Claude Fable 5.2 and Mythos 5.2 for PCB Design and Robotics
Updated 10 November 2026. Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026. They are the same underlying model. Fable 5.2 is the public model, with cyber and biology safeguards. Mythos 5.2 is the same capability set for organizations in Anthropic’s Cyber and Life Sciences verification programs.
For hardware teams, the useful news is narrower than the launch headlines. Fable 5.2 designed three working electronic prototypes with Diode, scored 69 percent on Diode’s internal PCB Gym, and scored 83.5 on public BenchCAD 2.1. On RoboDojo Sim, Anthropic’s own test runs put Fable 5.2 at 64.9%, ahead of Opus 5.5 and GPT-6 Astra on that harness. None of those results replace schematic review, electrical test, or a factory DFM gate.
In short
-
Same model, two doors. API IDs are
claude-fable-5-2andclaude-mythos-5-2. Context is 1 million tokens. Max output is 128K. List price is $8 per million input tokens and $40 per million output tokens. Fast mode is $16 and $80, up to 2.5 times token speed. - PCB evidence is a prototype study, not a volume NPI. Three four-layer boards (USB speakerphone, voice-assistant hub, handheld virtual pet) were built from Claude’s files. Diode reports no electrical rework. Anthropic also says the boards skipped electrical validation, antenna tuning, thermal test, certification, and mass-production design.
- Robotics evidence splits in two. A short official clip shows off-the-shelf arms driven through Claude Code. The 64.9% number is RoboDojo Sim, a camera-to-action simulation, not a count of those physical demos.
- Use Fable when the task is open-ended hardware or long agent work. Anthropic still points daily coding at Opus 5.5 and the KiCad MCP stacks we compared earlier.
What shipped on 9 November 2026
Anthropic’s launch note says Fable 5.2 and Mythos 5.2 generate tokens up to 30% faster than the 5.1 pair, at a lower list price. CNBC reported the price as about 20% below the 5.1 generation. Fast mode is optional and costs twice the standard rate.
Fable 5.2 is on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Mythos 5.2 uses the same platforms, but only after verification. Anthropic’s docs say Fable 5.2 needs 30-day data retention unless Anthropic authorizes zero data retention. Enterprise Frontier Safeguards is the path they describe for customers who want monitoring inside their own infrastructure.
Anthropic’s own guidance on model choice is practical. Start most coding on Opus 5.5. Move up to Fable or Mythos when the job is ambiguous, long, and technical. Move down to Sonnet 5.5 or Haiku 5.5 when the task is routine. Fable can also sit as an advisor that smaller models call, or as an orchestrator that delegates.
Claude Fable 5.2 and Claude Mythos 5.2 are available now:
9 November 2026
Today Anthropic showed Claude Fable 5.2 controlling robotic arms to solve a puzzle cube. OpenAI’s Dactyl took minutes with a robotic hand in 2019. When will a robot beat a 2.51 second human record with a general-purpose hand?
10 November 2026
Three devices, one paragraph each
The electronics study is separate from the model-launch benchmarks, and it is the one hardware buyers should read. Anthropic and Diode gave Fable 5.2 a short brief per product. The speakerphone brief asked for a USB-C puck with microphones, a speaker, Bluetooth, and a PCB antenna. Claude wrote the specification, schematic, placement, routing, 3D enclosure, firmware, and a parts list. Diode assembled the boards and ran the firmware under Claude’s direction. Anthropic says the speakerphone was still working after weeks of use, and that the hub and pet were also still working. They also say continued use and commercial certification are unknown.
This clip is Claude’s YouTube video “Claude Fable 5.2 designs a working voice-assistant hub.” Treat it as a lab demo. It is not a production test report.
What the model actually changed on the boards
- Order of routing. On the hub, Claude routed USB and Ethernet differential pairs first, matched lengths, and added tuning zigzags.
- Pin swaps instead of vias under the touch wheel. On the four-layer speakerphone, it reassigned chip pins so tracks under the touch electrodes stayed on one layer. A via there would punch through an electrode.
- Antenna keep-out. It kept copper and holes away from the Bluetooth antenna zigzag on every layer.
- Self-rewrite. When hub traces tangled, it deleted them and rerouted. Diode’s earlier models often stopped at “95% routed” boards that still could not be built.
- Custom router. KiCad’s interactive router drew one track at a time. For long runs off a 100-pin connector, Claude wrote a grid-search router in about 30 minutes. Anthropic says each long track then took about a second. It later increased clearance between an audio clock and Ethernet.
- DFM sort. Checks flagged six thin solder-mask slivers. Claude widened the openings. A ground-to-ground flag was intentional, and Claude left it.
- Mechanical fit. The speakerphone board would not drop straight in, because the mute switch hit the rear wall. Claude tilted the board about 25 degrees in CAD and wrote that step into the assembly notes.
