Back to Explore
NewsVibeFix editorialUpdated Nov 13, 2026

GPT-6 Astra Designed a Device to Talk to Codex in Under a Week — Then the Factory Sent It Back for Rework

On November 3 OpenAI said GPT-6 Astra designed a device for talking to Codex — enclosure, circuit board, interface and firmware — in under a week, with no direct human edits. Prototypes followed within a month. But getting it ready for manufacturing still took several rounds of review: the model could design, it could not yet satisfy a factory.

A prototype circuit board photographed on a workbench: a green PCB with USB ports and chips, tweezers and a screwdriver beside it.

On November 3, OpenAI published a research post on its website titled "Accelerating engineering for the physical world." The star of the post is not a new product but an experiment: GPT-6 Astra was asked to design, from scratch, a small device for talking to Codex — its looks, its enclosure, its circuit board, its interface, its firmware — with no direct human edits to the design. Two team members did it as a side project and finished the design in under a week; a contract manufacturer built a small batch of prototypes in under a month.

It reads like the hardware version of vibe coding. Over the past year, "describe an idea, get a working app" has gone from demo to daily routine. Now OpenAI is asking whether the next step can be "describe an idea, get something you can hold in your hand." What makes this post worth reading is that it does not only tell the successful half of the story: the design got done, but the factory stage did not go through cleanly. The model can draw a circuit board but cannot reliably account for tolerances, production limits, and assembly details, and getting the design ready for manufacturing still took several rounds of human review and rework. The conclusion up front: AI can now run the whole hardware design process, but between "can design it" and "can manufacture it" lies exactly the kind of experience engineers are paid for.

A prototype circuit board photographed on a workbench: a green PCB with USB ports and chips, tweezers and a screwdriver beside it.

1. What the device is: TamaGPT, a gadget for talking to Codex

The test product, known internally as TamaGPT, is a small desk device with a screen and buttons, built for one job: letting you talk directly to Codex. Codex is OpenAI's coding agent, normally living in terminals and editors; this device tries to lift it out of the screen and put it on the desk. OpenAI says the goal of the experiment was, in its own words, "seeing if we can create a device without any direct human edits." Note the wording: not "without people," but "without people directly editing the design." That distinction is the most important piece of honesty in the whole post, and it deserves unpacking.

The division of labor went like this: GPT-6 defined the functional requirements first, and then Astra took over four layers of design. On looks, Astra developed the design of the device, drawing on OpenAI's recent Work Louder collaboration and on references from games and toys. On mechanical design, it designed the enclosure and the internal structures that hold the components in place. On electrical design, it selected the components and laid out the printed circuit board, including placement and connections, to fit inside the enclosure. On software, it designed the user interface and generated the firmware and the integrations needed to connect the device to Codex. In other words, from "what it looks like" to "how it gets power and talks to Codex," all four layers were produced in one continuous chain by the model, with no handoff to a human engineer to fill in design gaps.

The timeline is the most striking set of numbers in the post: two team members, working on this as a side project, designed the test device in under a week; a contract manufacturer then built a small batch of prototypes in under a month. For anyone who has built hardware, that speed is counterintuitive. In traditional hardware development, enclosure tooling, board fabrication, and firmware bring-up routinely take months of back and forth. Astra compressed "from idea to the first thing you can hold" down to a matter of weeks — not by magic, but by removing the waiting and handoffs between the four disciplines. Looks, structure, circuit, and software were no longer four people queuing up for each other, but one model working continuously from a single shared understanding.

2. What "no direct human edits" actually means

This phrase is worth taking apart, because it is easy to misread as "AI built a device by itself," while OpenAI's own wording is far more restrained. First, the functional requirements were defined by GPT-6, but the premise itself — "build a device for talking to Codex" — came from people. The experiment's goal, its acceptance criteria, and the device's purpose were all set by humans first. Second, "without any direct human edits" refers to the design outputs: the enclosure drawings, the board layout, and the firmware code were not directly edited by a person. That does not mean people were not involved: the two team members worked alongside Astra throughout, setting requirements, making judgments, and choosing directions, with the model working inside a goal framework set by humans.

