Claude Fable 5.2 designs three electronic prototypes with Diode | Magica
Anthropic says Claude Fable 5.2 designed three working devices with Diode
Editorial Team
••📖7 min read
Anthropic says its new AI model produced circuit boards, enclosures and firmware for three prototypes that needed no electrical fixes. The demonstration avoided factory revisions, according to the company, while leaving certification, manufacturing scale and total costs unproven.
Anthropic says all three Claude-designed prototypes worked without electrical modifications after Diode built them.
The model handled circuit layouts, enclosures and firmware; human engineers manufactured, assembled and tested the hardware.
These are prototypes; certification, durability and total development costs remain unproven.
AI developer Anthropic said on November 9 that Claude Fable 5.2, its newly released model for demanding technical work, designed a USB speakerphone, a home voice-assistant hub and a handheld virtual pet that worked without repairs to their circuitry. Diode Computers, a circuit-board design and manufacturing company that has worked with Anthropic since 2025, physically built the prototypes.
The practical advance claimed in Anthropic’s research account is that the designs reached fabrication without the usual exchanges over manufacturing adjustments. Avoiding that back-and-forth could shorten development, but three working prototypes do not establish reliability or a repeatable saving. Diode is a commercial partner that helped Anthropic improve Claude’s electronics skills, so this is a demonstration by collaborators, rather than an independent product evaluation.
Diode built these three prototypes from Claude’s designs. Their operation without electrical modifications is company-reported; certification and long-term durability remain unproven. Image: Anthropic. Source: Anthropic.
From a short prompt to a manufacturable board
Each project began with a short description of the device and its main requirements. For the speakerphone, that meant a USB-C connection, microphones, a speaker and Bluetooth with an antenna on the circuit board.
Anthropic says Claude wrote specifications, selected components, drew schematics, positioned parts and routed the copper connections on the printed circuit boards. It also modeled 3D-printable enclosures, wrote firmware and prepared parts lists. According to Anthropic’s model announcement, each four-layer board was ready for manufacturing in under a day, using available components from US distributors.
Diode then manufactured the boards and followed Claude’s guidance for testing and assembly. Anthropic says human prompting was minimal, including short instructions to clean up schematics or continue. The speakerphone remained functional after weeks of use, and the other two devices were also still working when the company reported the results.
UC Berkeley and NVIDIA are releasing tools for AI agents to assemble and improve robot programs in simulation. A warehouse operator reports a useful production update, while unsuccessful fixes and benchmark results show why testing, speed and existing infrastructure still matter.
Editorial Team
Diode’s contribution predates this demonstration. Anthropic’s December 2025 account described a joint effort to improve chip reference designs—the supporting circuitry needed to make a chip operate. Engineers were already using Claude Code, Anthropic’s coding agent, to generate these designs from documentation before review. Diode develops Zener, a language for describing circuit-board schematics, and software that automates work in the KiCad board-design program.
Anthropic said the collaboration informed training improvements to its Sonnet 4.5 model and subsequent Claude models. The new test extends that work across board layout, enclosures and firmware.
The improvement was in revising the design
Anthropic identifies geometric reasoning as the biggest improvement over its earlier tests. Previous models misplaced supporting components, oriented connectors incorrectly and often failed to reconsider layouts that could not be completed. Anthropic says Diode had produced many boards with those models that were 95% complete but unusable.
Fable 5.2 instead revised its own work. On the speakerphone, it changed chip pin assignments so connections beneath touch controls could remain on one copper layer without crossing. On the voice-assistant hub, it repeatedly deleted and rerouted tangled connections.
The model also changed tools when necessary. Diode’s software exposed the router in KiCad, a circuit-board design program, but drawing long connections individually from a 100-pin connector was cumbersome. Anthropic says Claude wrote a grid-based routing algorithm in about half an hour that could lay a long track in roughly a second, then adjusted it to protect sensitive audio signals from nearby Ethernet connections.
Manufacturing checks supplied another constraint. Claude widened openings in the hub’s solder mask, the protective coating on the board, to eliminate six strips too thin to manufacture. It also rejected a false warning about a deliberate ground connection. The firmware also reflected the hardware’s power budget: the speakerphone dims its LED ring as audio gets louder to stay within its designed current limit.
Company-reported routing revision in the hub: the later layout leaves the microphone’s ground ring clear and grounded. Graphic: Anthropic. Source: Anthropic.
Better benchmark scores, incomplete economics
Anthropic’s research account reports a score of 69% for Fable 5.2 on Diode’s internal PCB Gym benchmark, compared with 25% for Fable 5. That test asks models to place components and reroute connections on boards whose layouts have been removed. It measures a defined design task, rather than the probability that a new consumer device will pass certification.
The company also reports stronger results on the public BenchCAD 2.1 benchmark, which tests reconstruction of parts, assemblies and circuit boards from drawings and images. Its launch account describes the measure as a 0–100 score with partial credit, rather than a count of fully successful devices.
Model
Reported BenchCAD 2.1 score, with tools at Max effort
Claude Fable 5.2
83.5
Claude Opus 5.5
65.3
Claude Fable 5.1
44.4
Those comparisons support Anthropic’s claim of progress over its own models. They do not establish lower end-to-end engineering costs. Fable 5.2’s standard API price is $8 per million input tokens and $40 per million output tokens. Both are 20% below Fable 5.1’s respective $10 and $50 rates, which the launch report attributes to Bloomberg. Anthropic’s announcement lists an optional fast mode at twice the standard rates: $16 per million input tokens and $80 per million output tokens, for claimed token generation up to 2.5 times faster. Neither rate reveals the cost of building these devices: the research account provides no total token bill, parts cost or full human-labor accounting.