- Acoustics. A 25 mm microphone port resonated near 3.4 kHz. Claude shortened the ports to about 10 mm with a wedge enclosure, moving the resonance to about 8.9 kHz.
Diode manufactured all three boards from Claude’s files without a round of shop-floor redlines. Anthropic discloses a commercial relationship with Diode, so treat the “no redlines” line as a vendor-partner result, then check it on your own stackup. Our hardware manufacturing team still wants the constraint sheet, the netlist diff, and the fab capability notes before an EVT build.
PCB and CAD scores
Two numbers are doing most of the work in social posts. They measure different things.
| Benchmark | What it asks | Fable 5.2 | Comparison | Who ran it |
|---|---|---|---|---|
| PCB Gym | Place parts and reroute boards whose layout was removed | 69% | Fable 5: 25% | Diode, internal. Anthropic has a commercial relationship with Diode. |
| BenchCAD 2.1 (with tools) | Rebuild 3D parts, assemblies, and boards from drawings and renders. Score 0 to 100. | 83.5 at Max effort | Opus 5.5: 65.3. Fable 5.1: 44.4 | Public benchmark. Scores reported by Anthropic. |
BenchCAD 2.1 is the public cousin of that chart. Anthropic’s launch note plots accuracy against cost and says Fable 5.2 at Max effort leads Opus 5.5 and Fable 5.1 on that hardware set. R&D World noted the list price fell about 20% versus 5.1, while full-effort cost per task can still rise because the model spends more tokens. Price per million tokens and price per finished board are different budgets. Track both.
BenchCAD 2.1, Max effort, as reported by Anthropic
If you already run KiCad through MCP, read this next to our KiCad and Altium MCP comparison. Fable 5.2 did not replace Konnect or kicad-mcp-pro. In the Diode study it used the KiCad router Diode exposed, then wrote a second router when the built-in tool was too slow for long tracks. That is the same pattern as the JLCPCB Claude plus kicad-mcp-pro case: the model is only as safe as the tools you expose, and a custom autorouter still needs your clearance rules.
Robotics: simulation score and a physical clip
Anthropic’s robotics section has two artifacts. Keep them separate when you quote them on X or in a design review.
The clip, published on Claude’s YouTube channel on 9 November 2026, shows two off-the-shelf arms driven through Claude Code. Anthropic describes a puzzle cube, a bottle cap, and a drawing of a robot. RobotsBeat notes on-screen labels for sped-up footage and removed pauses. A 23-second edit is not a cycle-time study. It also is not evidence that Mythos 5.2 was the controller. Anthropic attributes the robotics comparison to Fable 5.2.
RoboDojo Sim asks a model to control a simulated two-arm robot from camera images. Anthropic says it used each model’s highest available effort, and that the runs were internal.
| Model | RoboDojo Sim average | Effort |
|---|---|---|
| Claude Fable 5.2 | 64.9% | Max |
| Claude Opus 5.5 | 39.4% | Max |
| Gemini 4 Argon | 34.2% | High (RobotsBeat’s reading of Anthropic’s chart) |
| GPT-6 Astra | 29.8% | Max |
| Claude Fable 5.1 | 27.6% | Max |
RoboDojo Sim, Anthropic internal runs
Alex Albert at Anthropic predicts robotics will follow AI’s bitter lesson. Fable 5.2 scored 64.9% on RoboDojo Sim, up from 27.6% for Fable 5.1. A separate robotics paper found general models could move real arms but were not precise enough inside a practical action budget. That paper did not test Fable 5.2.
10 November 2026
An earlier public robot-arm study, on Fable 5 versus Fable 5.1 with YAM arms, is a useful caution. At medium effort and a 20-call budget, Fable 5.1 completed “block into bowl” on 8 of 20 trials (40%) and “puzzle into groove” on 2 of 20 (10%). Fable 5 completed the bowl task once and the puzzle task never. That study did not include Fable 5.2. A jump in simulation does not automatically become a jump on your arm, your camera, or your action budget.
If the product is a manipulator, the model is one block in the stack. Joints, drivers, and gearboxes still dominate cost and reliability. Our actuator guide for humanoid hardware and the robotics components range cover that hardware. Physical AI engineering is the engagement when you need the controller, the board, and the fixture in one program.
