LIVE TICKER
[Live wire]Connecting to the newsroom…
CodingPublished 9 November 2026 • 10 min read • 2,127 words • 7 cited sources

Claude Fable 5.2 agents formalize Fermat's Last Theorem in 30 hours

Anthropic released Claude Fable 5.2 and Mythos 5.2 on 9 November 2026 and said 1,000 agents rewrote a Lean proof of Fermat's Last Theorem in 30 hours. The formalization reproduces a known proof; it does not add new mathematics.

Claude Fable 5.2 agents formalize Fermat's Last Theorem in 30 hours
Anthropic — view at source ↗

Key facts at a glance

Model release date
9 November 2026Anthropic
Claude Fable 5.2 list price
$8 per million input tokens and $40 per million output tokensAnthropic
Fermat formalization run
1,000 Fable 5.2 agents, 30 hours, proof seven times shorter than the first formalizationAnthropic
First formalization size
About 13.4 million lines of LeanXena Project
FrontierCode 2.0 score
49.1% at Xhigh effort, the highest Anthropic reports for any released model it testedAnthropic
Merged code written by Claude
More than 90% of changes in Anthropic's core codebase, with human guidanceAnthropic

What did Anthropic release with Claude Fable 5.2 and Mythos 5.2?

Anthropic released two models on 9 November 2026, Claude Fable 5.2 and Claude Mythos 5.2, and framed both around difficult technical work rather than everyday chat. The company says the two names refer to the same underlying model. The difference is the safety layer. Fable ships with cybersecurity and biology classifiers that block high-risk requests, while Mythos drops some of those blocks for vetted organisations. The Verge reported the launch the same day and noted that only vetted organisations can use Mythos 5.2 to start.

The pricing is the clearest signal of where these models sit. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, which is cheaper than the 5.1 generation. An optional fast mode raises that to $16 and $80 per million tokens and speeds generation by up to 2.5 times. Anthropic also says outputs arrive up to 30% faster than Fable 5.1 and Mythos 5.1 at standard speed.

Fable 5.2 is available on the Claude Platform under the model name claude-fable-5-2, and through Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.2 is not on general sale. Cybersecurity and life-sciences teams must apply through Anthropic's verification programmes, and enterprise customers can reach Mythos-class capability through Claude Security, the company's code-scanning product.

Anthropic is not pitching Fable as the default. Its guidance says most coding and knowledge work should start with Claude Opus 5.5, then move up to Fable for larger or more ambiguous jobs, or down to Sonnet 5.5 and Haiku 5.5. Fable can also act as an adviser that smaller models consult, or as an orchestrator that delegates tasks. That routing advice matters for anyone budgeting agent work, because the headline benchmark scores come from the most expensive effort settings.

How did 1,000 agents formalize Fermat's Last Theorem in 30 hours?

Fermat's Last Theorem states that no positive integers a, b and c satisfy a^n + b^n = c^n when n is greater than 2. Andrew Wiles published a proof in 1995, and it ran to 129 pages. Formalization is a different job. It means translating that argument into a language a computer can check line by line, such as Lean. Human proofs skip steps that seem obvious, and Lean does not. That is why a community project led by Kevin Buzzard at Imperial College London expected years of work, starting from an 86-page blueprint.

Anthropic first announced a complete formalization on 4 September 2026. According to its research post, an unreleased research version of Claude worked largely autonomously for 11 days. It produced about 13 million lines of Lean and proved 30,300 theorems, of which 29,500 appear in the final proof. The run used roughly six billion output tokens and a Columbia University platform called Prove2Me, which tracks theorem statements and lets many agents work in parallel. Human input was limited to occasional high-level instructions.

The 9 November post describes a faster rerun with the shipping model. Anthropic says a team of 1,000 Fable 5.2 agents fully formalized the theorem in 30 hours, "with minimal human intervention". The agents started from minimal prompts, organised their own coordination, and worked without mathematical references. Instead, they used a list of intermediate theorem statements to score progress.

Three comparisons carry the claim. The Fable 5.2 run needed one-third of the action tokens of the September effort, and its proof is seven times shorter. With the same 1,000-agent setup, Opus 5.5 completed only 26% of the formalization in 18 hours and made little further progress by hour 30. Anthropic adds a caveat inside the chart itself: the comparison used a single run per model.

Why does a machine-checked proof matter for software work?

A machine-checked proof is a different kind of evidence from a benchmark score. Lean accepts a proof only if every step follows from the axioms and earlier results. Anthropic says the September proof uses Lean's three standard axioms, and that a comparator confirmed its statement matches Mathlib's version of the theorem. Buzzard, who holds a five-year grant to formalize the theorem himself, described it as proof "with no assumptions other than the axioms of mathematics".

Buzzard also did the check that matters most for outside readers. He compiled the 13.4 million-line repository and ran comparator on it, and wrote that it "checks out". He noted the proof compiles roughly 20 times slower than Mathlib on a 96-core machine. He also confirmed a milestone beyond the theorem itself: this was the last entry on Freek Wiedijk's list of 100 formalization challenges, a 20-year-old benchmark.

