Claude Fable 5.2 hits 69% on PCB layout test, up from 25%

Claude Fable 5.2 designed a speakerphone, a voice-assistant hub and a handheld virtual pet from short prompts, and Diode Computers, a board maker, built all three from its files into prototypes whose circuitry needed no fixes, Anthropic said on November 9. The measured part is three working prototypes and a layout score from an internal benchmark Diode ran. None of the boards has been through electrical validation, antenna tuning, thermal testing or certification, and Diode’s engineers built and tested them by hand because Claude was not connected to test equipment or to the machines that make and assemble boards.

For each device Claude wrote the specification, drew the schematics, placed the components, routed the copper between them, designed a 3D-printable enclosure, wrote the firmware and prepared the parts list. The speakerphone prompt asked for “a puck that plugs into a computer over USB-C and includes microphones and a speaker, supporting Bluetooth with a PCB antenna.” All three devices still work after weeks of use. Anthropic says it does not yet know how they will hold up with continued use or whether they would pass commercial certification, and they were not designed for mass production.

Which score is which

PCB Gym takes finished board designs, removes their layouts and scores how well a model places the parts and reroutes the connections. Diode ran it at Anthropic’s request: Fable 5.2 scored 69%, Fable 5 scored 25%. The public figure comes from BenchCAD 2.1, which tests whether a model can rebuild 3D parts, assemblies and circuit boards from engineering drawings and rendered images. There Fable 5.2 scored 83.5%, nearly double Claude Fable 5.1.

The two benchmarks use different baselines. The more dramatic jump, from 25% to 69%, is measured on the internal test and spans two releases, from Fable 5 to Fable 5.2, while the public BenchCAD gain is set against Fable 5.1 alone.

Ripping up and rerouting

Anthropic puts the biggest change in geometric reasoning: where parts go and how the copper runs between them. Earlier models placed supporting parts too far from the chips they served or oriented connectors the wrong way. On the hub, Fable 5.2 routed the most demanding connections first, USB and Ethernet, one pair of tracks at a time, keeping both tracks in each pair the same length and correctly spaced and adding small zigzags that lengthen a track so the signals arrive together.

Over the past year Diode designed many boards with earlier models that were 95% complete but still unusable, because those models rarely revisited placement choices that made the routing impossible to finish. Fable 5.2 changed course without being told to. On the four-layer speakerphone board it reassigned signals to different pins on two chips so the tracks under the touch controls could stay on one layer without crossing. When the back of the board got too crowded, it swapped signals on interchangeable pins and laid out a whole corner again instead of adding extra jumps between layers. On the hub, it noted the copper was “getting too tangled to trace visually” and decided: “I should just rip these traces [tracks] and start fresh.” It deleted and rerouted connections dozens of times before an arrangement worked.

The hub also got a router Claude wrote itself. KiCad’s built-in router, available through the software Diode provided, draws one track at a time between two points Claude picked, and the hub had long tracks running from a 100-pin connector. Claude judged that doing them one by one would need too much step-by-step guidance, so it wrote an algorithm that splits the board into a grid and searches it for a clear path. It took about half an hour to build and laid each long track in roughly a second. When an audio clock track ended up too close to the Ethernet tracks, Claude changed the router to leave more room around sensitive tracks.

Checks before the factory

Claude wrote code to check that a 3D model of each board would fit its enclosure without collisions. The speakerphone board could not be lowered straight in because the mute switch hit the back wall; tilted by about 25 degrees during installation, it fits, and Claude added that step to the assembly instructions. For the hub, it calculated that a 25 mm microphone port would resonate near 3.4 kHz, “right in the voice band,” and switched to a wedge-shaped enclosure that shortened the ports to about 10 mm and pushed the resonance to about 8.9 kHz, at the edge of the speech range. The speakerphone firmware dims the LED ring as the audio gets louder, to stay under the current limit Claude had designed into the board.

Asked whether the virtual pet would run without a battery, Claude read the charger’s datasheet and answered that it would run on USB power, predicting that the charging light would flicker because the charger keeps recharging a small capacitor in the battery’s place.

Claude also ran the manufacturability checks, which compare a board’s geometry with what the factory can make, and the sourcing checks on whether distributors have the parts. On the hub, six slivers of solder mask (the protective coating on the board) came out too thin; Claude widened the mask openings without moving any parts. It ignored a flagged connection between two grounds, reasoning that the checker had not recognized it as deliberate and that it would not affect performance. Diode manufactured all three boards straight from Claude’s files with no rounds of adjustments, and the Diode team said that surprised them.

Where it started a year ago

The work with Diode began a little over a year ago with narrower tasks, such as reading a chip’s datasheet and working out which parts the chip needs to run. In Anthropic’s write-up of that round, Claude produced reference designs in Zener, Diode’s open-source language for describing schematics, built on Starlark, a strict subset of Python. Its common failures then were missing datasheet details on how circuits should be configured, misreading reference schematic images and misusing Zener. Diode’s engineers preferred Sonnet 4.5’s designs 8 of 10 times in blind head-to-head comparisons. The models of that period could not reliably lay out a board, so the test then was engineers’ preference between schematics. This time it was boards that a factory built and that still run.

Anthropic is not alone here: OpenAI recently wrote about using its models to design a handheld device. Anthropic’s next step is to see whether Claude can design denser, faster and more complex boards, check them in simulation and learn from measurements taken on the finished devices, the part of the loop that for now runs through Diode’s engineers.