AI for hardware and CAD design: what's real in 2026

By Evals Review editorial team · Published 5 Nov 2026 · Last updated 5 Nov 2026

Short answer: In 2026, AI can draft real hardware designs: OpenAI says two staff used GPT-6 Astra to design a prototype device's enclosure, circuit board, interface and software in under a week. But engineers still review every output, because the model missed manufacturing limits, tolerances and assembly details, so treat these tools as fast drafters, not sign-off engineers. [OpenAI]

Based on vendor announcements, press coverage and published benchmarks, not hands-on tested. No sponsored or affiliate links on this page. How we review.

TL;DR

  • Real today: AI models drafting CAD models, circuit boards, drawings and simulation setups, with engineers checking the results. [OpenAI]
  • Still human: manufacturability, tolerances, assembly details and sign-off. OpenAI says its own model missed these. [OpenAI]
  • Unproven: vendor benchmark scores you can't rerun, and any claim that AI replaces an engineering team.
  • Evaluate a tool on your own parts, with your own checks, and count the review rounds.

What's real today, with sources

OpenAI's GPT-6 Astra prototype device

On 3 Nov 2026 OpenAI said ChatGPT and Codex "are helping engineers build for the physical world", across mechanical design, electrical engineering and simulation. [OpenAI on X] Two OpenAI team members worked with GPT-6 Astra, as a side project, to design a prototype device for talking to Codex in under a week. Astra designed the enclosure, circuit board, interface and software. [OpenAI on X] OpenAI's write-up adds that Astra selected the electrical components, placed them on the board and generated the firmware, and that a contract manufacturer built a small batch of prototypes in under a month. [OpenAI]

Inside OpenAI's hardware teams

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, and that model usage across those teams, measured in tokens, has grown 23-fold since February. [OpenAI] That is a usage figure reported by OpenAI, not a quality measure.

Chip Ganassi Racing: 3D scans to editable CAD

OpenAI is working with engineers at Chip Ganassi Racing to turn 3D scans into editable CAD models and manufacturing drawings. [OpenAI on X] OpenAI says this is most useful when the original design files are unavailable, and that the team's engineers review the models and drawings for accuracy and manufacturability. [OpenAI]

Mazda: exploring vehicle designs through simulation

OpenAI announced a research collaboration with Mazda's aerodynamics and crash-safety engineers to explore new vehicle designs using airflow and structural simulations. [OpenAI on X] OpenAI says it will share more progress and results in the coming months, so there are no published results yet. [OpenAI]

A $50 DIY Muse pendant, vibe-coded with Claude

The Verge's Sean Hollister built his own take on Meta's upcoming Muse Charm for $50, using a Waveshare ESP32-S3 1.75C touchscreen pendant and Meta's open-source Muse Gadget code, then had Claude add features such as spoken replies, a conversation history, a phone-camera remote shutter and a working watch face. [The Verge] Meta's SDK is Apache-2.0 licensed and lists the Waveshare ESP32-S3-Touch-AMOLED-1.75C as a supported board. [Meta GitHub] The build was widely shared on X. [X] This is firmware and software on an off-the-shelf board, not new hardware design, and Hollister says he has no coding experience and has "no idea if any of this code is good". [The Verge]

Where humans still check the work

  • Manufacturing readiness. Getting OpenAI's device ready 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. [OpenAI]
  • Every output in industry pilots. OpenAI says engineers review the outputs and help identify where the models need to improve. [OpenAI on X] Its collaborators' engineers review outputs for accuracy and practical use. [OpenAI]
  • Code you can't read. In the DIY pendant, the builder couldn't judge the code and kept the device on guest Wi-Fi as a precaution. [The Verge] Meta warns that tinkering can brick boards or void warranties. [Meta]

What's hype or unverified

  • Vendor benchmark scores. OpenAI reports that on BenchCAD 2.0 a newer internal model averaged 0.875 versus 0.594 for GPT-6 Astra at its highest reasoning setting. [OpenAI] These are OpenAI's own numbers on an unreleased model. BenchCAD itself is a public academic benchmark, [arXiv] but its 2.0 Core set behind reported scores is shared only under an evaluation agreement; a separate 100-case public set is open. [Hugging Face] Unverified by us.
  • "AI designed a product with no human edits." OpenAI framed the device as a test with no direct human edits to the design, but the same post says manufacturing needed several human review rounds. [OpenAI] Both are true; the headline alone is not the whole story.
  • "A non-coder built a better product than Meta." That is social-media framing. The Verge's build lacks the retail Muse Charm's cameras and 5G, and Meta hasn't shipped the Charm yet, so there is nothing to compare like for like. [The Verge] Unverified.
  • Mazda and Chip Ganassi Racing outcomes. No performance results have been published for either collaboration as of 5 Nov 2026. Unverified.