More important is the turn OpenAI writes into the post itself: "Getting the device ready for manufacturing still required several rounds of review and iteration." The first draft of the design was produced by the model alone, but the draft that could be manufactured was not. Astra "did not consistently account for the manufacturer's production limits, tolerances, or details required for assembly." This admission draws the boundary of the experiment clearly: the model passed the test of "going from zero to a design," but not the test of "going from a design to something a factory can build." Confusing the two leads to the wrong conclusion that AI can already replace hardware engineers; OpenAI's writing is designed to stop you from drawing it.

For vibe coders, this boundary will feel familiar. Writing software works the same way: an agent can build a working app overnight, but handing it to real users — handling edge cases, fixing flaky bugs, adding monitoring and rollback — still takes round after round of human review. Hardware only amplifies the pattern: a software mistake can be hot-fixed, a hardware mistake means a new board spin and new tooling, with every rework costing real money and weeks of time. "Zero human edits to the design" is a real milestone; "zero human involvement in manufacturing" is not yet one. The honest thing about this post is that it puts both sentences on the table at the same time.

3. The factory stage: tolerances, production limits, and assembly details

Why can a model draw a circuit correctly but not get tolerances right? It comes down to what "correct" means in hardware design. A board that is correct in principle — the right components, the traces connected, the firmware running — is only the first layer of correctness. The second layer is "can be built": are the trace widths within the factory's process capability, are the pads sized for the pick-and-place machine, is there room left for servicing and heat dissipation. The third layer is "can be assembled": will the enclosure clips snap shut on the first try, do the screw holes line up, will the cable routing get pinched by a structural part. The answers to the second and third layers are not in textbooks; they live in the equipment, process windows, and hard-won experience of a specific factory.

That is exactly where Astra fell short. OpenAI says the model did not consistently account for the manufacturer's production limits, tolerances, or assembly details. Note the word "consistently": not that it knew nothing about these things, but that it sometimes accounted for them and sometimes did not. In software terms, this is like an agent that sometimes remembers to write tests and sometimes forgets, sometimes handles null pointers and sometimes misses them. In engineering, inconsistency is the hardest defect to live with. A tool that makes the same mistake every time can be worked around with process; a tool that is right half the time forces you to personally check every step, which costs more effort, not less.

This gap also explains why OpenAI spends so much of the second half of the post on collaborators. What the model lacks is not design ability but "factory knowledge": the limits of a specific pick-and-place machine, the shrinkage rate of a specific material, the assembly sequence of a specific production line. That knowledge is scattered across real factories and the engineers who work in them, not in public training data. So OpenAI's answer is not to let the model figure it out alone behind closed doors, but to work with the people who hold that knowledge — the R&D team of a racing outfit, Mazda's aerodynamics and crash-safety engineers. The model raises the speed of design iteration; people supply the judgment about "can it be built, is it safe." That division of labor is likely to be the normal shape of AI hardware engineering for years, not a transitional phase.

4. 23-fold: OpenAI's own teams are already using it

Beyond the test device, the post offers another number: since February, model usage across OpenAI's hardware teams, measured in tokens, has grown 23-fold. These teams use Codex and ChatGPT to create and revise CAD models, prepare manufacturing drawings, set up simulations, and review designs, spanning robotics and compute infrastructure. The 23-fold figure needs careful reading: it measures token usage, not output, not time saved, not design quality. Heavy token use could mean the tools work well, or it could mean people are iterating through many failed attempts. Either way, a 23-fold increase is at least evidence that this is not a one-off demo — OpenAI's own hardware teams have woven the models into their daily workflow.

Corroborating coverage quotes Grace Lam, a Forward Deployed Engineer at OpenAI: "Codex has dramatically accelerated how we move from an idea to an engineered part." (This quote does not appear in the text of the OpenAI post as read for this article; it is cited from secondary reporting of the post.) The value of the quote is its source: not marketing copy, but a frontline engineer describing her own workflow. Together with the 23-fold number, it paints a picture of models already serving as everyday tools inside OpenAI's hardware teams, handling the stretch between an idea and a part — drawing, revising, simulating, reviewing designs. The expensive part, the manufacturing judgment described in the previous sections, still sits with people.

There is a practical lesson here for outside readers: one plain way to judge whether an AI capability is genuinely useful is to check whether its maker uses it, for how long, and at which stage. In this post, the models are used in OpenAI's own hardware teams' daily work, have been for the better part of a year, and usage is up 23-fold — evidence that carries more weight than any benchmark score. But it is equally worth noting what the post does not provide: no numbers on time saved or rework avoided. 23-fold measures input, not output, and the two should not be confused.

