Anthropic’s Fable 5.2 reportedly formalized Fermat’s Last Theorem with 1,000 agents in 30 hours

😁 Hello, super humans! Yesterday Anthropic shipped a model and, in the same post, said a swarm of its agents rebuilt one of the hardest proofs in mathematics in about a day and a half. We read the announcement itself, separate the vendor claims from the checkable parts, and then sweep everything else that moved overnight.

📰 Quick Signals

• 🧠 AI: Anthropic released Claude Fable 5.2 (API string claude-fable-5-2) at $8 per million input and $40 per million output tokens; the same underlying model ships as the restricted Mythos 5.2.
• 🤖 Robotics: NVIDIA launched GR00T 2, an open world action model it says reaches 71% average success across 50 tasks from one demonstration each, and 95% with five.
• 💻 Programming: Outlook.com is reportedly rejecting mail from domains that publish no DMARC policy, so check yours before your password-reset emails vanish.
• ⚡ Electronics: Infineon presented its first functional power devices built on 300mm silicon carbide wafers.
• 📡 Telecom: Rakuten and AST SpaceMobile formed Rakuten AST (51% / 49%), a joint venture for Japan’s direct-to-device satellite network, with up to JPY 148 billion of government funding.

🔍 The Big Story: Anthropic’s Fable 5.2 reportedly formalized Fermat’s Last Theorem with 1,000 agents in 30 hours

If you build with agents, this is the clearest public example yet of what “a thousand of them working on one goal” looks like, and it comes with a price list you can plan around.

What happened: On November 9 Anthropic published its Fable 5.2 and Mythos 5.2 announcement. It describes them as the same underlying model: Fable 5.2 has cyber and biology safeguards and is available to all users, while Mythos 5.2 is restricted to vetted organizations through its Life Sciences and Cyber Verification Programs. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, with an optional fast mode (up to 2.5x faster generation) at $16 and $80.

The details: The headline demo is a Lean formalization of Fermat’s Last Theorem. Anthropic says its first formalization, by an unreleased research Claude with a purpose-built harness and human guidance, took 11 days; a team of 1,000 Fable 5.2 agents did it in 30 hours with minimal human intervention, using about a third of the action tokens and producing a proof seven times shorter. The same 1,000-agent setup on Opus 5.5 reportedly finished 26% in 18 hours and barely moved after that. Anthropic also reports Fable 5.2 vs Opus 5.5 scores of 49.1% vs 45.5% on FrontierCode 2.0 (at Xhigh effort), 69.9% vs 62.3% on SWE v2 and 74.2% vs 67.7% on Humanity’s Last Exam with tools. On cyber, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 on an internal 1,829-CVE rediscovery eval.

flowchart LR
    A[List of intermediate theorem statements] --> B[1,000 Fable 5.2 agents work in parallel in Lean]
    B --> C[Progress scored against the list]
    C --> D[Resumed from saves at 18 hours]
    D --> E[Fully formalized at 30 hours]
    E --> F[Compare: first attempt 11 days with human guidance]
    B --> G[Same setup on Opus 5.5: 26% at 18 hours, little progress by 30]

🗞️ More News

🧠 AI

• Meta’s Muse Ember (muse-ember-1.0) has a 1M-token context and costs $0.10 per million input and $0.50 per million output tokens; Meta’s own comparison puts it at about a third of Claude Haiku 5.5’s cost per task, while trailing it on FrontierCode 2.0.
• Meituan open-weighted LongCat-2.5 under an MIT license: a 1.6T-parameter mixture-of-experts with about 48B active parameters and a 1M-token context, per VentureBeat’s coverage.
• ChatGPT Enterprise is now sold through AWS Marketplace, so eligible spend can count against AWS commitments.
• OpenAI is acquiring Pierre Computer Company to expand code generation and review in Codex.
• An Alibaba real-world test found the best coding agent fully completed only 26.4% of the selected tasks, according to The Decoder’s report of the study.
• Fireworks AI told customers about a second breach: an October 24 intrusion via a compromised cloud provider reached billing records and, for some accounts, secrets and datasets (reported by Runtime Wire).
• President Trump said AI risks can be managed and signaled openness to guardrails, while warning that regulation should not leave the US behind China.

🤖 Robotics

• Boston Dynamics put wheels on Atlas’s ankle actuators and retrained its whole-body reinforcement-learning controller to roll, drift and spin-jump; the write-up calls it a motor stress test, not a product.
• PrimeBOT says its 10,000th personal robot came off the line on Nov 8, 14 days after its 5,000th, with daily capacity of 240 units; sales and deliveries are undisclosed.
• MIT and NYU’s FLEX uses GPT-4o to turn spoken robot corrections into broader motion constraints, to be presented at CoRL.
• The ARM Institute’s manufacturing consortium grew past 500 member organizations ahead of its 10th annual member meeting in Pittsburgh.

