OpenAI’s Astra CAD Work Still Needs Manufacturing Review

According to JetsonsAge’s analysis, OpenAI on November 3, 2026 announced a research collaboration with Mazda and expanded its work with U.S. racing team Chip Ganassi Racing. The projects give its models distinct assignments: exploring vehicle shapes through simulation for Mazda and converting scans of existing racing parts into editable CAD models and manufacturing drawings for Chip Ganassi Racing. Both are research workflows in which partner engineers review the outputs.

An internal Astra-assisted device shows why that review matters. Two OpenAI team members designed a prototype for talking to Codex in under a week, and a contract manufacturer built a small batch in under a month. Preparing the device for manufacturing still took several rounds of review and revision because the model missed production limits, dimensional tolerances and assembly details. For engineering teams, the demonstrated capability is assistance with design work that still needs expert inspection before manufacture.

What the Astra device demonstrated

The device exercise covered more than a single CAD shape. GPT-6 helped define functional requirements, while Astra designed the enclosure and its internal supports, selected and placed printed circuit board components, and generated firmware and software integrations for Codex. The team set out to create the design without direct human edits. That constraint tested the model’s ability to coordinate mechanical, electrical and software work; it did not remove people from the manufacturing handoff.

The small prototype batch establishes that a physical device was built. It also exposes a specific boundary: geometry that fits together in a digital design can leave a manufacturer with unresolved questions about permitted variation, production methods or the order in which components must be assembled. The under-a-week figure covers the design exercise, while the under-a-month figure covers the prototype batch. Neither is a measured time for a validated production run.

BenchCAD measures design tasks, not factory readiness

In OpenAI’s engineering account, engineer Grace Lam describes simulation setup as requiring “far less manual work”; the same account gives a BenchCAD 2.0 average of 0.875 for a newer internal model and 0.594 for GPT-6 Astra at its highest reasoning setting. The higher score belongs to the newer internal model. BenchCAD 2.0 tests reconstruction of three-dimensional parts and assemblies and identification of circuit-board connections.

OpenAI’s hardware and infrastructure teams also use Codex and ChatGPT to create and revise CAD models, prepare manufacturing drawings, set up simulations and review designs. Those activities show where the tools already enter an engineering workflow. The benchmark measures selected design tasks, while the device exercise follows a design into physical prototyping. A score on reconstruction and connections cannot establish how often a supplier accepts a design, how much correction work remains or whether a simulation predicts a finished part’s behavior.

Mazda and Chip Ganassi Racing test different handoffs

Chip Ganassi Racing’s workflow turns a scan of an existing part into editable CAD geometry for engineering review.

Chip Ganassi Racing’s research team is testing a route from three-dimensional scans of existing parts to editable CAD models and two-dimensional manufacturing drawings. A scan captures the surface of a part; an editable solid gives engineers geometry they can change, while a drawing conveys requirements for making it. The route is especially relevant when original design files are unavailable. Racing engineers review the resulting models and drawings for accuracy and manufacturability before relying on them.

Mazda’s aerodynamics and crash-safety engineers are exploring vehicle designs with airflow and structural simulations. Astra in Codex helps the aerodynamics work by modeling new nose shapes, simulating airflow and refining promising variations. Styling and cost remain part of the assessment alongside performance. This is an exploration of design choices through simulation, with Mazda’s engineers evaluating the results; the announcement provides no measured drag reduction or validated vehicle design from the collaboration.

The distinction matters for assessing progress. Chip Ganassi Racing must determine whether a recovered model accurately represents an existing part and can support a manufacturing drawing. Mazda must judge whether a simulated shape remains useful once other vehicle requirements are considered. OpenAI has invited engineering firms and software vendors to help develop the work around their tools and engineering problems.

Independent CAD research identifies persistent errors

A conceptual GPT-6 Astra CAD assembly shows how mating and parameter errors can persist despite plausible geometry.

An October 17 preprint by Qiang Zou and colleagues evaluated GPT-6 Astra on CAD understanding, generation, editing and evaluation. It found successful cases involving nontrivial geometry alongside recurring errors in freeform shapes, parameter dependencies and mating relationships between parts. Iterative agent refinement improved some results but did not consistently resolve critical errors, and long completion times further constrained practical use.

Those failure modes affect different stages of the announced work. A flawed freeform surface can change a proposed vehicle shape; a broken parameter dependency can make later edits unreliable; an incorrect mating relationship can leave an assembly that looks plausible but does not fit. The preprint is an independent assessment of CAD tasks, separate from OpenAI’s BenchCAD comparison and the partner trials. Its results identify reasons for inspection without supplying a success rate for Mazda, Chip Ganassi Racing or the device prototype.

Where engineers still have to sign off

The prototype revisions and CAD findings point to review points at the handoff between generated output and an engineering decision. The following checks are an editorial synthesis of the disclosed failure modes and partner workflows, not a published certification procedure:

Inspect freeform surfaces and solids for the intended shape, valid geometry and continued editability.
Change controlling dimensions and confirm that dependent features update as intended.
Verify part mating, clearances, tolerances and assembly order against the proposed manufacturing process.
Match manufacturing drawings and circuit-board connections to the approved design, and examine the inputs and assumptions behind any simulation.

For a U.S. engineering team, the useful comparison is the full time from a design request to a reviewed result, including corrections made after a model generates its first answer. The next meaningful evidence from the partner work will be a completed scan-to-drawing or simulated design iteration accompanied by the engineers’ review effort and the quality of the resulting design.