluated
Reported by OpenAI: Preparing the device for manufacturing still required several rounds of review and iteration. Astra did not consistently account for the manufacturer’s production limits, tolerances or assembly details. Although the experiment explored creating a device without direct human edits, the account does not establish a manufacturing process free of human review.
OpenAI also describes an expanded collaboration with Chip Ganassi Racing to test converting 3D part scans into editable CAD models and 2D manufacturing drawings, with the partner’s engineers reviewing accuracy and manufacturability. It separately announces a research collaboration with Mazda on vehicle design and evaluation involving aerodynamics and crash safety.
Our assessment: The evidence supports a report of internal adoption and ongoing partner evaluation. It does not establish general access to a complete hardware-design workflow. GPT-6 Astra is the model used in the account, while Codex is the agent tool in the engineering workflow; this evidence is not a release announcement for model weights, a standalone harness or an inference server, nor does it establish production readiness.
Team decision: pilot reviewable tasks and retain manufacturing approval gates
Advice, not tested by kuoo.uk: Teams with CAD tooling, reference designs and engineering review capacity could consider a bounded assistance pilot on one part or drawing set. Teams without qualified reviewers or documented manufacturing constraints should defer automated generation of manufacturing deliverables.
Track three categories of results:
- Artifact consistency: Check dimensions, interfaces and agreement between models and drawings against the reference design, rather than merely confirming that files open.
- Manufacturing constraints: Have engineers verify tolerances, assembly requirements and supplier process limits, retaining human approval before submission for manufacturing.
- Actual workload: Record generation, review and rework time separately, then compare the total with the existing process. Token consumption and first-output speed should not substitute for a workload assessment.
The bounded decision is to expand only if review and rework remain acceptable and critical constraints can be checked consistently. Any reduction in manual modeling must be weighed against additional verification and integration work.
Evidence and limits
This report relies solely on the supplied excerpt of OpenAI’s official article, dated November 3, 2026; November 4 is the retrieval date. Usage, design timelines and collaboration progress are vendor-reported and not independently verified. We have not run the tools, reproduced the device design or validated manufacturing results. The material does not establish general workflow availability, exact tooling configurations or full costs.
AI-assisted drafting with a separate model review. This is source-based analysis, not independent performance testing.