Designing an electronic device means designing and simulating a circuit, laying out the board, modeling the enclosure, writing firmware, and checking that parts are in stock and that a factory can build it. This post, published on Anthropic's Science blog on 9 November 2026, the date of the Claude Fable 5.2 system card, reports that Fable 5.2 carried three devices through all of these steps with little prompting, and that Diode Computers, a company that designs and manufactures circuit boards, built working prototypes from its files. The post is a case study with two benchmark results. Its acknowledgments describe it as a collaboration between Anthropic researchers and engineers, "including Mike Cretella, Connor Jennings, and Brent Yi, and Craig Chen and the team at Diode Computers." Footnote 1 states: "Anthropic has a commercial relationship with Diode." The post's benchmark numbers also appear, with their settings, in the Fable 5.2 system card (BenchCAD) or come from Diode's internal evaluation (PCB Gym).
What was done
The earlier work
The post dates the collaboration to a little over a year before the post, when Anthropic and Diode began training Claude on simple circuit tasks such as reading a chip's datasheet and working out which supporting parts it needs. An article on claude.com dated 12 December 2025, by Diode and Anthropic staff, describes that work: Claude wrote reference designs for chips in Diode's Zener language, and Diode's electrical engineers preferred Sonnet 4.5's designs to Sonnet 4's in 82% of blind head-to-head comparisons. According to the post, the models of that time "couldn't reliably lay out a circuit board" and struggled to design components that fit together.
The three devices and the prompts
Fable 5.2 was asked to design a speakerphone, a voice-assistant hub and a handheld virtual pet. Each began with a short prompt giving the kind of device and its main hardware requirements. For the speakerphone the prompt asked for "a puck that plugs into a computer over USB-C and includes microphones and a speaker, supporting Bluetooth with a PCB antenna." Claude then wrote the specification and drew the schematics. It placed the components on a printed circuit board (PCB) and routed the copper connections between them. It also designed a 3D-printable enclosure, wrote the firmware and prepared the parts list. Diode assembled the boards and ran the firmware "under Claude's direction."
Footnote 2 describes the human prompting during the work as minimal and "limited to short directives (e.g., 'clean up the schematic' and 'continue')." One exception was a question from the Diode team, "Will it work if there is no battery?", which Claude answered from the battery charger's datasheet. It predicted that the pet would run on USB power but that the charging light would flicker.
What it reports
The prototypes
According to the post's introduction, Claude directed human engineers to build prototypes that "worked without any fixes to their circuitry." Diode manufactured all three boards directly from Claude's design files, "without any back-and-forth over adjustments" (section "Designing for real-world manufacturing"). The Diode team, the post says, was surprised not to need the usual rounds of manufacturing changes. The speakerphone still worked after weeks of use, and the other two devices are "still in good working order." The post adds that, as early prototypes, it is not yet known how they will hold up or whether they would pass commercial certification.
Layout and revision
The post attributes the main improvement over earlier models to geometric reasoning about where to place parts and how to route connections (section "Reasoning through geometric problems"). Its examples:
- Routing order. On the voice-assistant hub, Claude routed the high-speed USB and Ethernet pairs first. It kept each pair's two tracks equal in length and correctly spaced, and added small zigzag "tuning patterns" so that signals arrive together.
- Self-revision. On the speakerphone's four-layer board, Claude reassigned signals to different chip pins so that the tracks under the touch controls could stay on one layer without crossing. Later it re-laid a crowded corner rather than adding extra jumps between layers. On the hub, it noted the connections were "getting too tangled to trace visually," wrote "I should just rip these traces [tracks] and start fresh," and rerouted them dozens of times (section "Revising designs"). The post contrasts this with earlier models, which, it says, rarely revised their layouts; Diode had designed many boards with them that were "95% complete but still unusable."
- Its own router. When KiCad's built-in router, which drew one track at a time, proved too slow for long tracks from a 100-pin connector, Claude wrote a grid-search router in about half an hour that laid each long track in about a second (section "Using software and documentation").
Mechanical and manufacturing checks
Claude wrote code to check that each board fit its enclosure. It found that the speakerphone's board would fit only if tilted about 25 degrees during installation, and added that step to the assembly instructions. For the hub, it calculated that a 25 mm microphone port would resonate near 3.4 kHz, "right in the voice band." It changed the enclosure to a wedge shape that shortened the ports to about 10 mm and raised the resonance to about 8.9 kHz (section "Keeping track of constraints and downstream effects").
Claude also ran design-for-manufacturing checks, which compare a board's geometry with a factory's limits, and sourcing checks on part availability. On the hub, it fixed six solder-mask slivers flagged as too thin. It also set aside a flagged connection between two grounds that it judged deliberate; the post says it "correctly reasoned" that the connection would not affect performance.
Benchmarks
The post gives two benchmark results (section "Reasoning through geometric problems"):
- PCB Gym, described as "an internal benchmark" run by Diode, scores how well a model places components and routes connections on boards whose layouts have been removed. Fable 5.2 scored 69% and Claude Fable 5 25%. The chart, captioned "Source: Diode's internal PCB layout evaluation," plots score against total API cost, at maximum reasoning effort unless stated. Read from the chart, which prints no values: Opus 5.5 about 58%, Fable 5.1 and GPT-6 Astra about 49%, Sonnet 5.5 about 44%, and Opus 5 about 34%. Fable 5.2 at maximum effort cost about $2,200 in total, against about $6,800 for Fable 5; two further Fable 5.2 points, at low and medium effort, sit at about 58% and 61% for roughly $700 to $800 (also read from the chart). The post does not say how many boards PCB Gym contains or how it scores them.