Fable is also not Anthropic’s default recommendation for every engineering-related coding task. The company advises users to begin most coding and knowledge work with its Opus 5.5 model and move to Fable for more ambitious technical problems. A hardware report’s price comparison lists Opus at $4 per million input tokens and $20 per million output tokens, half Fable’s standard rates. How many tokens each model needs to finish a task still matters. Anthropic says Fable 5.2 is available through its own platform and Amazon, Google and Microsoft cloud services.
The launch has a financing backdrop, too. The launch report relays Bloomberg’s account that Anthropic’s stock-market listing could come as soon as November.
Nor is this the first reported working AI-designed board. A report on the demonstration and its precedents describes an independent designer’s working Fable 5 board in September, after substantial design corrections. It also relays OpenAI’s account of a November 3 experiment with its GPT-6 Astra model: a handheld device for talking to its Codex coding agent required multiple manufacturing revisions. These examples used different projects and workflows; neither the earlier independent designer’s result nor the rival companies’ demonstrations provide a controlled comparison of engineering productivity.
Company-reported BenchCAD 2.1 scores with tools at Max effort, on a partial-credit scale of 0–100. These results do not establish certification readiness or total development costs. Chart: Magica; data: Anthropic. Source: Anthropic.
The next test is beyond a working prototype
The three devices were not designed for mass production. Anthropic says they have yet to undergo electrical validation, antenna tuning, thermal testing or certification, and their longer-term durability is unknown. Claude was not connected to measurement instruments or board-making and assembly machines, leaving the physical work with Diode engineers.
Access to the model also does not reproduce the entire demonstrated workflow. The hardware report says Diode’s software interface to KiCad was not released. That limits what an engineer can infer from simply having Fable 5.2 and a circuit-design program available.
Anthropic’s stated next step is to test denser, faster and more complex boards, use simulation, and learn from measurements on finished devices. The evidence still needed is whether the reported absence of manufacturing revisions holds for more complex devices—and whether the combined model, fabrication and engineering bill improves on an ordinary development process.
Cloud Pathways lets customers coordinate large TPU workloads from one Python client, extending the system Google used for Gemini training. General availability on GKE follows earlier customer access; substantial reported usage and internal recovery results still leave customer cost savings unmeasured.
NVIDIA’s GR00T 2 predicts video and robot actions together to transfer skills across bodies. Company tests show gains from human video, but selected tasks, corrective data and recovery agents qualify its strongest success rates.
Meta is reportedly building a consulting channel for corporate AI sales after announcing an enterprise platform led by former MongoDB CEO Chirantan Desai. Its hiring remit also covers advertising and messaging, while prospective partners already work with rival AI suppliers.
OpenAI’s $200 Chromatic: Codex Edition adds AI-assisted development to an existing cartridge-game platform. The public bundle includes a programmable cartridge, but creation requires a computer, established tools and playtesting; the store now lists a November 19 restock.
PwC US plans a December launch for a critical-thinking and judgment certificate, with completion targeted by June 2027. It will assess staff through performance reviews as the firm seeks new ways to develop skills once practiced through routine work.
Linnea Quigley and Beverly Randolph say Aftermath departed from the musical they agreed to make, exposing a gap between a marketed cast reunion and performers’ control over their screen identities.
Anthropic’s Claude Fable 5.2 and Mythos 5.2 bring lower token prices and stronger technical performance. The release also limits capabilities that could accelerate frontier AI development, as new evidence puts the pace of research close to a sensitive policy boundary.
Meta’s smallest Muse model matches Claude Haiku 5.5’s rounded intelligence score at about one-third of the benchmark task cost. The case for delegating agent work depends on caching, task difficulty and higher charges for long requests.
SpaceX proposes a gas-fired power plant and super-compute facility on about 444 acres in Starbase, Texas. A December hearing will consider local incentives, while the public notice leaves the tax concession and operating demands unspecified.
The California state senator elected to Nancy Pelosi’s open seat wants a federal AI regulator, independent safety checks and penalties for developers. His California record shows both the reach and limits of that agenda; political support from an Anthropic-funded group adds scrutiny.
Apple Creator Studio subscribers can buy extra Logic Pro assistant usage for $49.99 in the U.S. The optional packs last up to a year, but Apple’s public support document does not quantify their capacity.
Boston Dynamics has fitted its electric Atlas humanoid with scooter wheels to test learned controls against real motor limits. The research feeds into the walking robot as Hyundai plans factory deployment, but the demonstration offers no quantified reliability gains.
Nominal raised $240 million at a $3 billion valuation, three times its March valuation, to expand its hardware engineering software into AI agents and lifecycle applications. Its defense foothold and company-reported workflow gains provide a foundation, but do not establish the newer agents’ performance or economics.
The former Megaspeed executive sought backing for a data centre project worth several hundred million dollars, citing GPU access and Alibaba business. The reported pitch adds commercial detail to scrutiny of her ties to Nvidia’s chief, while espionage and chip-smuggling allegations remain unproven.
President Donald Trump offered an unnamed executive’s alleged reversal as evidence that AI risks can be managed. His account identifies neither the conduct at issue nor the legal basis for a jail threat, leaving its value as a safety precedent unproven.
The Texas startup is building factory-made data centers for AI customers, with testing reportedly planned for early 2027. Its financing backs a manufacturing strategy whose cost and power-access advantages remain to be demonstrated.
Baidu says its automated research system has improved inference for the forthcoming ERNIE-5.5-Flash, alongside public benchmark gains. The results show bounded engineering progress, while the cost of finding those improvements remains undisclosed.
Nvidia’s shares fell as other chip stocks rebounded after its board opened an independent inquiry into reports of a foreign intelligence effort targeting CEO and co-founder Jensen Huang. The board retains confidence in him and says it has no reason to believe secrets were compromised; China rejects the spying allegation.
