OpenAI Shows GPT-6 Astra and Codex Doing CAD, PCB Design and Simulation: BenchCAD 2.0, Mazda and Chip Ganassi

OpenAI says its newest models can now carry real engineering work, from drawings and simulations to circuit boards, and it backs that with one concrete experiment: a hardware device designed by GPT-6 Astra with no direct human edits. In its November 3 research post, OpenAI also reports that model use across its hardware teams grew 23-fold since February, and that a newer internal model scores 0.875 on a new BenchCAD 2.0 test against 0.594 for Astra.

This is a company write-up, so most numbers come from OpenAI alone. This guide explains what was shown, what is measured and what is not, and what engineers can do today. For background on the models involved, see our GPT-6 Astra showcase and the Astra vs Claude Fable 5.1 comparison.

TL;DR: what OpenAI claims and what is checkable

The device experiment: what "no direct human edits" means

OpenAI used GPT-6 to define the functional requirements and then had Astra handle four layers. The aesthetics drew on OpenAI's earlier Work Louder collaboration, which we covered in the Codex Micro keyboard post, plus references from games and toys. For mechanical design Astra designed the enclosure and internal structures. For electrical design it chose components and laid out the PCB, including placement and connections, to fit the enclosure. For software it designed the user interface and generated firmware and integrations for Codex.

The caveats are as important as the claim. OpenAI says getting the design ready for manufacturing "still required several rounds of review and iteration." Astra "did not consistently account for the manufacturer's production limits, tolerances, or details required for assembly." That is the gap between a design that looks right on screen and a part a factory can make. Anyone who has sent a PCB to fabrication knows design rules, minimum feature sizes and assembly constraints cause most rework.

Two points the post leaves open: the exact tools the model used to drive CAD and ECAD software, and how many review rounds were needed. OpenAI links the enclosure and Work Louder pages but publishes no design files in the text we could read.

How OpenAI's own hardware teams use Codex

OpenAI says 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. Token use across those teams grew 23-fold since February.

Three named OpenAI staff are quoted. Grace Lam, a forward deployed engineer, says Codex speeds the path "from an idea to an engineered part," including converting 3D designs into production 2D drawings and setting up simulations and finite element analysis. Qiong Wang, a technical staff member, says the tools help organize cross-functional constraints, explore architectures and run design reviews in ECAD environments, with potential to "take a board from concept to Gerber." Jackie Lee describes using Codex for architecture, engineering and construction work: checking information across drawings and specifications and flagging inconsistencies, with architects, engineers and BIM professionals reviewing results.

These are testimonials from OpenAI employees about OpenAI's own work. They show direction, not independent evidence. The 23-fold figure measures tokens, which tracks usage and not quality or hours saved.

BenchCAD 2.0: the one number you can compare

OpenAI says BenchCAD 2.0 tests whether models can reconstruct 3D parts and assemblies and identify connections in circuit boards. A newer internal model scored an average of 0.875, against 0.594 for GPT-6 Astra at its highest reasoning setting.

What that means: if the numbers are comparable, the unreleased model closes much of the gap to a perfect score on this test. What we cannot check: the benchmark's task list, its scoring rules, whether it is public, or how other labs' models do. OpenAI itself says the engineering lifecycle "includes many other elements beyond design." Treat the jump as a signal about the next model, not a measure of today's product. For a public circuit-design test to compare against, see our coverage of the EEBench circuit board benchmark.

Partners: Chip Ganassi Racing and Mazda

Chip Ganassi Racing. OpenAI is expanding its work with the racing team's R&D group to turn 3D scans of existing parts into editable CAD models and 2D manufacturing drawings. That is useful when the original design files are missing. CGR engineers review outputs for accuracy and manufacturability. R&D engineer Matthew Ortman is quoted saying the model has been useful going from raw scan data to a solid body, working with Codex directly in CAD, where feedback he gives is applied directly.

