Designing a circuit board has long been a craft of patience: place the parts, route hundreds of copper tracks, check the factory can build it, then start over when something does not fit. On November 9, 2026, Anthropic published results suggesting an AI model can now handle most of that loop. Working with the board maker Diode Computers, it asked Claude to design three gadgets from short prompts, and all three prototypes ran without any fixes to their circuitry.
This article summarizes Anthropic’s own research post. The devices are early prototypes, not products, and the company is clear about that. Even so, the experiment shows how quickly AI is moving from writing code to shaping physical hardware.
What Anthropic and Diode tested
According to Anthropic’s research post, the team asked Claude Fable 5.2 to design a USB speakerphone, a home voice-assistant hub and a battery-powered handheld virtual pet. Each started with a short prompt. The speakerphone prompt, for example, asked for a puck that plugs into a computer over USB-C, with microphones, a speaker and Bluetooth through a printed antenna.
From there Claude wrote the specification, drew the schematics, placed the components, routed the copper connections, designed a 3D-printable enclosure, wrote the firmware and prepared a parts list. Diode’s engineers assembled the boards and ran the firmware under Claude’s direction. Anthropic says the speakerphone still works after weeks of use, and the other two devices are also in good working order.
The effort builds on an earlier collaboration. A little over a year ago, Anthropic and Diode tried simpler tasks such as reading a chip datasheet and working out which supporting parts a chip needs. Those older models could not reliably lay out a board or design parts that fit together.
Where the model improved: geometry and routing
The biggest change, Anthropic reports, is geometric reasoning. Earlier models put supporting parts too far from the chips they served or turned connectors the wrong way. Fable 5.2 produced more compact layouts and routed the most sensitive connections first, including high-speed signal pairs, high-current paths and noise-sensitive lines.
On the voice hub, Claude began with the USB and Ethernet connections. It kept each pair of tracks the same length and correctly spaced, added small zigzag tuning patterns so signals arrive together, and arranged the power supply parts for better efficiency.
Anthropic backs this with two benchmarks. On Diode’s internal PCB Gym, which scores how well a model places components and reroutes boards whose layouts have been removed, Fable 5.2 scored 69 percent against 25 percent for Fable 5. On the public BenchCAD 2.1, which tests rebuilding 3D parts and boards from drawings, it scored 83.5 percent, nearly double Claude Fable 5.1.
Revising its own work
Another notable shift is self-correction. Anthropic says earlier models rarely revised placement choices even when routing was clearly doomed, and Diode ended up with many boards that were 95 percent complete but unusable. Claude now changes course unprompted.
On the four-layer speakerphone board, it reassigned signals between pins on two chips so tracks under the touch-sensitive controls could stay on one layer. When a corner of the board got crowded, it swapped interchangeable pins and relaid the whole corner. On the hub, it judged the copper connections were getting too tangled to trace and deleted and rerouted them dozens of times until one arrangement worked.
Using tools, datasheets and manufacturing checks
Claude also consulted manufacturer datasheets. Asked whether the pet would run without a battery, it read the charger datasheet and predicted USB power would work but the charging light would flicker. For long tracks from a 100-pin connector, it wrote its own grid-based router in about half an hour, able to lay each long track in roughly a second, then tuned it to leave more room around an audio clock line.
Downstream effects were tracked too. Claude found the speakerphone board could not be lowered straight into its case because the mute switch would hit the back wall, and worked out a tilt of about 25 degrees as an assembly step. For the hub it estimated that a 25 mm microphone port would resonate near 3.4 kHz, in the voice band, so it proposed a wedge-shaped enclosure that shortened the ports to about 10 mm and moved the resonance to about 8.9 kHz.
It ran design-for-manufacturing and parts-sourcing checks itself and sorted the findings. It fixed six solder-mask slivers that were too thin by widening openings, and correctly ignored a flagged ground connection that was deliberate. Diode built all three boards straight from Claude’s files, with no rounds of manufacturing changes, which the team said surprised them.
What it means
For hardware startups and product teams, the promise is speed: a first prototype that works sooner, with engineers spending time on validation rather than layout. For education and makers, it hints at a future where a short written brief yields a buildable design.
The limits matter. Anthropic notes the prototypes have not been through electrical validation, antenna tuning, thermal testing or certification, were not designed for mass production, and Claude was not connected to test equipment or assembly machines, so people built and tested everything. The company also discloses a commercial relationship with Diode, and says OpenAI has described a similar handheld-device effort.
Conclusion
The next steps Anthropic names are denser, faster and more complex boards, checked in simulation and improved from measurements of finished devices. If that works, AI could become a routine partner in hardware design. For now, three small gadgets that simply worked are a credible first sign.