Can an AI system take an electronics idea from a short brief to a device that actually works? Anthropic says Claude Fable 5.2 helped design three such prototypes with circuit-board company Diode: a USB-C speakerphone, a voice-assistant hub and a handheld virtual pet. Diode's engineers built the hardware and ran the firmware. The results suggest a model can contribute across a complicated design workflow, while leaving important questions about safety, durability and commercial readiness unanswered.
What did Claude do in the electronics project?
Anthropic says each project began with a short description of the device and its main requirements. Claude then wrote specifications, selected components, produced circuit schematics, laid out and routed four-layer printed circuit boards, modeled enclosures in CAD, wrote firmware and prepared parts lists. It also revised designs when it found a problem. Diode's engineers assembled the boards and ran the firmware under Claude's direction. Anthropic says the prototypes worked without electrical changes to their circuit boards.
The speakerphone example shows why this is more than drawing a circuit diagram. Its board had to fit a small enclosure with microphones, a speaker, USB-C and Bluetooth. Anthropic describes Claude routing high-speed USB and Ethernet pairs with matched lengths and spacing, accounting for the power supply, and figuring out that the board had to be tilted into the enclosure so a mute-switch knob would clear the case. The task links electrical layout to mechanical fit and assembly instructions.
What do the reported benchmark scores mean?
Anthropic also reports a 69% score for Fable 5.2 on Diode's internal PCB Gym benchmark, compared with 25% for Fable 5. The test removes layouts from existing boards and scores how well a model places components and routes connections. On the public BenchCAD 2.1 benchmark, Anthropic reports 83.5%, nearly twice the score of Fable 5.1. These numbers describe specific engineering tasks. They do not measure how often a new design will work in every setting, pass safety standards or survive years of use. The report is from Anthropic and its project collaborator; VIBPS has not independently reproduced the tests.
The model's reported improvement was especially visible in geometric decisions: where parts sit, whether traces can be routed, and how the board fits with nearby components. Anthropic says earlier models sometimes left boards nearly complete but unusable. Fable 5.2 was more likely to change its approach when it could not finish a layout. That is a meaningful design behavior, though it still needs engineers to inspect the files and build the result.
Why are these still prototypes?
Anthropic is explicit that all three devices are early prototypes. They have not gone through electrical validation, antenna tuning, thermal testing or commercial certification, and they were not designed for mass production. Claude was not connected to the lab instruments or factory machines; people at Diode physically made and tested the devices. A prototype that powers on and works in a partner's test is useful evidence of progress, but it is not the same as a product ready for customers.
The distinction matters for makers because a hardware product has to survive more than a successful first build. Components must remain available, boards must be manufacturable at scale, radios and power systems must meet requirements, and enclosures need to tolerate heat and repeated use. Anthropic says the speakerphone continued working after weeks of use, but that observation does not replace the broader tests it lists as outstanding.
What can software and hardware makers take from it?
The project points to a practical use for AI in engineering: coordinating several artifacts that have to agree, from requirements and parts lists to board files, enclosure geometry and firmware. It also shows where human review stays central. Engineers supplied physical measurements and manufacturing knowledge, assembled the boards, and checked behavior on real devices. For teams experimenting with AI-generated code or companion apps, the same habit applies: treat generated work as a testable draft, and verify it in the environment people will actually use.
How to test an AI-built app before inviting users
Anthropic's demonstration is a report on three collaborations and selected benchmark tasks, not an independent evaluation of AI hardware design as a whole. It gives a concrete view of what Fable 5.2 contributed and what engineers still had to do. The next milestone is not just producing a plausible board file; it is showing that designs can pass the physical checks, certifications and manufacturing constraints that turn a prototype into a dependable product.