Mazda. OpenAI announces a new research collaboration with Mazda's aerodynamics and crash-safety engineers, using simulations of airflow and structural behavior to explore performance while accounting for styling and cost. In one example, Astra helps explore ways to reduce drag by modeling new nose shapes, simulating airflow and refining the most promising ones. Mazda principal engineer Akihiro Nakata says exploring dozens of variants helps show "which features matter for aerodynamics and where designers have room to explore." OpenAI says it will share more results "in the coming months," so no drag numbers exist yet.

Note that OpenAI's page carries a usage restriction on the Mazda material, so we do not reproduce its images here.

Why this matters for builders and engineers

Coding agents are moving into non-code files. Codex has been a software tool. These examples place it inside CAD and ECAD environments, where the output is a solid body, a board layout or a drawing. The same pattern showed up in our look at how Codex now sits behind Git at OpenAI: the agent is becoming infrastructure inside the company's own work.

Review is the bottleneck. In every example a human engineer reviews. OpenAI says so for CGR, Mazda and the BIM work. The speed gain is in drafting and exploring options, and the risk is in accepting an output that looks plausible. The device experiment shows exactly where: tolerances and assembly.

The benchmark race now includes physical work. OpenAI also released a LifeVal everyday tasks benchmark this week, and its cost-efficiency push with GPT-6.1 Sol is about cheaper runs. Engineering tasks need long, tool-heavy runs, so cost per task will matter here too.

Where this sits in the history of AI for hardware

AI help for hardware is not new, but it has been narrow. Autorouters and generative-design tools have long suggested trace paths or lightweight shapes inside one program. What OpenAI describes is different in scope: one general model that reads requirements, picks parts, lays out a board, shapes an enclosure and writes the firmware that ties them together. That is the "end-to-end system design" the post names, and it crosses the boundaries between tools that engineers normally bridge by hand.

The earlier public data point is the EEBench circuit test linked above, which measured how models handle board design in isolation. OpenAI's own device chains the steps together. The honest result, working prototypes plus a list of manufacturing gaps, is a more useful signal than a clean success story. It tells tool vendors where to focus: design rules, tolerances and assembly notes, the unglamorous constraints that decide whether a design ships.

For software vendors the opening is plain. OpenAI says it wants to work with "software vendors involved in the design, simulation, manufacturing, or production of complex physical products." A model that drives CAD through its normal interface needs stable automation hooks, readable file formats and fast simulation loops. Vendors with good APIs will be easier for agents to use, and vendors without them risk being routed around.

What to do now

1. If you are a hardware engineer, try Codex or ChatGPT on drafting tasks with low risk: converting a 3D model to a 2D drawing, writing a design-review checklist, or checking a spec against a drawing. Keep your normal sign-off.
2. If you run an engineering firm or sell design software, OpenAI invites you to fill in its collaborator form to shape training and tooling across mechanical, electrical, electromechanical and mechatronics work.
3. Check manufacturing constraints yourself. Give the model your fabricator's design rules as input, then verify with the fabricator's own checks.
4. Do not read the 23-fold figure as productivity. Ask for time-to-design and rework-rate numbers when vendors pitch you.

Open questions

• Which tool interfaces does the model use inside CAD and ECAD software, and can third parties use them?
• Will BenchCAD 2.0 be published so other labs can be scored?
• When will the newer internal model ship, and will it carry the same safety limits as Astra?
• What do the Mazda drag and crash results show once published?

Our read: this is a credible direction with honest caveats from OpenAI, but a long way from "a model designs a product." The strongest evidence is that the manufacturing gaps are named. The weakest is the usage number. Facts here are accurate as of November 3, 2026 and come from OpenAI's post.

Related reading

• GPT-6 Astra best demos and showcase
• GPT-6 Astra vs Claude Fable 5.1
• OpenAI Codex Micro and Work Louder keyboard
• EEBench: AI circuit board design benchmark
• OpenAI Pando: Codex at the Git backend
• OpenAI LifeVal benchmark
• GPT-6.1 Sol and Astra cost efficiency
• Source: OpenAI, Accelerating engineering for the physical world; earlier: OpenAI Work Louder collaboration page, Ars Technica on the Codex keyboard