How to evaluate an AI design tool: a checklist

A checklist, not a score. Run it on your own work before you trust any demo.

  1. Use your own parts and boards. Pick 5 to 10 real past jobs with known-good answers: a bracket, a scan-to-CAD part, a two-layer board. Vendor demos are picked to look good.
  2. Check outputs with deterministic tools. Run design-rule checks on boards, rebuild and measure CAD in your own tool, and rerun simulations. Prefer benchmarks graded this way, such as BenchCAD (voxel overlap against reference parts, no LLM judge) [GitHub] and RWE-bench (physics-checked verifiers).
  3. Test manufacturability, not just geometry. Give it your manufacturer's tolerances, minimum wall thickness, trace widths and assembly constraints, and see whether it respects them. This is exactly where OpenAI says its model fell short. [OpenAI]
  4. Count review rounds and edit time. Measure hours from prompt to a design your engineer would sign off, and how many correction rounds it took, against your current process.
  5. Check editability. A dumb mesh or flat drawing is less useful than a parametric model with sensible features and constraints. Change one dimension and see what breaks.
  6. Check traceability. Can it explain why it chose a component or dimension, and cite the datasheet or standard?
  7. Check data handling. Know where your CAD files and schematics go, and whether they are used for training, before you upload proprietary designs.
  8. Keep a human sign-off gate. Every source on this page keeps an engineer in the loop. Do the same, and log what the reviewer had to fix.

For the general version of this process, see our 15-step eval checklist and how to choose an eval for your use case.

Comparing the consumer AI agents mentioned here? See Claude Duet vs Meta Muse vs OpenAI Dots.

FAQ

Can AI design a hardware product on its own in 2026?

Not end to end. OpenAI says GPT-6 Astra designed the enclosure, circuit board, interface and software of a prototype Codex device in under a week, but getting it ready for manufacturing still took several rounds of human review, because the model did not consistently account for production limits, tolerances or assembly details.

What did OpenAI announce for CAD and engineering on 3 Nov 2026?

OpenAI said ChatGPT and Codex are helping engineers with mechanical design, electrical engineering and simulation. It described the Astra prototype device, a Chip Ganassi Racing project turning 3D scans into editable CAD models and 2D manufacturing drawings, and a Mazda research collaboration using airflow and structural simulations. It also asked engineers, firms and software vendors to work with it.

Did someone really build a Muse Charm-style pendant for $50?

Yes, as reported by The Verge's Sean Hollister on 31 Oct 2026. He used a $50 Waveshare ESP32-S3 1.75C board with Meta's open-source Muse Gadget code and had Claude add features. He says he has no coding experience and doesn't know whether the code is good, and his version has no camera or 5G.

How should I evaluate an AI CAD or PCB tool?

Test it on your own parts and boards, check every output with your normal tools (design-rule checks, simulation, a drawing review), measure how many review rounds it takes to reach a manufacturable design, and prefer published benchmarks that grade with deterministic checks, such as BenchCAD or RWE-bench, over vendor demos.

Did Evals Review test these tools?

No. This page is based on OpenAI's announcement and posts, The Verge's article, Meta's repository and published benchmarks, read on 5 Nov 2026. We have not run GPT-6 Astra, Codex or Claude on hardware tasks ourselves.

How we checked: on 5 Nov 2026 we read OpenAI's 3 Nov 2026 blog post and X thread, The Verge's 31 Oct 2026 article, Meta's muse-gadget-sdk repository and the BenchCAD paper, repository and dataset pages. We did not run any of these tools ourselves. Last updated 5 Nov 2026.

Sources