OpenAI Says Its Models Can Now Design Hardware End-to-End

OpenAI says two of its employees used a model called Astra to design a working electronic device — enclosure, printed circuit board, and firmware included — in under a week, with contract-manufactured prototypes delivered in under a month. The claim comes from an OpenAI announcement, and the company itself concedes the result required several rounds of human review before anything could actually be built.

The disclosure is part of a broader statement on what OpenAI calls "accelerating engineering for the physical world." The company says its models now support mechanical engineering, electrical engineering, and end-to-end system design, and that internal teams already use them to test ideas and iterate faster. No independent verification of any claim exists yet; every figure in the announcement comes from OpenAI itself.

The stakes are straightforward. Engineering workflows tie together mechanical designs, electronics, software, and manufacturing requirements through specialized tools that have historically resisted automation. If frontier models can genuinely compress that loop — preparing drawings, setting up simulations, coordinating cross-discipline changes — they touch a market of professional CAD, EDA, and simulation software worth tens of billions of dollars. If they cannot, this is another vendor narrative built on cherry-picked internal demos.

What did the Astra test actually produce?

The centerpiece of the announcement is a device users can employ to talk to Codex, OpenAI's coding agent. OpenAI says it used GPT-6 to define the functional requirements, then let Astra — described as delivering "state-of-the-art performance in multimodal and computer use workflows" — handle the rest without direct human edits.

According to OpenAI, Astra's responsibilities broke down into four areas:

• Aesthetics and design: Astra developed the device's look, drawing on OpenAI's recent Work Louder collaboration and on references from games and toys.
• Mechanical design: Astra designed the enclosure and the internal structures holding components in place.
• Electrical design: Astra selected components and designed the PCB (printed circuit board), including component placement and connections, to fit within the enclosure.
• UI/UX and software: Astra designed the user interface and generated the firmware and software integrations connecting the device to Codex.

Two team members built the test device with Astra in under a week, as a side project. A contract manufacturer then produced a small batch of prototypes in under a month. OpenAI did not disclose the batch size, the cost, or whether the prototypes functioned as intended.

The company did disclose the failure modes, which is where the announcement earns some credibility. "Getting the device ready for manufacturing still required several rounds of review and iteration," OpenAI wrote. "Astra did not consistently account for the manufacturer's production limits, tolerances, or details required for assembly. These gaps are helping us identify what future models need to do better."

In plain words: the model produced a design, but a design is not a manufacturable product, and humans closed that gap.

How much are OpenAI's own engineers using these tools?

The announcement includes one internal adoption figure. OpenAI's hardware and infrastructure teams — spanning robotics and compute infrastructure — use Codex and ChatGPT to create and revise CAD (computer-aided design) models, prepare manufacturing drawings, set up simulations, and review designs. Since February, model usage across those teams, measured in tokens, has grown 23-fold.

Token counts measure volume of use, not value of output. A 23-fold increase tells us OpenAI's engineers are consuming far more model compute; it does not tell us how much of that output survived engineering review. OpenAI does not claim otherwise, but the distinction matters for any firm benchmarking its own adoption against this number.

What does BenchCAD 2.0 measure?

OpenAI also published a benchmark result. BenchCAD 2.0 tests whether models can reconstruct 3D parts and assemblies and identify connections in circuit boards — tasks that combine coding with engineering knowledge rather than pure text generation.

A newer internal model achieved an average score of 0.875 on BenchCAD 2.0, against 0.594 for GPT-6 Astra at its highest reasoning setting. OpenAI framed the gap as evidence that capability is improving. The unnamed model has no release date, no external evaluation, and no stated availability — a familiar pattern in frontier-lab announcements, where the strongest numbers belong to systems customers cannot yet buy.

"While the engineering lifecycle includes many other elements beyond design, we view improvements in capability here as early evidence of the ability of AI models to positively impact iteration speed for professional workflows," OpenAI wrote. Note the hedging: design is one stage of a lifecycle that includes validation, certification, tooling, and production — none of which the benchmark touches.

Who is testing this outside OpenAI?

OpenAI says it works with "a small group of industry collaborators" on real engineering problems, with their engineers reviewing model outputs for accuracy and practical use. The company named two of them.

The first is an expanded collaboration with Chip Ganassi Racing (CGR), the motorsport organization. OpenAI and CGR's R&D team are testing workflows that convert 3D scans of existing parts into editable CAD models and 2D manufacturing drawings — a scenario OpenAI calls "particularly useful when the original design files are unavailable," a common problem in legacy manufacturing and racing programs. CGR engineers review the models and drawings for accuracy and manufacturability, and OpenAI says their feedback identifies where models reduce manual work and where they fall short. No results or metrics from the CGR work appear in the announcement.

The second is a new research collaboration with Mazda. OpenAI's research and Forward Deployed Engineering teams are working with Mazda's aerodynamics and crash-safety engineers on developing and evaluating new vehicle designs, using simulations of airflow and structural behavior. The goal, per OpenAI, is exploring performance improvements while accounting for styling and cost. Results will be shared "in the coming months" — meaning none exist in public form today.

What is OpenAI not saying?

Several material questions go unanswered. OpenAI does not say whether Astra is a product, a research preview, or an internal codename; does not name the contract manufacturer or the internal BenchCAD model; does not quantify how many human review rounds the test device needed; and does not report any collaborator-measured outcome, only the structure of the collaborations. The announcement also functions as a recruiting pitch — OpenAI explicitly invites engineers, engineering firms, and software vendors in design, simulation, manufacturing, and production of complex physical products to fill out a contact form and help shape model training.

"We're working with collaborators to understand where these capabilities can make the biggest difference and what needs to improve," OpenAI wrote. That sentence describes a research program in its data-collection phase, not a shipping capability.

The credible reading is narrow: OpenAI's models can now generate complete first-pass hardware designs fast, internal usage is growing quickly, and at least two named industrial partners are kicking the tires. Whether those designs survive contact with production limits, safety validation, and cost constraints at scale is precisely the question OpenAI's own disclosures leave open — and the one the Mazda and CGR results, when they arrive, will have to answer.