Where Fable 5.2 sits against other published scores
Anthropic’s launch table uses each model’s highest available effort unless noted. Safeguards are on for Fable and Opus scores unless the row says otherwise. Mythos rows are the reduced-safeguard twin, which matters for cyber and biology, not for ordinary PCB layout.
| Evaluation | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% Xhigh | 32.4% Xhigh | 45.5% | 42.8% | 41.1% Medium |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% (Mythos 75.3%) | 52.6% | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1 offline, partial | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim | 64.9% | 27.6% | 39.4% | 29.8% | not in this row |
| Humanity’s Last Exam, with tools | 74.2% | 65.6% | 67.7% | 57.2% | not listed |
| Chartography, no tools | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
| Remote Labor Index, high effort | 41.3% | 17.9% | 21.3% | 20.8% | not listed |
On FrontierCode 2.0, Anthropic says High-effort Fable 5.2 is about level with Opus 5.5 at Max (45.3% versus 45.5%) at roughly half the cost per task, and ahead of Astra’s 42.8% at similar cost. Xhigh effort reaches 49.1%. Gemini 4 Carbon still leads the published AutomationBench cell at 53.9% versus Fable’s 51.4%. If your agent mostly clicks desktop tools, that row matters as much as BenchCAD.
Mythos-only rows in the same note: CyScenarioBench 76.3% (Mythos 5.1 was 61.7%), CVE rediscovery 89.6% (Mythos 5.1 was 78.3%). Those tasks are blocked on Fable by design. Do not plan a firmware security audit on Fable 5.2 and expect Mythos behavior. Firmware bring-up that is ordinary embedded work can stay on Fable, then move to a verified Mythos seat only if the task is exploit-class. For production firmware scope, see embedded systems and firmware.
What to pilot, and what to keep human
- Sandbox board. One repo, read-only KiCad or Altium MCP first, then writes. Compare Fable 5.2 with Opus 5.5 on the same netlist and count ERC, DRC, and netlist diffs. Method is in the earlier MCP article.
- Intent checks the model skipped in public prototypes. Anthropic lists electrical validation, antenna tuning, thermal test, and certification as not done. Add those before you quote “it worked.”
- Factory rules. Match trace, space, drill, and mask sliver limits to your contract manufacturer, not only to KiCad defaults. Claude fixed six mask slivers because a checker existed. If your checker is missing, the model will not invent your fab’s limits.
- Robot pilots. Log completions, retries, and tokens per finished pick. The YAM-arm study showed a model can look busy and still miss the groove.
- Do not send Mythos prompts at a PCB layout problem unless you already have verification access and a reason that is cyber or life sciences. The layout study used Fable.
Techwall takes OEM programs that need DFM, production firmware, and a Hong Kong commercial owner. If you want a scoping call on a board or a robot controller, contact us. An engineer replies within one business day with a scoped consultation.
FAQ
Are Claude Fable 5.2 and Claude Mythos 5.2 different models?
No. Anthropic says they are the same underlying model. Fable 5.2 adds safeguards and is generally available. Mythos 5.2 is for verified cybersecurity and life-sciences organizations. API price and context window match.
Did Claude Fable 5.2 design a production PCB?
It designed three working prototypes with Diode: a USB speakerphone, a voice-assistant hub, and a handheld virtual pet. The boards were four-layer, ready for that prototype build in under a day, and needed no electrical modification after assembly. They were not certified, thermally qualified, or designed for mass production.
What is PCB Gym versus BenchCAD 2.1?
PCB Gym is Diode’s internal test of placement and rerouting after the original layout is removed. Fable 5.2 scored 69%, versus 25% for Fable 5. BenchCAD 2.1 is a public test of rebuilding parts, assemblies, and boards from drawings and images. Fable 5.2 scored 83.5, versus 65.3 for Opus 5.5 and 44.4 for Fable 5.1.
Can Fable 5.2 control a robot?
Anthropic published a clip of Fable 5.2 driving two off-the-shelf arms through Claude Code, including a puzzle cube, a bottle cap, and a drawing. On RoboDojo Sim it scored 64.9% in Anthropic’s runs, versus 39.4% for Opus 5.5 and 29.8% for GPT-6 Astra. The clip is edited. The percentage is a simulation average, not a factory OEE number.
Should we switch KiCad agents from Opus 5.5 to Fable 5.2?
Pilot both on one sandbox board. Fable 5.2 leads the published hardware scores. Opus 5.5 remains Anthropic’s recommended daily model, and High-effort Fable is close to Max-effort Opus on FrontierCode at lower cost per task. Measure issues per hour and dollars per merged change before you move a release branch.
Is a zero-DRC AI board ready to order?
No. DRC checks the rules file you loaded. It does not check design intent, signal integrity, antenna match, or your factory’s process window. The Diode prototypes still need the same review a human layout would get.
Sources
- Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026
- Anthropic, Claude can design electronic prototypes from start to finish, 9 November 2026
- Claude Platform, Fable 5.2 model overview
- Claude Platform, Mythos 5.2 model overview
- BenchCAD 2.1
- RobotsBeat, Fable 5.2 robotics test, 9 November 2026
- YouTube, Claude electronics design clip
- YouTube, Claude Fable 5.2 controls physical robot arms
- Anthropic on X, launch post
- Robocurve, Fable 5 versus Fable 5.1 on YAM arms