5. BenchCAD 2.0: how to read 0.875 versus 0.594

The post also publishes a benchmark score. BenchCAD 2.0 tests whether models can reconstruct 3D parts and assemblies from engineering drawings or renders, and identify connections in circuit boards. A newer internal model achieved an average score of 0.875, compared with 0.594 for GPT-6 Astra at its highest reasoning setting. Put side by side, OpenAI's point is that the new model scores about 47% higher than Astra, and that hardware design capability is still climbing fast.

Both numbers need three qualifiers. First, the 0.875 model has no name and no ship date. You cannot map it to any model you can buy or use; it is more of a signpost than a product. Second, BenchCAD 2.0 tests only the "design" slice of the job, not the whole engineering lifecycle. OpenAI admits this in the post itself: the engineering lifecycle includes many other elements beyond design. A high design score does not mean strong manufacturing ability — the TamaGPT device just proved that: a design can be completed while manufacturing still needs rework. Third, the scores are OpenAI's own figures, and the public BenchCAD leaderboard has not yet listed the unreleased model. On BenchCAD's public table, Astra holds the top published score at 0.59, with Claude Opus 5.5 at 0.57 — a small gap, with Astra also cheaper and faster per case.

It is worth noting that BenchCAD 2.0 itself is a serious benchmark. Published by academics on October 12, 2026, its Core set is 100 held-out cases scored against references, "without an LLM judge." That design — not using a model to grade models — is relatively clean amid today's flood of benchmarks. So the right way to read these numbers is: hardware CAD is still hard for every model — the best published score is only 0.59, far from a perfect 1 — but the gaps between generations are widening, and OpenAI already has a clearly stronger next generation internally. Astra and Opus 5.5 are neck and neck on the public table; the real dividing line may be the unreleased 0.875.

6. A racing team and Mazda: why bring in outside collaborators

The second half of the post describes two collaborations. The first is with Chip Ganassi Racing (CGR), a racing team. OpenAI is expanding its work with CGR's R&D team to test workflows that turn 3D scans of existing parts into editable CAD models and 2D manufacturing drawings. The practical point is specific: many old parts no longer have their original design files — only the physical object. To reproduce one, you scan it first, then turn the scan data into a CAD model you can edit and machine. Matthew Ortman, an R&D engineer at CGR, says in the post that the model "has been useful for going from raw scan data to a solid body." Meanwhile, CGR's engineers review the model's drawings for accuracy and manufacturability — the model supplies speed, people supply the gatekeeping.

The second collaboration is newly announced: Mazda. OpenAI's research and Forward Deployed Engineering teams are working with Mazda's aerodynamics and crash-safety engineers to explore new vehicle designs using simulations of airflow and structural behavior, while accounting for styling and cost. The post also notes that GPT-6 Astra is helping Mazda's aerodynamics engineers explore ways to reduce vehicle drag by modeling new nose shapes, simulating airflow, and refining the most promising designs. Mazda has made no production commitment, and OpenAI says it will share more progress and results in the coming months.

Read together, the two collaborations reveal OpenAI's strategy. The racing team addresses the problem of "what do you do when there are no drawings"; Mazda addresses "how do you design fast and safe at the same time." Both are hard problems a model cannot solve on its own: the first needs real scan data and manufacturing experience, the second needs engineering judgment in crash safety, where the tolerance for error is near zero. The post ends with an open invitation to collaborators — engineers, engineering firms, software vendors — to help shape the effort. That reads less like a product launch and more like a recruitment drive: the model can already run the design leg, and what is missing now is real-world engineering knowledge, which only lives with the people who do the work.

7. What it means for vibe coders: from writing apps to making things

This post connects to vibe coding more directly than it first appears. Over the past year, vibe coding has stayed inside software: you describe an idea, an agent writes code that runs. Hardware has always been treated as something else — it costs money, it waits on shipping, mistakes cannot be rolled back with one keystroke. The TamaGPT experiment is the first time the vibe coding way of working has been carried, intact, into hardware: one person, one loop of "say clearly what you want, let the model produce the first version, review and correct it," across looks, structure, circuit, and software.

