OpenAI shows off AI-designed hardware and car simulations

We've gotten used to AI models churning out software code and marketing copy. But what happens when you ask an AI to design a mechanical enclosure, lay out a circuit board, or simulate airflow over a race car?

OpenAI just shared a detailed look at how its frontier models, specifically GPT-6 Astra and Codex, are pushing into mechanical, electrical, and physical product engineering.

A prototype designed in under a week

To see how far Astra could go in real-world computer-use workflows, OpenAI tried building a physical gadget without direct human design edits. The goal was to create a handheld device users could use to talk directly to Codex.

Two OpenAI staffers took this on as a side project. In under a week, Astra had mapped out the device's aesthetics, modeled the mechanical enclosure, picked the electrical components, routed the PCB, and generated the firmware. A contract manufacturer had physical prototypes in hand in less than a month.

There's a catch, of course. Getting the design actually fabricated took multiple rounds of review and tweaking. Astra struggled to consistently account for real-world manufacturing limits, tight tolerances, and small assembly details. As OpenAI noted, models still have a lot to learn about the messy physics of a factory floor.

23-fold token growth and car simulations

Despite those rough edges, internal adoption inside OpenAI is accelerating fast. Hardware and infrastructure teams working on robotics and compute clusters are now routinely using Codex and ChatGPT to modify CAD files, generate production drawings, and review specs.

Since February, model usage across OpenAI's hardware groups has surged 23-fold in token volume. On BenchCAD 2.0, a benchmark testing 3D part reconstruction and circuit connections, an internal model scored 0.875, blowing past GPT-6 Astra's 0.594 at maximum reasoning.

Now external partners are putting the tech through its paces. Chip Ganassi Racing is using the models to turn 3D scans of existing vehicle components into editable CAD files and 2D manufacturing drawings, especially when legacy files don't exist. Meanwhile, Mazda is teaming up with OpenAI researchers to simulate aerodynamics and crash safety for new car concepts.

Why it matters

We think this is one of OpenAI's smartest and most grounded technical updates in months. Instead of promising fully autonomous factories overnight, it highlights exactly where AI speeds things up and where physical realities still push back.

For engineers and hardware startups, these models won't replace a veteran machinist anytime soon, but they could dramatically lower the barrier to fast prototyping. Bridging industrial design, PCB layout, and firmware usually requires a whole squad of specialists. Compressing that initial back-and-forth into days changes the economics of bringing physical products to life.

For the broader tech landscape, CAD giants and simulation software vendors should be paying close attention. AI is moving beyond the browser window and into industrial design toolchains. The main bottleneck is obvious: while hallucinating code just means an error in terminal, hallucinating a mechanical tolerance means scrapped metal and blown budgets.