OpenAI Speeds Up Hardware Design: Prototype Built in Under a Month

An engineer gets a 3D scan of a part, but the original CAD model doesn't exist. Usually manual work follows: reconstruct the geometry, produce a drawing, check dimensions, prepare the model for manufacturing. OpenAI claims that Codex and GPT-6 Astra can already take over a significant part of this chain, from building CAD models and printed circuit boards to setting up simulations and finding discrepancies between documents.

The practical point here is not the catchy phrase "AI designs hardware." It is the shortened time between an idea and the first physical prototype. In an OpenAI experiment, two employees developed a device for communicating with Codex in less than a week, working on it as a side project. Astra helped define the look, design the enclosure and internal mounts, select electronic components, route the printed circuit board, create the interface and write the firmware. A contract manufacturer produced a small batch of prototypes in under a month.

What It Looks Like in Real Work

Imagine a team building a desktop controller. The engineer specifies the dimensions, the list of components, the connector layout and the cooling requirements. The model creates an initial enclosure, places the board and mounts, prepares drawings, and then changes related elements after a comment: for example, it moves a port and frees up space for a cable.

Previously, each such change could require edits in several tools at once. The agentic workflow tries to maintain links between mechanics, electronics and software. This is especially useful where there are many repetitive operations: converting a 3D model into a set of 2D drawings, setting up a standard simulation, reconciling specifications, or logging detected conflicts in a BIM project.

OpenAI reports that since February, token consumption by its hardware teams has grown 23-fold. This shows demand inside the company, but does not prove savings in time or money: the publication has no controlled comparison of timelines, cost and error counts before and after adoption.

Where the Model Still Makes Mistakes

The prototype had to be revised several times. Astra did not always account for a specific manufacturer's constraints, part tolerances and the information needed for assembly. In other words, the model could produce a geometrically convincing design that still cannot be sent to a machine or production line without verification.

This is an important boundary. A bug in a web interface is usually fixed with an update. A mistake in a board, enclosure or power component turns into a lost batch, a delivery delay, and sometimes a safety risk. That is why the engineer remains responsible for checking materials, loads, clearances, manufacturability and regulatory compliance.

On the BenchCAD 2.0 test, OpenAI's new internal model scored 87.5%, while GPT-6 Astra at maximum reasoning level scored 59.4%. The test checks reconstruction of three-dimensional parts and assemblies, as well as recognition of connections on boards. The gap is notable, but the results of the internal model and the test itself should for now be treated as developer data, not independent confirmation of readiness for industrial use.

What Teams Can Do Right Now

Don't start with a promise to "automate the whole department." Choose one measurable operation: drawing preparation, CAD reconstruction from a scan, simulation setup, or finding mismatches between a model and a specification. Record the baseline time, number of returns and cost of fixes, then compare them with the process involving the model.

All agent outputs must pass engineering review before manufacturing. It is also useful to give the model the rules of the specific production: minimum thicknesses, available materials, tolerances, equipment limits and assembly requirements. The more precise these boundaries, the lower the chance of getting a spectacular but unmanufacturable design.

The main takeaway: AI is becoming a working layer on top of CAD, ECAD, BIM and simulators. It speeds up the preparation of options and routine conversions, but the decision to release a physical product must for now remain with the engineer, who is responsible not for a convincing picture but for a part that actually comes off the line.