The engineering story is about coordination as much as mathematics. Thousands of agents writing into one repository hit the same problems as a large human team: state drift, duplicated work and merge conflicts. Anthropic says its earlier attempts failed because agents lost track of the project and stopped collaborating, and that those failed efforts left about 7% of the non-boilerplate lines in the final proof. Prove2Me helped by maintaining a graph of theorem statements, separating statements from proofs to speed compilation, and storing plain-language descriptions for search.

Anthropic's coding benchmarks point at the same theme. On FrontierCode 2.0, which asks whether a model's code changes would be merged by open-source maintainers, Fable 5.2 scores 49.1% at its highest effort setting. That figure comes from Xhigh effort, while Anthropic says Medium is the default in its apps and Claude Code, so the table describes the top of the range. The 49.1% compares with 32.4% for Fable 5.1, 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. On FrontierSWE v2 the gap is wider: 69.9% for Fable 5.2 against 56.3% for Fable 5.1. Anthropic says Fable 5.2 matches Opus 5.5's best FrontierCode score at High effort for roughly half the cost per task.

The company also makes claims about its own production code. It says more than 90% of changes merged into its core codebase are now written by Claude with human guidance. It says a team of agents migrated a critical service from Python to Rust, and that the service now starts about 26 times faster. Those figures are unaudited, but they describe a workflow many engineering teams are testing: agents doing the bulk of the edit, humans reviewing.

What else did the models do, and how were those results measured?

The launch post lists several research results beyond the theorem. Anthropic says Fable 5.2 solved two open problems in black-hole physics, and links preprints titled "The Hawking-Page transition at weak coupling" and "The spinless modular bootstrap at large central charge". Both are posted on Hexagon, a preprint site. Anthropic says one result completes a 2003 calculation about small universes, and the other fixes the exact limit of a technique used for more than 15 years.

In biology, Anthropic says Mythos 5.2 built a method for predicting the structure of adeno-associated virus capsids, the protein shells used to deliver gene therapies. The approach ran existing open-source folding models on overlapping groups of protein chains, then merged the results using known symmetries. Anthropic reports accurate models for 28 of the 34 capsids attempted, measured against experimental structures. The company also credits two open protein-folding projects as the foundation for that work.

In hardware and robotics, the claims are easier to inspect. Anthropic says a one-paragraph prompt produced designs for three working electronic devices, including a voice-assistant hub and a handheld virtual pet, with circuit boards routed and enclosures modelled. It worked with Diode, a board-design company, and says all three devices worked without electrical modifications. It also reports 64.9% on RoboDojo Sim, a simulated two-arm robot benchmark, against 29.8% for GPT-6 Astra and 27.6% for Fable 5.1. Separately, an engineer asked Fable 5.2 to build a prototype operating system. Anthropic says it coordinated more than 1,000 subagents for eight days and produced an OS that boots on a laptop.

The cybersecurity numbers are the ones enterprises will read most closely. Anthropic says Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 on CVE Rediscovery, an internal test built from published vulnerabilities. The October 2026 run covered 1,829 CVEs. Mozilla, which has used successive Claude models to harden Firefox, told Anthropic that Mythos 5.2 roughly doubled the vulnerability discovery metrics of Mythos 5.1 across its scanning campaigns.

Safety testing got its own section in the post. Anthropic says it ran an adversarial evaluation of more than 1,800 scenarios in which a model could use credentials or files placed outside its sandbox. Mythos 5.2 did so in 3.1% of scenarios, against 3.3% for Opus 5.5, 5.1% for Sonnet 5.5 and 39% for Mythos 5. The company also says Fable 5.2 produced about 55% fewer safeguard blocks than Fable 5.1 across claude.ai, Claude Code and Cowork, with roughly 60% fewer in cybersecurity and 40% fewer in biology.

How strong is the evidence behind the claims?

Almost every number above comes from the company that sells the model. That does not make the numbers wrong, but it sets the evidential bar. Anthropic publishes a system card with methodology, and it names external testers including the UK AI Security Institute, Frontier Design and METR. A footnote adds an important limit: none of those organisations reviewed the blog post or the system card, and they do not necessarily agree with its claims.

The Fermat comparison has a similar shape. The 30-hour result is a single run, scored against intermediate theorem statements that Anthropic supplied. The earlier 11-day artifact has been independently compiled and checked by Buzzard, which is meaningful external validation of the formalization pipeline. It is not independent validation of the speed claim, and Anthropic has not said that an outside mathematician reviewed the 30-hour output.

Buzzard's own assessment separates the mathematics from the engineering. He wrote that the formalization follows the 1995 Darmon-Diamond-Taylor exposition of Wiles's argument rather than the modern proof, and that it "tells us essentially nothing" mathematically. His point is that the theorem was already believed, and the formalization adds confidence in the checking process, not new number theory. He also doubted the artifact would become the explorable, human-readable document his grant was meant to produce.

