Testing Claude on real-world electronics design

Anthropic collaborated with circuit-board design and manufacturing company Diode Computers to evaluate how artificial intelligence handles physical electronics engineering from concept to working hardware. In Anthropic Research's report, the team demonstrated Claude Fable 5.2 designing three fully functional hardware prototypes: a USB-C speakerphone, a home voice-assistant hub, and a battery-powered handheld virtual pet. Operating from high-level natural language prompts, the model produced complete technical specifications, drew circuit schematics, determined physical printed circuit board component placements, routed multi-layer copper tracks, drafted 3D-printable enclosures in CAD software, and wrote the operational device firmware without requiring human circuit fixes.

This reporting details how frontier models handle physical engineering challenges, geometric spatial reasoning, and manufacturing constraints under rigorous hardware benchmarks. The findings serve hardware engineers, electrical designers, embedded firmware developers, and technical leaders evaluating artificial intelligence for physical product design and industrial CAD workflows. Earlier experiments conducted roughly a year prior saw models fail to position components correctly or close connections reliably, whereas the updated system delivered physical prototypes that remain operating after weeks of testing.

01Geometric reasoning across multi-layer circuit boards

Claude Fable 5.2 demonstrated significant spatial planning improvements over previous model iterations when determining component locations and routing copper traces. Earlier models routinely placed supporting parts too far away from host chips or misoriented physical connectors, resulting in broken layouts.

During the voice-assistant hub design, the model prioritized high-speed and noise-sensitive signals first, routing differential pairs for USB and Ethernet tracks with identical lengths and proper spacing. It independently introduced tuning patterns featuring zigzags to synchronize signal arrivals while arranging power supply components to boost operational efficiency.

The model validated these spatial capabilities across dedicated hardware engineering benchmarks. Diode evaluated the model using PCB Gym, an internal benchmark that strips component placements and wiring from existing circuit boards and scores the model's ability to reconstruct viable layouts.

Claude Fable 5.2 scored 69 percent on PCB Gym, up from 25 percent scored by Claude Fable 5. On BenchCAD 2.1, a public evaluation measuring 3D part reconstruction and board assembly from drawings and renderings, Fable 5.2 scored 83.5 percent, nearly doubling the performance of Claude Fable 5.1.

02Autonomous revisions and trace rip-up decisions

Prior models rarely restructured earlier layout errors, frequently leaving boards 95 percent completed yet fundamentally unusable because blocked trace channels prevented final routing. Claude Fable 5.2 recognized dead ends independently and reorganized board architecture without user intervention.

While designing a four-layer printed circuit board for the USB speakerphone, the model reassigned signals between interchangeable pins on two host chips. This swap allowed copper tracks situated underneath touch-sensitive controls to remain cleanly across a single layer without crossing paths.

When rear routing density created congestion, Claude swapped additional pins and completely routed the entire board corner again rather than stacking via layer transitions. On the voice-assistant hub project, the model flagged copper lines as becoming excessively tangled to trace visually before stripping out traces and rerouting connections dozens of times. While routing near a bottom-port MEMS microphone, the model removed a power track looping around a ground ring and eliminated a via between microphone pads to maintain cleaner local footprints.

03Algorithmic router generation and datasheet analysis

When handling component constraints, Claude read published technical manufacturer datasheets to predict hardware behaviors. Asked whether the handheld virtual pet could operate without a battery installed, the model verified the power management chip's documentation. It calculated that the device would power on via USB-C but correctly deduced that the charging indicator LED would flicker continuously because the internal charger would cycle voltage across a small decoupling capacitor.

Software tool friction prompted the model to write custom design automation software. KiCad's built-in interactive router required step-by-step point guidance that proved too slow for routing long track bundles from a 100-pin connector on the voice-assistant hub.

Claude spent thirty minutes programming a custom grid-based pathfinding router that laid each connection in approximately one second. When noticing an audio clock line running close to Ethernet signals, the model modified its routing code to expand copper clearance around noise-sensitive tracks.

04Mechanical interference detection and acoustic modeling

Spatial planning extended to mechanical CAD assemblies, where Claude drafted enclosures while checking for collisions between physical parts. During the speakerphone enclosure design, the model calculated that the completed board could not slide down vertically into its 3D-printable housing because the mute toggle switch struck the chassis wall. Claude determined the board required a 25-degree tilted entry angle during installation, verified the fit within the 3D model, and wrote the angle into the assembly manual.

For the voice-assistant hub enclosure, Claude analyzed acoustic physical dimensions and computed that a 25 mm microphone port would create a resonance spike near 3.4 kHz. Because that frequency sits within critical human vocal bands, the model noted the design could distort incoming speech before adjusting dimensions. These checks prevented common prototyping oversights where electronic layouts assemble electronically but fail mechanical integration or acoustic performance standards.

05Prototype manufacturing results and validation limits

Diode Computers fabricated the boards, 3D-printed the enclosures, and flashed the firmware generated directly by Claude Fable 5.2. Prototypes for the speakerphone, home assistant hub, and virtual pet all powered up and remained functioning over weeks of lab bench testing. Engineers completed assembly directly from Claude's parts lists, schematics, and mechanical models without making corrective modifications to the printed copper circuitry.

Despite operational success, Anthropic emphasized that the hardware units remain early-stage functional prototypes. The testing did not assess long-term operational durability, environmental degradation, or thermal wear under extended deployment cycles. Furthermore, neither Anthropic nor Diode submitted the devices for formal commercial regulatory certifications, including FCC electromagnetic emissions testing or consumer hardware compliance standards required for commercial production.