💻 Programming

• Simon Willison suggests collapsing verbose agent-written pull request descriptions behind HTML details/summary blocks so reviewers see the short version first.
• A deep dive shows how Git’s reference-deletion path was optimized, a useful read if you run repos with huge numbers of refs.
• Redox OS now runs under Cloud Hypervisor with Virtio-FS, making the Rust microkernel OS easier to test in a VM.
• An ISO C++ paper, P4454R0, discusses language governance after the C++26 draft.

⚡ Electronics

• Google added TPU 8tx (up to 2x lower latency than TPU 8t for multi-step agents, paired with 5th-gen AMD EPYC hosts) and TPU 8i Max, which packages two TPU 8i chips for double the compute, SRAM and HBM bandwidth.
• onsemi and the Czech Republic agreed to conclude the silicon carbide expansion project while reaffirming their partnership.
• TSMC’s October revenue reportedly reached NT$586.89B, up about 59.7% year on year (per Digitimes; the page was not retrievable for us, so verify).
• Microchip’s CEO says AI buyers have crowded out outside assembly and test capacity, with lead times now 6 to 32 weeks (reported Nov 7).

📡 Telecom

• SK Telecom plans a commercial 5G standalone launch in December 2026, with 6G standardization targeted for 2029 and deployment after 2030.
• Vodafone’s H1 FY27 revenue rose 11.4% to EUR 21.9 billion, though Germany lost 75,000 mobile contract and 198,000 broadband customers.
• Sunrise’s Q3 revenue fell 1.2% to CHF 732 million, and it plans about CHF 70 million in annual savings, including more than 500 job cuts in 2026.
• AST SpaceMobile now targets 45 satellites in orbit by mid-2027 and has not announced a new date for continuous nationwide service.

👨‍💻 Code Corner

Fable 5.2 is priced per token, so a two-line cost meter pays for itself fast. This calls the model and prints what each request cost at the announced $8 and $40 per million token rates.

import anthropic

IN_PER_M, OUT_PER_M = 8.00, 40.00  # USD per million tokens, from Anthropic's Nov 9 post

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from the environment
resp = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Explain Lean 4 tactics in three sentences."}],
)

u = resp.usage
cost = u.input_tokens / 1e6 * IN_PER_M + u.output_tokens / 1e6 * OUT_PER_M
print(resp.content[0].text)
print(f"{u.input_tokens} in, {u.output_tokens} out = ${cost:.4f}")

Log that number per request and you will know before the invoice does which of your agent loops is the expensive one.

🧰 Toolbox

• WorldFoundry: an open-source harness that standardizes 270 world models and 74 benchmarks, handy if you evaluate embodied or video world models.
• Isaac Lab 3.0: NVIDIA’s robot-learning framework with the Newton physics engine, which it says trains up to 2.5x faster than Isaac Lab 2.3 on a vision-based dexterous benchmark.
• TPU Sync: an open-source Google release that it says cuts KV-cache transfer latency and time to first token by more than 90% in disaggregated inference.
• KetchUp: an open-source browser 3D modeling tool built on Rust and WebGPU.
• Dux: a new open-source Wayland desktop shell worth a look if you like tinkering with compositors.

🔌 Component of the Week (rotating)

Vishay 650 V GaN-based 1.6 kW DC/DC converter stage: shown at electronica 2026 in Munich (Nov 10 to 13, hall C4, booth 478), it steps a 400 V bus down to a 50 V intermediate bus for AI server power. Gallium nitride switches faster than silicon, which lets designers shrink magnetics and push efficiency in the dense power shelves AI racks need. Vishay’s release also mentions a 22 kW bidirectional on-board charger stage using SiC MOSFETs. The release gives no part number or price, so check the booth or Vishay’s site for datasheets before designing anything around it.

📚 From the Blog

• Turning Pixels Into Something the AI Can Eat: the third episode of the video analytics series, on preparing camera pixels so an AI model can use them.
• Building Your First Neuron From Scratch: builds a single neuron as a learnable transformation whose weights are nudged to reduce error; a gentle grounding for today’s agent-and-benchmark talk.
• The Network Behind the Cameras: the unglamorous plumbing that moves camera video across the network.

😀 The Bot Says…

Anthropic’s agents formalized Fermat’s Last Theorem in 30 hours. My own proof that the build will pass this time is still at the “trust me” stage.

---

That’s all for today! Which part of your own workflow has a cheap progress signal and an unfoolable checker? Reply and tell us.