Anthropic says Claude designed three devices built without circuit fixes
The experiment covered circuit boards, cases and firmware. Human engineers handled the physical builds; the devices have not undergone commercial validation.
Loading page…
The experiment covered circuit boards, cases and firmware. Human engineers handled the physical builds; the devices have not undergone commercial validation.
Listen to this story
Anthropic’s November 9 experiment shows Claude Fable 5.2 taking electronics prototypes from device brief through board layouts, firmware and manufacturing files; Diode Computers built three designs directly from those files without requesting changes. The result suggests AI can reduce iteration between a hardware concept and an assemblable prototype, but it does not establish product readiness: the devices skipped electrical validation, antenna and thermal testing, certification, and mass-production design, while longer-term durability remains unknown.
Claude scored 83.5% on BenchCAD 2.1, nearly double Fable 5.1, though that benchmark measures a different task from designing a complete buildable device.
For the hub, Claude wrote a grid-search routing algorithm in about half an hour; it could then route each long track in roughly one second.
Claude adjusted the hub case to move microphone-opening resonance above voice frequencies and dimmed the speakerphone’s LED ring as audio grew louder.
A speakerphone, a voice-assistant hub and a handheld virtual pet moved from short prompts to working hardware in Anthropic’s latest electronics experiment. In research published November 9, the company says Claude Fable 5.2 designed all three devices and guided human engineers through building prototypes that needed no fixes to their circuitry.
The tests, conducted with circuit-board maker Diode Computers, began with a simple description of each device and its main hardware requirements. For the speakerphone, that meant a USB-C puck with microphones, a speaker and Bluetooth. Claude produced specifications, circuit diagrams, board layouts, a 3D-printable case, firmware—the software running on the device—and a parts list.
Anthropic contrasts those builds with earlier tests in which models struggled to arrange components that fit together. Supporting parts could end up too far from their chips, or connectors could face the wrong way. Diode’s work with earlier models produced many boards that were 95% complete but still unusable, because the remaining connections could not be finished.
Fable 5.2 instead prioritized connections carrying high-speed signals, high currents or signals sensitive to interference. It also revised layouts without being asked. On the speakerphone, it reassigned signals to different chip pins so copper tracks beneath the touch controls could stay on one board layer without crossing.
Alongside the working prototypes, Anthropic reports an 83.5% score on BenchCAD 2.1, nearly double Fable 5.1’s result. That public benchmark tests reconstruction of 3D parts, assemblies and circuit boards from engineering drawings and rendered images. It measures a different task from producing a complete device that engineers can assemble.
The hub required long connections from a 100-pin connector. Claude initially used the KiCad design program’s routing tool, drawing one track at a time between chosen points. It then wrote its own routing algorithm to search a grid for clear paths. Anthropic says that took about half an hour; the resulting tool could route each long track in roughly a second.
The case design also changed for electrical and acoustic reasons. Claude calculated that the hub’s original microphone openings would resonate within the voice-frequency range. It chose a wedge-shaped case with shorter openings to push that resonance higher. For the speakerphone, it wrote firmware that dims the LED ring as audio gets louder, keeping the device within its designed current limit.
Before manufacturing, Claude checked board geometry against factory limits and checked whether suppliers had parts in stock. On the hub, it fixed six strips of protective solder mask flagged as too thin by widening openings. It also dismissed a flagged ground connection after determining it was intentional. Anthropic says Diode manufactured all three boards directly from the files without requesting adjustments.
The physical work remained human. Claude was not connected to test equipment or circuit-board manufacturing and assembly machines, so Diode engineers built and tested the prototypes. Anthropic says the speakerphone still worked after weeks of use, and the hub and virtual pet were also working. Longer-term durability remains unknown.
The company’s limitations are substantial: the devices have not undergone electrical validation, antenna tuning, thermal testing or certification, and were not designed for mass production. The result is a demonstration of prototype design and assembly—not evidence that these designs are ready to become commercial products.
Loading discussion...
Join the conversation
Explain which evidence you would weigh more and why.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.