- BenchCAD 2.1, a public benchmark of rebuilding parts, assemblies and circuit boards from drawings and images, scored by executed geometry with no model judge (repository README). The post reports 83.5% for Fable 5.2, "nearly double the score of its predecessor, Claude Fable 5.1." The system card gives the settings (Section 8.13.2, pp. 232 to 233). There, 83.5% is the score with tools (a shell, file editor and image zoom, without limits on rounds or tokens), against 44.4% for Fable 5.1. Without tools, using Anthropic's own implementation of the benchmark protocol, the scores were 75.4% and 36.3%. Each score is a mean over five runs of the 100-problem Core set. The BenchCAD leaderboard lists these numbers as "credit to Anthropic," taken from the system card.
Who checked the results
- The prototypes were built and tested by Diode engineers, because Claude was not connected to test equipment or to fabrication machines (section "Limitations and next steps"). The tests reported are that the devices work and keep working; the post lists electrical validation, antenna tuning, thermal testing and certification as not done.
- PCB Gym is Diode's internal evaluation, not public, run by a partner with a commercial relationship with Anthropic.
- BenchCAD 2.1 is public and its scorer is open, but the Fable 5.2 scores were run by Anthropic and reported in its system card; the leaderboard shows no independent run of Fable 5.2.
- Selection. The post describes three devices and does not say whether other designs were attempted.
The prompts and instructions used
The post quotes the speakerphone's opening prompt and the short directives in footnote 2, and reports that Diode's prompting was minimal. It does not publish the full prompts, the tools and design software given to the model beyond KiCad and Diode's software, or any system prompt. It contains no comparison of different instructions.
Limitations and open questions
The post's own limitations section says the devices are prototypes. They have not been through electrical validation, antenna tuning, thermal testing or certification, and they were not designed for mass production. Engineers had to build and test them physically. The post plans next to test denser and faster boards, checks in simulation, and learning from measurements on finished devices.
Further points a reader may need:
- The BenchCAD comparison is with tools. The 83.5% and the "nearly double" comparison are with-tools scores; without tools the system card reports 75.4% against 36.3%. The post does not say which setting it quotes.
- Different predecessors. The PCB Gym comparison is with Fable 5 (25%); Fable 5.1, also on the chart, scored about 49% (read from the chart).
- A partner's benchmark. PCB Gym's tasks, size and scoring are not published.
- Outside comparison. The post notes that "OpenAI recently wrote about using its models to design a handheld device." That OpenAI post (3 November 2026, "Accelerating engineering for the physical world") reports a newer internal model at 0.875 on BenchCAD 2.0, against 0.594 for GPT-6 Astra, and says its own device "still required several rounds of review and iteration" before manufacture. The BenchCAD leaderboard lists the 0.875 on a separate board for customized harnesses (Codex, 2.0 scorer) and says scores from different harnesses are not comparable. Setting the two device accounts side by side is this note's own.
Research questions
- How do the prototypes perform under electrical validation? Measurements of signal integrity, antenna performance and thermal behavior on the three boards would test whether "worked without any fixes" holds beyond functional use.
- What is the success rate across designs? Reporting every device attempted, with how many reached a working prototype, would show how typical these three are.
- What does PCB Gym measure? Publishing its task count, board types and scoring would let others interpret the 69% and the cost axis.
- How much did Diode's engineers contribute? A log of every human message and every physical intervention during assembly would make "minimal prompting" measurable.
- Would the designs pass a human design review? Blind review of Claude's boards and comparable human-designed boards by electrical engineers, as in the 2025 reference-design study, would compare quality as well as function.
How it connects
- The Claude Fable 5.2 and Mythos 5.2 system card reports the BenchCAD 2.1 scores with their settings (Section 8.13.2); this post quotes the with-tools score.
- Introducing our Science Blog set out the blog; Long-running Claude for scientific computing is another Science-blog case of Claude carrying a technical project over many steps with occasional human input. Both connections are this note's own.
References
- Anthropic. (2026, November 9). Claude can design electronic prototypes from start to finish. https://www.anthropic.com/research/claude-electronics-design record
- Anthropic. (2026, November 9). System Card: Claude Fable 5.2 & Claude Mythos 5.2, Section 8.13.2 (BenchCAD 2.1), pp. 232 to 233. https://www.anthropic.com/claude-fable-5-2-mythos-5-2-system-card record · 2 pages
- Asnaghi, D., Khazan, L., Jennings, C., Hershey, D., and Marwell, N. (2025, December 12). Making Claude a better electrical engineer. Claude. https://claude.com/resources/articles/making-claude-a-better-electrical-engineer record
- BenchCAD Team. BenchCAD 2.1 release notes and README (tag v2.1). https://github.com/BenchCAD-org/benchcad-2-main/releases/tag/v2.1 record
- BenchCAD leaderboard, read 13 November 2026. https://benchcad.com/leaderboard.html record
- OpenAI. (2026, November 3). Accelerating engineering for the physical world. https://openai.com/index/accelerating-engineering-for-the-physical-world/ record