What happened
In a November 9, 2026 research post, Anthropic describes a collaboration with Diode Computers, a company that designs and manufactures circuit boards. Anthropic says it began testing Claude on basic electronics design more than a year earlier, when the models struggled to lay out boards reliably and to design components that fit together. In the latest experiment, the team asked Claude Fable 5.2 to design a speakerphone, a home voice-assistant hub, and a handheld virtual pet, using short prompts about each device’s purpose and main hardware requirements.
For the speakerphone, the prompt called for a puck that connects to a computer over USB-C and includes microphones, a speaker, Bluetooth, and a PCB antenna. Claude wrote specifications, created schematics, selected and placed components, routed copper connections, designed a 3D-printable enclosure, wrote firmware, and prepared a parts list. It used CAD software to model the designs and checked component fit. For example, the speakerphone’s mute switch would hit the enclosure wall if the board were lowered straight in; Claude worked out that the board could be tilted by about 25 degrees during installation and added that step to assembly instructions.
Anthropic says Claude also reasoned about electrical and manufacturing constraints. For the voice-assistant hub, it routed USB and Ethernet differential pairs while keeping their tracks matched in length and spacing, and wrote its own routing algorithm when the built-in KiCad router would have required too much step-by-step guidance for long connections. It adjusted the router to leave more room around sensitive tracks. It also ran design-for-manufacturing and sourcing checks, fixed six solder-mask slivers flagged as too thin, and judged one reported ground connection to be an intentional design choice. Diode manufactured all three boards from Claude’s files without requesting adjustments.
The prototypes continued to work after assembly, and Anthropic says the speakerphone remained in use for weeks. On Diode’s internal PCB Gym benchmark, which tests component placement and rerouting on boards with layouts removed, Fable 5.2 scored 69%, compared with 25% for Fable 5. On the public BenchCAD 2.1 benchmark, which tests reconstruction of 3D parts, assemblies, and boards from engineering drawings and rendered images, it scored 83.5%, nearly double Claude Fable 5.1’s result. The post does not describe a commercial product or an access change; it reports a research collaboration and prototype work.
What the archive adds
New Runtime’s earlier coverage includes Claude Fable 5.2 support in llm-anthropic 0.31 on November 10, 2026, and Claude Fable 5.2 and Claude Mythos 5.2 on November 9. Those notes covered model availability and formalization; this is a new fact about Fable 5.2 applied to electronics design, rather than a restatement.
Trend signal
This supports harness architecture outlives model choice only weakly: the demonstration depends on Claude using CAD, KiCad routing, datasheets, custom code, and manufacturing checks, not on model reasoning in isolation. It is evidence that the surrounding tools and workflows are central to extending a model into a practical engineering task, though one vendor-led prototype exercise cannot establish a durable product boundary.
Why now
Anthropic frames the work as a comparison with tests begun more than a year earlier, when its models struggled with board layout and component fit. The current experiment is also positioned alongside recent work by OpenAI on designing a handheld device, according to Anthropic. The immediate timing is therefore a reported capability progression, not a newly announced hardware product.
Connections to other events
Anthropic says OpenAI recently wrote about using its models to design a handheld device. The connection is a shared direction—models contributing to physical hardware design—but the post does not provide a direct comparison between the two efforts. New Runtime’s earlier coverage of Fable 5.2’s release and formalization work provides model context, while this experiment tests a different task: coordinating design files and constraints that must survive physical assembly.
Why it matters
For teams building with AI, the important result is not simply that a model generated schematics or firmware. The reported workflow crossed several boundaries: interpreting component datasheets, laying out and routing a board, checking whether parts fit inside an enclosure, adjusting designs when routing became tangled, and accounting for factory limits. These are coupled tasks where a locally plausible answer can fail downstream.
The demonstration suggests a practical way to evaluate engineering agents: judge whether they can revise a plan when constraints collide and produce artifacts that people can build, not just whether they can explain a design. Yet Diode’s engineers still assembled and tested the boards, and the devices have not passed electrical validation, antenna tuning, thermal tests, or certification. Teams should treat this as evidence of useful prototyping assistance, not a substitute for engineering review or product qualification.
Limits and uncertainty
Anthropic’s account is a company-reported collaboration with Diode, with which it has a commercial relationship. The three devices are prototypes, not mass-production designs. Anthropic says they have not undergone electrical validation, antenna tuning, thermal testing, or certification, and that engineers had to build and physically test them because Claude was not connected to test equipment or board-manufacturing machines. The post does not establish how the work compares with unaided engineers on time, cost, reliability, or repeated designs. The benchmark figures are reported by Anthropic and Diode; independent replication and performance on denser, faster, more complex boards remain open questions.
Skill to build now
This week, pick one bounded hardware task and make downstream constraints explicit in the brief: component availability, board geometry, enclosure fit, manufacturing limits, and the checks that would count as passing. Ask an AI system to produce inspectable artifacts and a revision log, then have an engineer review each critical decision. The experiment’s useful lesson is to test whether the system can detect and correct conflicts, not merely produce a first draft.
Reader takeaway
Claude Fable 5.2 reportedly helped take three electronics designs from short prompts to working prototypes, including board layouts, enclosures, and firmware. That is a meaningful expansion of AI-assisted engineering, but not proof of production readiness: human engineers still built and tested the devices, and the prototypes remain unvalidated. The useful benchmark is whether an AI workflow can make artifacts that survive real constraints and expert review.