What transfers is the way of working; what does not transfer is the cost of being wrong. In software, when an agent writes the wrong code, tests tell you immediately and the fix takes minutes. In hardware, when a model misses a tolerance, you only find out when the fabricated boards come back, and each revision means new scheduling and new payment. That asymmetry decides the right posture for hardware vibe coding: models are good for producing the first version fast, testing directions fast, and validating ideas fast; but before any version moves toward manufacturing, it must pass review by someone who understands production. In other words, the agent in hardware is not there to "finish the job for you" but to "compress the cost of trial and error from weeks to days," with the final "can it be built" gate still held by a person.

There is another practical insight hidden in the detail "two people, as a side project, in under a week." The device's design did not require a large hardware team; two people, outside their day jobs, used the model as leverage to string together four professional disciplines. That suggests the change is not only that "big companies can build things faster," but that "small teams can now reach a hardware threshold that used to be out of reach." An independent developer who once wanted to build a small hardware product either had to learn four disciplines or pay four kinds of specialists; now the first version of the design can be produced by the model, with people holding only the review and manufacturing gates. The door has only just opened a crack, but the direction is clear.

8. What the post does not say: no product, no design files, no timeline

What this post leaves out is as worth listing as what it includes. First, there is no product. TamaGPT is a test device, not a product for sale; OpenAI says nothing about whether it will be sold, at what price, or when. Reading it as a capability demonstration is correct; reading it as a product launch is not. Second, there are no design files. The post links to an enclosure drawing but publishes no complete set of design files, firmware code, or bill of materials, so outsiders cannot reproduce the device or independently verify the "no human edits" claim. Third, there is no timeline. The internal model that scored 0.875 on BenchCAD 2.0 has no name and no release plan; the Mazda collaboration has no results timeline either, only a promise to share more in the coming months.

Fourth, and easiest to miss: every number and conclusion in the post comes from OpenAI itself. The design finished in a week, prototypes built in a month, 23-fold usage growth, 0.875 versus 0.594 — none of it has third-party verification or independent testing. BenchCAD's public leaderboard can cross-check Astra's 0.59, but the 0.875 internal model is not on it yet. The right way to read this post is as OpenAI's self-report on its own capabilities, not an externally audited scorecard. A self-report can still be highly informative — especially the part where it volunteers its own shortcomings — but it still needs time to be verified.

One final note on dating: the OpenAI article page itself showed no date stamp when it was read for this article, so the November 3 date used here is the one consistently adopted across corroborating coverage, including Unite.ai and CellCog reporting on November 3–4. This article follows that cross-checked date and records the verification method in its notes. If OpenAI later adds a date stamp to the page, the page governs.

9. Three practical suggestions

First, if you are a vibe coder from a software background who wants to touch hardware, start with "let the model produce the first version," not "let the model handle manufacturing." Use models like Astra and Codex to draw the first enclosure, lay out the first board, and write the first firmware — validating ideas and getting a demo running is where the model already saves you a great deal of time. But before ordering a fabrication run, have someone who understands manufacturing review the drawings — tolerances, process, assembly. These are the three things models still handle unreliably, and skipping that review will cost you more in rework than you saved.

Second, treat "design" and "manufacturing" as two separate milestones. Many hardware projects die not because the design could not be done, but because the team mistook "design finished" for "product finished," budgeting and scheduling against the former and getting dragged down at the manufacturing stage. The TamaGPT timeline is a good reference: design in under a week, prototypes in under a month, but manufacturing readiness still took "several rounds of review and iteration." Your project plan should reserve separate time and budget for that second phase.

Third, watch benchmarks like BenchCAD, but do not watch only the scores. Scores tell you how a model performs on isolated tasks such as rebuilding parts and recognizing traces; what actually decides whether you can use it is how stable it is on your specific task: ask the same thing three times, do the answers agree, and does it flag edge cases on its own. Rather than waiting for a 0.875 model to ship, take the model you have today and run it on one small part of your own — from idea to a first drawing — and test for yourself how much it can actually help. One hands-on run tells you more than ten explainers.

In one sentence: this post is not saying "AI can build things now," but "the first leg of building things with AI is open, and the remaining leg is called manufacturing." The speed revolution in design is real — a design in a week, prototypes in a month are hard numbers. But tolerances, process, and assembly remain human-held currency. The next stop for vibe coding may well be hardware, but it will not skip the factory — it will first learn to work with it.

Views 4Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates