AI News
AI News AgentResearchAnthropic4 min read

Claude designs electronic prototypes from start to finish

Anthropic says Claude Fable 5.2 can design complete electronic prototypes, from the board and enclosure to the firmware and parts list. Three devices fabricated by Diode worked without changes to their circuits, although they are not yet ready for production or certification.

Claude can now design and help build complete electronic devices: from the electrical schematic and circuit board to the enclosure, firmware, and bill of materials. Anthropic tested this capability with Diode Computers on three prototypes: a speaker with microphones, a voice assistant hub, and a wearable virtual pet.

The advance matters because creating hardware is not just about connecting parts. You also have to decide where to place each component, route the copper connections, check that everything fits inside the enclosure, write the device software, and make sure a factory can produce it.

Claude Fable 5.2 handled much of that work with minimal instructions. Diode fabricated the circuits from its files, and all three prototypes worked without requiring changes to their electrical connections.

From circuit to functional device

To create the speaker, for example, Claude received a simple description: a disk-shaped device that connected over USB-C, included microphones and a speaker, and had Bluetooth with an antenna integrated into the board.

From there, the model:

  • Wrote the device specifications.
  • Drew the electrical schematics.
  • Chose the position of the components on the printed circuit board, known as a PCB.
  • Routed the connections between the components.
  • Designed a 3D-printable enclosure.
  • Programmed the firmware, the software that controls the hardware.
  • Prepared the list of required parts.

Diode engineers then fabricated the board and ran the firmware following Claude's instructions. The speaker was still working after several weeks of use.

Improvements in board design

The main difference compared with earlier versions is spatial reasoning. Designing a PCB requires placing parts close to the components they support, orienting connectors correctly, and routing copper traces without creating crossings or interference.

Claude Fable 5.2 started with the most delicate connections, such as those carrying fast signals, high currents, or noise-sensitive data. In the voice assistant hub, for example, it routed the USB and Ethernet connections first, kept the traces for each pair together and the same length, and added small zigzags to synchronize when the signals arrived.

In an internal Diode test called PCB Gym, the model scored 69%, compared with 25% for Claude Fable 5. In BenchCAD 2.1, a public test that evaluates the reconstruction of parts, assemblies, and boards from technical drawings and images, it reached 83.5%, almost twice the score of Claude Fable 5.1.

It was also able to correct its own work. When a layout left too little space to finish the connections, Claude switched signals between interchangeable pins, rebuilt entire sections of the board, and rerouted the traces. At one point while designing the voice hub, it even deleted and redesigned the connections dozens of times until it found a valid layout.

It does more than draw: it checks the consequences

Claude consulted component datasheets, the technical documents provided by their manufacturers, to make specific decisions. When analyzing the virtual pet, it concluded that the device would work over USB even without a battery, although the charging light would probably blink.

It also considered the physical design. It detected that the speaker's board would hit the back wall of the enclosure during installation. The solution was to tilt it by about 25 degrees during assembly. In the voice hub, it modified the shape of the enclosure to reduce resonance from the microphone ports and improve their performance in the voice frequency range.

To solve a routing problem, Claude wrote its own router: an algorithm that searches for open paths on a grid to connect the different points on the board. The program could route some long connections in approximately one second instead of requiring step-by-step instructions.

What changes for you

This does not mean you can ask Claude for a product and receive one ready to sell. It does mean an idea can reach a first physical version sooner, with less manual work in the early stages.

A small team could use a model like this to quickly create a controller, sensor, or connected device and identify problems before manufacturing many units. AI handles more technical decisions, while engineers are still needed to build, measure, and validate the result.

The three devices are still prototypes. They have not undergone complete electrical testing, thermal validation, antenna tuning, or certification, and they were not designed for mass production. Claude also was not directly connected to measurement equipment or manufacturing machines.

The next step will be to see whether it can design denser, faster, and more complex boards, simulate how they work, and learn from measurements taken from finished devices. The important signal is not that AI will suddenly replace engineers, but that it can already take on a much larger part of the journey from an idea to a working prototype.