Claude Fable 5 Designs a Working PCB From Scratch, but the Experiment Cost $450

Claude Fable 5 has been used to design a complete custom printed circuit board that was later manufactured and shown working with real hardware. The project suggests that AI can now handle a large part of the PCB design process, although the cost and mistakes involved show that the technology is still far from replacing experienced engineers.

The creator set a strict rule for the experiment. The board would receive no manual design edits or verification before manufacturing. Any design problem that appeared before production had to be handled by the AI.

The goal was to build a compact board around a Raspberry Pi Pico 2350, four physical buttons, a 1.54 inch E ink display, I2C connectivity and additional GPIO access. The PCB also needed to match the physical dimensions of the display.

AI handled the board from schematic to manufacturing

Earlier attempts using other advanced AI models reportedly failed to produce a usable design. Claude Fable 5 made more progress and was able to move through the schematic, component placement and routing stages with the available PCB design tools.

The process was not error free. The AI selected incorrect components at some stages, forcing additional rounds of correction before the final board could be sent for manufacturing.

Despite those problems, five assembled boards were eventually produced for around €130, or roughly $150.

The completed hardware was connected to the intended E ink display and shown operating normally. The interface included functions such as Album, E Reader, Notes and Stopwatch.

Software for the device was created separately using Claude Opus 5. According to the creator, that code worked correctly with the finished board.

The API bill is the biggest limitation

The manufacturing cost was relatively modest compared with the amount spent on AI access.

From the beginning of the project to the final working design, around $450 was spent on API credits. That means the AI processing cost was roughly three times the price paid to manufacture five assembled boards.

For a hobby project, that expense may still be acceptable to someone experimenting with new tools. It becomes harder to justify if the same approach is used repeatedly or for larger commercial projects.

The experiment also does not mean PCB expertise is no longer necessary. A trained engineer is more likely to notice problems involving electrical safety, signal integrity, component tolerances, thermal behaviour and manufacturing requirements before money is spent on physical boards.

What this project demonstrates is that AI assisted hardware design is becoming practical enough to produce working devices. The technology can already perform tasks that previously required significant manual work inside PCB design software.

Cost, reliability and verification remain major limits. As AI models improve and inference prices fall, similar workflows could become more useful for rapid prototypes, personal electronics projects and early hardware development.