Level 1 — Absolute Beginner

OpenAI makes computer programs that can think and write. One of them is called Astra.

Two workers at the company asked Astra to design a small machine. The machine lets you talk to another program called Codex. Astra designed the box, the electric board inside, and the software.

The two workers finished in less than a week. It was not even their main job. Then a factory made a few of the machines in less than a month.

But there was a problem. Astra did not always follow the factory rules. People had to check the design again and again before the factory could build it.

a program

A set of instructions that makes a computer do something.

to design

To decide how something will look and work before making it.

a machine

A thing with parts that does work for you.

a box

A container, here the outside part of the machine.

software

The programs inside a computer or a machine.

a factory

A building where things are made.

a problem

Something that is wrong and needs to be fixed.

to check

To look at something to see if it is right.

Level 2 — Elementary

On Tuesday OpenAI published a post about using its models to design physical products. The most eye catching example was a small device that lets a person talk to Codex, the company's coding assistant.

Two team members built the device with the help of GPT-6 Astra, and they did it in under a week while treating the work as a side project. According to the company, Astra chose the look of the device, designed the case and the parts that hold everything in place, picked the electrical components, laid out the printed circuit board, and wrote the software that connects the device to Codex. A contract manufacturer then produced a small batch of prototypes in under a month.

The company does not hide the weak point. Getting the device ready for the factory still took several rounds of review by people. Astra did not consistently account for the manufacturer's production limits, the tolerances it works to, or the small details needed for assembly. OpenAI says those gaps show what future models must do better.

The post also describes other work. Since February the company's hardware teams have increased their use of these models 23 times over, measured in tokens. OpenAI is testing the same ideas with a racing team, which turns three dimensional scans of old parts into usable design files, and it announced a new research project with Mazda on car aerodynamics and crash safety.

a device

A small machine or piece of equipment made for a job.

an assistant

Someone or something that helps you with your work.

a side project

Work done in spare time, next to your main job.

a component

One of the parts that make up a machine.

a circuit board

The flat plate inside a device that connects the electrical parts.

a prototype

An early version of a product, built to test the idea.

assembly

The act of putting the parts of something together.

aerodynamics

The study of how air flows around a moving object.

Level 3 — Intermediate

Software has been eating itself for years; hardware has been the stubborn exception, because atoms do not forgive the way code does. In a research post published on 3 November, OpenAI set out to test that boundary. The company says it asked whether a device could be created without any direct human edits, and used GPT-6 Astra to define the functional requirements and then design the enclosure, the printed circuit board and the software integrations for a device whose purpose is to let people talk to Codex, its coding agent.

The division of labour is worth spelling out, because it is broader than the usual demonstration. By OpenAI's account, Astra developed the look of the device, drawing on a previous collaboration and on references from games and toys; designed the enclosure and the internal structures that hold the components in place; selected the electrical components and laid out the board, including placement and connections, so that everything fitted inside the case; and designed the user interface before generating the firmware and the integrations needed to reach Codex. Two team members completed this in under a week as a side project, and a contract manufacturer built a small batch of prototypes in under a month.

The honest paragraph is the interesting one. Getting the device ready for manufacturing still required several rounds of review and iteration, because Astra did not consistently account for the manufacturer's production limits, tolerances, or the details required for assembly. Those are precisely the constraints that separate a rendering from a product, and the company frames the gaps as a map of what future models need to do better. On its internal BenchCAD 2.0 test, which asks a model to reconstruct three dimensional parts and assemblies and to identify connections on circuit boards, a newer unreleased model scored an average of 0.875 against 0.594 for Astra at its highest reasoning setting.

The wider programme is already running inside the company and outside it. Token usage across OpenAI's own hardware and infrastructure teams has grown 23-fold since February, applied to computer aided design models, manufacturing drawings, simulations and design review. Grace Lam, a forward deployed engineer at the company, said Codex has dramatically accelerated how they move from an idea to an engineered part. Chip Ganassi Racing is testing workflows that convert three dimensional scans of existing parts into editable design files, useful when the original files no longer exist, and a new research collaboration with Mazda puts the models to work on airflow and structural simulations for vehicle design.

stubborn

Hard to change or move, resisting effort.

an enclosure

The outer case that holds and protects a device.

an integration

A piece of software that connects one system to another.

a tolerance

The small amount of variation a manufactured part is allowed.

Level 4 — Advanced

The interesting claim in OpenAI's 3 November research post is not that a model wrote some code, but that it crossed into the physical world, where the feedback loop is measured in weeks and the penalty for being approximately right is a bin full of unusable parts. The company says it tested whether a device could be created without any direct human edits, and used GPT-6 Astra to set the functional requirements and then produce the enclosure, the printed circuit board and the software integrations for a small object whose function is conversational access to Codex, its coding agent. Two staff members did it in under a week, as a side project, and a contract manufacturer turned out a small batch of prototypes in under a month.

The scope of what the model handled is the part worth dwelling on, because most demonstrations of this kind quietly confine themselves to one discipline. Astra, on the company's account, produced the industrial design, drawing on an earlier hardware collaboration and on the visual grammar of games and toys; the mechanical design, meaning the enclosure and the internal structures that locate the components; the electrical design, including component selection and board layout with placement and routing constrained by the case; and the interface, firmware and integrations. Mechanical, electrical and software engineering are ordinarily three sets of people and three sets of file formats, and the coordination between them is where product schedules usually go to die.

Then comes the paragraph that gives the post its credibility. Readying the device for manufacture still took several rounds of review and iteration, because Astra did not consistently account for the manufacturer's production limits, its tolerances, or the details required for assembly. This is the right failure to report, and a familiar one to anyone who has handed a beautiful drawing to a factory. Design for manufacture is not an aesthetic judgement but an accumulation of knowledge about a particular plant, its machines and the things that go wrong on its floor, and it is exactly the kind of tacit knowledge that does not sit tidily in training data. OpenAI's own benchmark records progress rather than arrival: on BenchCAD 2.0, which asks for reconstruction of three dimensional parts and assemblies and identification of connections on circuit boards, an unreleased internal model averaged 0.875 against Astra's 0.594 at maximum reasoning effort.

Around the demonstration sits a deliberate commercial strategy. Internally, token consumption by OpenAI's hardware and infrastructure teams has risen 23-fold since February, spent on computer aided design work, production drawings, simulation setup and design review across robotics and compute infrastructure; Grace Lam, a forward deployed engineer, reports iterating faster across CAD and converting three dimensional designs into production ready two dimensional drawings with far less manual work. Externally, Chip Ganassi Racing is reconstructing editable models from scans of parts whose original files are lost, and its engineer Matthew Ortman describes going from raw scan data to a solid body with the manual tweaks applied directly. A new collaboration with Mazda applies the models to aerodynamic and crash safety simulation, where Astra explores nose geometries to reduce drag. Mazda's principal engineer Akihiro Nakata frames the value as holding design intent and aerodynamic performance together without treating either as a compromise, which is a more candid account of what engineers actually argue about than most corporate testimonials manage.