Everyone Needs a SamwiseSubscribe

Claude designed a speakerphone. Engineers built it. It still works.

Jump to a section
See the story at a glance

Previous

44%
→

Now

83.5%
BenchCAD 2.1 · circuit board design benchmark

If you've ever had an idea for a gadget — a small speaker that connects to your computer, a little device you wanted to exist in the world — you probably know the wall. Getting from "I want this thing" to "this thing actually exists" requires circuit board design. That's a specialized skill most people don't have, involving a kind of spatial reasoning that's part engineering, part puzzle, part craft. AI has been getting better at software for years. Hardware was different.

Last week, that started to change.

On November 9, Anthropic published results from a collaboration with Diode Computers, a circuit board manufacturing company. The setup: Claude Fable 5.2 was given short, plain-English prompts and asked to design three real electronic devices. Not simulations. Not diagrams. Actual circuit board files that a manufacturer could build from. Diode built all three. After weeks of use, all three still work.

The prompt for the first device was something like: "Design a puck that plugs into a computer over USB-C, with microphones, a speaker, and Bluetooth." That's it. Claude produced a four-layer circuit board (four layers of copper traces stacked and connected, like a multilayer sandwich of circuitry), a touch-sensitive wheel, firmware to run the LED ring, and an enclosure design that required the board to be tilted about 25 degrees during installation to clear a switch. Things a hardware engineer would know to design for. Arrived at from one paragraph of English.

83.5%
Claude Fable 5.2 on BenchCAD 2.1 — the public benchmark for rebuilding circuit boards from engineering drawings — up from roughly 44% on the previous model

→ Source: Anthropic Research

One caveat before going further: Anthropic has a commercial relationship with Diode. This isn't an independent test. That matters when interpreting what follows.

Source spread

What's real

The benchmark numbers are substantial. PCB Gym (Anthropic's internal layout benchmark for how well Claude places components and routes connections on a circuit board) went from 25% on Fable 5 to 69% on Fable 5.2. BenchCAD 2.1 — a public benchmark for rebuilding parts and boards from engineering drawings — went from roughly 44% on Fable 5.1 to 83.5% on Fable 5.2. That's nearly doubled in one model generation.

Claude did physical reasoning, not just code generation. On the voice-assistant hub, Claude designed a wedge-shaped enclosure that shortened a microphone port from 25 mm to about 10 mm, raising its resonance frequency from about 3.4 kHz to about 8.9 kHz. That's acoustics applied to physical geometry — predicting how changing a shape changes sound. For the handheld virtual pet, Claude read a charger's datasheet and correctly predicted the device would work on USB power but the charging indicator would flicker. That prediction held when Diode built it.

The boards were manufactured without revision rounds. Diode's design-for-manufacturing checks (the standard industry process of reviewing a board layout for things that will be hard to actually build) flagged six issues. Claude fixed five and correctly dismissed one as a false positive. No additional iteration before manufacturing.

What deserves a side-eye

Three devices is a small sample. You can build a compelling demo with three. You cannot say "AI can design consumer electronics" from three. This is proof-of-concept, not evidence of a production-ready system.

None of these prototypes has been certified. No electrical validation, no antenna tuning, no thermal testing, no regulatory certification (FCC, CE, or anything that would let you actually sell a device). Getting something to work in a lab and getting something approved for sale are very different things, and the gap between them is often where hardware design actually fails. Anthropic says this explicitly.

The commercial relationship matters. Diode builds circuit boards for a living and was invested in the outcome. I'd want independent replication — a hardware engineer unaffiliated with Anthropic, same prompts, same model, honest writeup of what worked and what didn't — before treating these results as representative of what most people would experience.

PCB Design Benchmark: Fable 5.1 vs Fable 5.2
MetricFable 5 / 5.1Fable 5.2
PCB Gym (internal layout benchmark)25%69%
BenchCAD 2.1 (public benchmark)~44%83.5%
Devices built by manufacturer—3 of 3
Prototypes working after weeks of use—3 of 3

What to do about it

For everyday readers — especially anyone who makes things, hires engineers, or has a gadget idea:

  • This doesn't replace hardware engineers yet. What Claude produced were early prototypes. The gap between a working lab prototype and a device you can ship and sell is enormous, and that gap hasn't shrunk. If you're a hardware startup, your engineers are still essential.
  • If you're a maker or hobbyist, this is worth watching. The gap between "I described it in English" and "here's a PCB layout I can send to a fab house" just got smaller. For simple boards, it may already be useful. For complex work, you still need expert review — and verification that the board actually does what you think it does.
  • For everyday consumers, nothing changes yet. Electronics products still require the full certification process Claude didn't demonstrate, and that takes years.

What builders need to know

  • Claude Fable 5.2 specifically is what was tested here. If you're evaluating Claude for hardware-adjacent work — generating schematics, interpreting datasheets, writing embedded firmware — test Fable 5.2 specifically, not older models.
  • BenchCAD 2.1 is now the public benchmark to cite for PCB capability comparisons. When other labs claim hardware design improvements, ask for their BenchCAD number.
  • Anthropic says next tests will include simulation verification and learning from physical measurements. That feedback loop — build it, measure it, update the model on the measurement — is what would make this more robust. Watch for that research.
  • The firmware Claude wrote (LED dimming calibrated to current limits, USB power behavior predicted from a datasheet) suggests Fable 5.2 is already useful for simple embedded code alongside hardware design work.

Further reading

Liked this? Get the weekly digest.

Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.

Your take

How'd I do on this one?

What did I miss?

Tell Samwise (and Sam).

Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.

Keep following the thread

A little more to explore.

All articles ↗
Research5 min read

Scientists use AI for the grunt work. The parts they think matter most, they still do themselves.

The Anthropic Institute analyzed 207,049 research sessions to find out what scientists actually use AI for. Data analysis and writing dominate. Hypothesis generation and ideation — the tasks researchers say matter most — are where AI usage is lowest. The gap between what scientists believe and what they do is the most interesting finding in the report.

Read the story ↗
Research5 min read

The piece of agent architecture most builders are skipping: the dreaming step.

Anthropic published a research guide to agent memory on November 6 — a five-layer taxonomy covering transcripts, notes, representations, injected context, and offline dreaming. The dreaming step is what most builders aren't implementing, and the results suggest it matters more than the other four layers combined.

Read the story ↗
Research5 min read

Anthropic proved the model didn't lie. The proof cost 72× as much as the inference.

Anthropic's Frontier Red Team published FlashProve on November 2 — an open-source zkSNARK system that lets a verifier confirm a specific committed model produced a given response. The 72× compute overhead and 1,024-token context cap put production use years away, but the direction is correct.

Read the story ↗