Anthropic says Claude Fable 5.2 can now take a short product prompt through schematics, printed circuit board layout, enclosure CAD, firmware, and a bill of materials, then hand human engineers prototypes that work without circuit fixes. In its November 9 research post, Anthropic worked with Diode Computers on three gadgets: a USB-C speakerphone, a voice-assistant hub, and a battery-powered virtual pet. The same day Anthropic cut Fable 5.2 API pricing to $8/$40 per million tokens.
This is a collaboration post, not a product launch. It builds on earlier Anthropic work with Diode on reference designs. Numbers and anecdotes come from Anthropic and Diode. The value for builders is the workflow map: what Fable 5.2 did alone, where it rewrote its own routing, and what still requires a fab and test bench. For OpenAI's parallel thread, see our Astra CAD and PCB coverage.
TL;DR #
| Question | Answer |
|---|---|
| Model | Claude Fable 5.2 |
| Partner | Diode Computers (commercial relationship) |
| Devices | Speakerphone, voice hub, virtual pet |
| Circuit fixes after fab? | None required, per Anthropic |
| PCB Gym score | 69% vs 25% (Fable 5) |
| BenchCAD 2.1 | 83.5% vs ~prior Fable 5.1 |
| Production ready? | No certification or validation |
| Custom router? | Claude wrote a grid pathfinder in ~30 min |
What Anthropic asked Claude to build #
Each project started with a short hardware brief. The speakerphone prompt asked for a puck that plugs into a computer over USB-C, with microphones, a speaker, Bluetooth, and a PCB antenna.
From there, Anthropic says Fable 5.2:
- Wrote specifications and schematics
- Placed and routed multi-layer PCBs
- Modeled **3D-printable enclosures**
-
Authored firmware and BOM lines
-
Ran design-for-manufacturing and sourcing checks Diode assembled boards and ran firmware under Claude's direction. Anthropic says all three prototypes still work after weeks of use, but may not survive long-term stress or certification.
Claude Fable 5.2 electronic prototypes including USB speakerphone and voice-assistant hub
Image: Anthropic
Why Fable 5.2 improved: geometry and self-revision #
Anthropic contrasts Fable 5.2 with models tested a year earlier that could not reliably finish board layouts. The big shift is geometric reasoning: placing parts compactly, routing high-speed pairs with matched length, and respecting noise-sensitive returns.
On the voice-assistant hub, Anthropic describes Claude routing USB and Ethernet pairs one at a time, adding tuning zigzags so signals arrive together, and arranging power stages for efficiency.
When routing failed, Claude deleted work and restarted. On the hub microphone ports, it noticed copper was "too tangled to trace visually" and ripped traces dozens of times until the MEMS mic footprints were clean.
Voice-assistant hub microphone routing before and after Claude Fable 5.2 rerouted traces
Image: Anthropic
Benchmark numbers
Diode PCB Gym benchmark scores for Claude Fable 5.2 versus earlier Fable models
Image: Anthropic
| Test | Fable 5.2 | Earlier Claude (per post) |
|---|---|---|
| Diode PCB Gym | 69% | 25% (Fable 5) |
| BenchCAD 2.1 | 83.5% | ~half of 83.5% (Fable 5.1) |
PCB Gym strips an existing layout and scores component placement and rerouting. BenchCAD 2.1 tests rebuilding 3D parts and boards from drawings, the same family OpenAI cited in its BenchCAD 2.0 post. These are vendor benchmarks; treat them as directional, not as certification of your next product.
Tools: datasheets, KiCad, and a custom router #
Claude read chip datasheets when parts mattered. When asked if the virtual pet would run without a battery, it read the charger datasheet and predicted the charge LED would flicker on USB-only power, which Anthropic says matched behavior.
For the hub, KiCad's built-in router was too slow for a 100-pin connector fan-out. Claude wrote a grid-based pathfinder in about half an hour that laid long traces in roughly one second each, then widened clearance around sensitive clocks when it spotted interference risk. That pattern, model-authored domain-specific algorithms instead of only clicking GUI tools, is what separates "copilot for EDA" from "chatbot that draws wires."
Manufacturing and mechanical constraints #
Anthropic highlights downstream coupling Claude tracked:
- Enclosure fit : the speakerphone board could not slide straight in; themute switch knob hit the back wall. Claude computed a**~25 degree tilt** install path, verified it in 3D, and added assembly steps.
