1. Three working devices from three short prompts
In a launch test, Anthropic gave its newest AI model, Claude Fable 5.2, a one-paragraph prompt for each of three electronic devices: a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet.
The model wrote the specifications for each one, picked parts that US distributors had in stock, and laid out a four-layer circuit board for each device. It also modeled each enclosure and wrote the software.
People at Diode, a circuit board company, built the boards from those files.

Anthropic put the result in plain terms. "Each board was ready for manufacturing in under a day, with minimal human input." "All three devices worked without needing any electrical modifications." Both lines come from the announcement.
Two details will sound familiar to anyone who has built from someone else's drawings.
Anthropic's write-up says Diode built all three boards "directly from Claude's design files, without any back-and-forth over adjustments." The model ran its own build checks first. On the hub, those checks flagged six slivers of solder mask, the protective coating on a circuit board, as too thin. The model widened the openings and moved no parts.
The model also caught a fit-up problem with the speakerphone. Its board could not drop straight into the enclosure, because the mute switch would hit the back wall. The model worked out that the board had to go in at a tilt of about 25 degrees, and it added that step to the assembly instructions.


2. The limits behind the launch claim
This is the most complete case we have read of AI designing physical products.
Tests that score this kind of work have moved quickly, too.
Source: Anthropic, Fable 5.2 announcement. Scores reported by Anthropic, with tools, at maximum effort.
We see some of this as hype, and Anthropic states the limits itself. The devices are prototypes. They have not been through electrical validation, thermal testing, or certification, and they were not designed for mass production.
Anthropic has a commercial relationship with Diode, and we found no check of the result by anyone outside the two companies. Treat it as a direction, and watch for the first outside team that repeats it. Earlier models, the write-up says, produced boards that were "95% complete but still unusable."
Your engineers and estimators may read this as a threat to their jobs. But the model had no hands. It did the design work, and people built the boards, assembled the devices, and tested them.
We expect people to keep the building, the checking, and the sign-off. AI does the first pass. People verify, decide, and sign.
3. More people will design products
We think writing software no longer holds up most business projects. AI drafts the code, and people review it. Software went from something a company hired a team for to something one person can direct. Physical product design, in our view, is starting down the same road.
Early evidence has also come from outside Anthropic. In September, a developer with no circuit board experience had the earlier Fable 5 design a board and ordered it for 26 euros per board.
His computer recognized it on first plug-in. The design also called for two parts that did not fit their pads, which is why people still need to check the work.
We expect a new kind of buyer to show up, perhaps a small company with an AI-drafted design and no purchasing department. They will want five pieces, then fifty, and many will not call.
Instead, they will ask an AI assistant who can make the part, and the assistant will read websites, send files, and compare answers. Someone still has to make each design, and that demand lands on manufacturers.
4. What this asks of a manufacturer
We call a company that can serve this new buyer agent-ready. A buyer's AI assistant can find what the company makes, get a checked price, and get an answer about a job, and a person at the company approves what goes out.
Our view is that many shops are not there yet, and that the gap sits in their records. When RSM surveyed mid-size companies in 2026, manufacturers named data quality (32%) and connecting to older systems (27%) as the top limits on wider AI use.
Typically, capabilities live in a salesperson's head, costs live in an estimator's spreadsheet, and job status lives with whoever you can reach by phone. An AI assistant cannot read any of that.
We expect the assistant to move on when it finds no listed materials or certifications, or when the site gives it no way to request a price.
Check one thing about your data before you use any of these tools yourself. Anthropic keeps prompts sent to Fable models for 30 days for safety monitoring unless a customer has a specific enterprise arrangement.
Before you send drawings or pricing to any model, ask the vendor how long it keeps prompts and who can read them. Then give the AI its own login, limited to what the job needs, log every action, and have a named person approve what leaves the building.
What to do this week
- Ask an AI assistant such as Claude or ChatGPT who makes what you make, in your region. Note whether your company comes up and what it says about you.
- Read your own website as a stranger would. Check whether it lists your processes, materials, size limits, and certifications in plain text.
- Count how long your last ten quotes took from request to send. Count how many small requests got no answer.
- Pick one question customers ask most, such as "where is my order?" Write down how many places someone looks to answer it.
- Decide who on your team would approve an AI-drafted quote or reply. Write down what that person would check every time.
For how cheaper AI changes which RFQs are worth answering, read The Quotes You Never Sent. For how to set up any AI role with clear limits, read The New AI Org Chart.
Questions people ask
What is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's top AI model, released November 9, 2026. Anthropic recommends it for the hardest technical work, with Opus 5.5 for most daily tasks.
Did an AI design working hardware?
Anthropic reports that Fable 5.2 designed three working electronic prototypes from one-paragraph prompts. The designs covered circuit boards, enclosures, and software, and a partner company, Diode, built them. The devices are prototypes. They have not been certified or validated for production, and we found no check by an outside party.
What does agent-ready mean for a manufacturer?
It means a buyer's AI assistant can find your company, get a checked price, and get an answer about a job, while a named person on your team approves what goes out.
Terms in this piece
- Model
- the AI itself, trained by a lab such as Anthropic, OpenAI, or Google. Products like Claude and ChatGPT run on top of models.
- Agent
- an AI model set up to take steps on its own, such as reading files or running tools, inside limits a person sets.
- Firmware
- the software that runs inside a physical device, such as a speakerphone.
- Agent-ready
- a company that an AI assistant can find, get a checked price from, and get answers from, with a person approving what goes out.



