AI is moving from documents into physical systems
OpenAI reported on November 3 that engineering teams are using AI across mechanical and electrical design, simulation, manufacturing documentation, and end-to-end device development. Examples in the report include converting scans into editable CAD, generating engineering drawings, and supporting aerodynamic and crash simulations.
The publication is a vendor account of its models and collaborators, so its claims should be evaluated as early operational evidence rather than a universal performance guarantee. The broader signal is clear: AI tools are entering workflows where an error can become scrap, delay, rework, or a physical safety issue.
A prototype reveals both speed and gaps
In one described exercise, a team used an internal model called Astra to help create an enclosure, circuit board, interface software, and a small production batch. OpenAI says the workflow reduced iteration time, but manufacturing still required several reviews and the model missed tolerances and assembly details.
That is the useful lesson. A design that looks correct on screen may fail because of stack-up, fastening, tooling access, material behavior, thermal conditions, regulatory constraints, or assembly sequence. AI can propose and transform work; it does not remove the need for engineering judgment.
Benchmarks are inputs, not release authority
OpenAI reports a score of 0.875 for a newer internal model on its BenchCAD 2.0 benchmark, compared with 0.594 for GPT-6 Astra. Those are vendor-reported results on a bounded benchmark, not evidence that either model can approve a complete product lifecycle.
Release decisions need several independent layers: requirements, calculations, simulation, design review, manufacturability review, security and safety analysis, prototype inspection, physical tests, configuration control, and accountable sign-off. The model’s output should remain traceable to the prompt, source data, tool version, reviewer, and resulting change.
Pilot where evidence is easy to measure
Start with a bounded task such as drawing cleanup, variant generation, component research, test-fixture design, or preparation of simulation inputs. Define the accepted source data, restricted design space, reviewer qualifications, and test plan before measuring speed.
Track engineering hours, defect escape, rework, manufacturability findings, test failures, and traceability—not only how quickly a first draft appears. A successful pilot gives engineers more time for high-value decisions while preserving the evidence and ownership required to release a physical product.