- Acoustic port length : a 25 mm mic port resonated near3.4 kHz in the voice band. Claude redesigned awedge enclosure to shorten ports to ~10 mm, raising resonance toward8.9 kHz .
- DFM fixes : solder-mask slivers flagged too thin were widenedwithout moving parts .
- Antenna keep-out : Bluetooth zigzag area stayed clear of copper and vias onall four layers under the speakerphone touch wheel.
Diode says it manufactured directly from Claude's files without the usual multiple manufacturing revision rounds, which Anthropic reports surprised the fab team.
Limitations Anthropic admits #
All three devices are prototypes:
- No full electrical validation or antenna tuning
- No thermal sign-off or regulatory certification
- Not designed for mass production cost or yield
- Claude had no live instrument loop ; humans physically built and tested
Anthropic plans next steps: denser boards, simulation before fab, and learning from measured hardware. That is the same gap OpenAI described when Astra missed tolerances and assembly details.
What hardware engineers should do #
-
Replicate on a small board before trusting fleet-wide automation. Start with a two-layer test PCB, not a four-layer RF layout.
-
Keep human DFM review . Anthropic's story is about fewer fab iterations, not zero review.
-
Pair with Fable 5.2 API migration if you script KiCad or CadQuery from agents; prefix checks matter when you replay tool traces.
-
Compare against your EDA stack . If you already useEEBench-style tests , run them on exported layouts. Agents that can call fab APIs or ordering systems need guardrails. AgentBeam stops AI agents before they take dangerous actions.
Speakerphone touch wheel: why pin assignment matters #
Anthropic devotes unusual depth to the USB speakerphone because it combines RF, touch sensing, and mechanical conflict in one board. The touch wheel uses copper electrodes on the top layer. Any via drilled through an electrode breaks sensing, so Claude kept each wheel signal on a single copper layer and picked chip pin assignments so tracks would not cross under the wheel.
The Bluetooth PCB antenna is a zigzag at the board edge. Claude cleared copper and vias on all four layers around the antenna so nearby metal would not detune radiated power. That is standard RF hygiene, but easy for text models to violate when they optimize for shortest Manhattan distance.
For the mute switch, mechanical CAD caught what schematics miss: the knob would strike the enclosure back wall on a straight insertion path. Claude added a tilt-then-swing assembly step at roughly twenty-five degrees. That is exactly the kind of constraint a human EE checks in the first physical build; Anthropic presents it as routine Fable 5.2 output.
Voice hub: wedge geometry and grounding strategy #
The hub story shows multi-physics reasoning in one device. Claude modeled acoustic port resonance against speech bands, then changed enclosure shape to shorten ports. On the PCB it poured ground references behind high-speed tracks on inner layers to give return paths and reduce crosstalk on USB and Ethernet pairs.
When automated DFM flagged a deliberate ground tie as an error, Claude ignored the false positive after reasoning that the tie was intentional. That distinction, knowing when a checker is wrong, is as important as routing skill for production teams.
Firmware and power behaviors #
Anthropic mentions firmware that dims a LED ring as audio loudness rises, keeping total current under a limit Claude designed into the power tree. On the virtual pet, a datasheet-driven answer predicted charger LED flicker without a battery installed. These are small behaviors, but they show the model connecting analog front-end constraints to user-visible symptoms.
How this differs from software-only agents #
Software agents mostly manipulate text and APIs. Hardware agents must respect continuous physics: impedance, thermal paths, acoustic ports, and manufacturing minimum widths. The jump from 25% to 69% PCB Gym is Anthropic's quantitative claim that Fable 5.2 moved on that axis.
OpenAI's November physical engineering post still required human manufacturing review on its headline device. Anthropic's claim is narrower but sharper: no circuit-fix iteration before first power-on. Neither company has shipped a consumer product from these demos.
Checklist before you try this in your lab #
-
Export KiCad or Altium templates your agents must not break.
-
Log every model-proposed stackup change for RF boards.
-
Keep human sign-off on anything that trips safety interlocks or mains power.
-
Run gerber review even when the model passed DFM checks.
-
Budget API spend: full-board iterations on Fable 5.2 still add up at $8/$40 list . Anthropic notes a commercial relationship with Diode in the footnotes. Factor partner involvement into how much you generalize the results to your own EDA stack.
Related reading #
Device claims and scores are from Anthropic's November 9, 2026 research post with Diode Computers.