OpenAI: GPT-6 Astra designed the entire Codex device by itself in one week

Hardware was considered the last stronghold that AI was reluctant to enter: there was too much physics, too many tolerances, and too many contractors. OpenAI says that stronghold has now cracked. According to the company, its hardware team tasked GPT-6 Astra with designing a device for talking to the Codex coding agent, and completed the project in under a week. What is more, it was a side project rather than the team's main assignment.

What exactly the model did

According to OpenAI, Astra handled the entire process itself. It designed the enclosure, selected the components, routed the printed circuit board, and wrote the firmware and interface. Previously, these stages were handled by different specialists: a mechanical engineer, an electrical engineer, and an embedded software developer. Now, a single agent keeps everything in context at once and coordinates decisions across layers. For example, if the board ends up a couple of millimeters wider, the enclosure is adjusted to fit it without a lengthy exchange between people.

A contract manufacturer then took over. It produced a small batch of prototypes in under a month. For comparison, the conventional cycle from an idea to the first working samples usually takes hardware startups several quarters.

Where the model stumbled

One important detail is easy to overlook: several rounds of revisions were required. Astra did not always account for manufacturing constraints, tolerances, and the finer points of assembly. This is precisely the difference between "the drawing is correct" and "the part can be molded, printed, and assembled by hand." The model handles digital geometry well, but a real factory floor knows things absent from its data: how plastic behaves, where a drill cannot reach, and the order in which components must be installed.

The benchmark and the numbers

OpenAI is also touting a leap in quality. A new internal model scores 87.5% on BenchCAD 2.0, while GPT-6 Astra reaches 59.4% at its highest reasoning level. The benchmark tests whether a model can reconstruct 3D parts and assemblies and identify connections on printed circuit boards. The company disclosed nothing about the internal model, so the figures cannot be verified. This is a familiar situation: the comparison was conducted using the company's own test and under its own conditions.

Another figure: token consumption by OpenAI's hardware teams has increased 23-fold since February. They use Codex and ChatGPT for CAD models, manufacturing drawings, and simulations in robotics and computing infrastructure. In parallel, the company is developing its own device, rumored to be something between a smart speaker and a smartphone.

What to do with this

First, this is OpenAI's own claim, with no independent confirmation so far. The "one week" refers to a side project carried out by a team highly proficient with the tools. The same result should not be expected from every team.

Second, the signal for hardware startup founders is real: the cost of the first iteration is falling, and the bottleneck is shifting from design to manufacturing feedback. Those who learn to feed contractor feedback back into the model quickly—tolerances, DFM reports, and real photos of defects—stand to gain the most.

Third, product teams can prototype hardware in parallel with software instead of waiting until the software is finished. A sensible first step is to take a simple task, such as an enclosure for a sensor or a carrier board, run it through the model, and send it to a contractor for a manufacturability review. This will quickly reveal exactly where the connection between your model and the manufacturing process breaks down.

The bottom line: AI is not yet replacing the factory floor or an experienced manufacturing engineer, but it is compressing the cycle from an idea to a physical object. That cycle has always been the main constraint in hardware.