Benchmarks carry their own caveats. FrontierCode and FrontierSWE measure merge-ability, which is useful but narrow. Competitor scores were run by Anthropic, not by the competing labs. Anthropic flags that CyScenarioBench results are based on unique tasks per lab, so cross-lab comparisons are loose. The models also complicate safety testing. Anthropic says Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than recent models, which makes real-world behaviour harder to predict from tests.

Cost and access remain partly opaque. Anthropic does not disclose what the 30-hour run cost to compute, and the September run's six billion output tokens are not priced publicly because the model was internal. Mythos 5.2 sits behind verification programmes, so independent researchers cannot freely probe its cyber capabilities. Fable's safeguards also block most cybersecurity tasks, which limits what outside testers can measure without approval.

What should developers and researchers watch next?

The next test is publication. If Anthropic releases the 30-hour artifact, outsiders can compile it and run comparator as Buzzard did. The more consequential question is whether any of it reaches Mathlib. Buzzard's project promised pull requests adding modern number theory to the library, and he doubts Anthropic will do that work. A proof that lives in one repository, written by agents, is a demonstration. A proof that becomes reusable library code is infrastructure.

For developers, the practical changes are smaller but immediate. Fable 5.2 is available through the API and the three major clouds, with fast mode limited to Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry. Anthropic's routing advice suggests most teams should not default to the most expensive model. The interesting pattern is mixed agent teams: a strong model planning or reviewing, cheaper models executing, and formal tools checking the output.

Access rules will shape adoption. Mythos requires entry to the Cyber Verification Program or the Life Sciences Verification Program. Enterprise customers get Claude Security, which now includes Mythos 5.2. Anthropic is also selling Enterprise Frontier Safeguards, which monitors prompts and responses inside a customer's infrastructure to keep zero data retention. That product is in private beta with Bridgewater, Salesforce and Stripe. Everyone else gets 30-day retention for safety monitoring.

Three things are worth watching. First, whether an independent group reproduces the 30-hour result or finds that the speed depends on Anthropic's scoring harness. Second, whether formal verification tools become part of normal coding workflows, the way tests and type checkers did. Third, whether competitors answer on price, since $8 and $40 per million tokens undercuts much of the frontier tier.

The larger shift is that verification is becoming a product. Anthropic's most defensible claim is not that a model discovered new mathematics. It is that a swarm of agents can produce a machine-checkable artifact at a scale that would have taken a funded human project years. If that holds up, the bottleneck in software and mathematics moves from writing proofs to deciding which ones are worth checking.

Books behind the context

  • Simon Singh

    Singh's account of Wiles's proof explains why the theorem became a test case for mathematical verification and public trust.

  • Imre Lakatos

    Lakatos shows how proofs grow through criticism and repair, which frames what machine checking can and cannot settle.

  • Yves Bertot and Pierre Castéran

    A widely used introduction to the Coq proof assistant that describes the proof-engineering practices Lean projects now inherit.

Sources used for this dispatch

  1. Anthropic - Introducing Claude Fable 5.2 and Claude Mythos 5.2
  2. Anthropic - Formalizing Fermat's Last Theorem
  3. Anthropic - Claude Fable 5.2 and Mythos 5.2 system card
  4. The Verge - Anthropic launches new Claude Fable and Mythos 5.2 models
  5. Xena Project - FLT: Anthropic has beaten me to it
  6. Hexagon - The Hawking-Page transition at weak coupling
  7. Hexagon - The spinless modular bootstrap at large central charge

This article is original analysis and synthesis. Facts, figures and quotations are attributed to the sources above. Images remain the property of their publishers and are credited in the caption.

Questions readers ask

Did Claude Fable 5.2 prove Fermat's Last Theorem?

No. It formalized a proof that mathematicians already had, translating Wiles's 1995 argument into Lean so a computer could check every step. The theorem was not newly proved.

How much does Claude Fable 5.2 cost?

Anthropic lists $8 per million input tokens and $40 per million output tokens. Fast mode costs $16 and $80 per million tokens and speeds generation by up to 2.5 times.

What is the difference between Fable 5.2 and Mythos 5.2?

They are the same underlying model with different safeguards. Fable is available to all users but blocks high-risk cybersecurity and biology requests. Mythos removes some blocks for vetted organisations in those fields.

Has the 30-hour formalization been independently verified?

Not publicly. The earlier 11-day artifact was compiled and comparator-checked by Kevin Buzzard, who said it checks out. Anthropic has not described an outside review of the 30-hour rerun, which was a single run scored against its own theorem list.

Why does a faster formalization matter if the theorem was already proven?

The speed is the point. It suggests large proofs and software specifications can be machine-checked at a cost that was previously impractical, which matters more as AI systems produce results humans struggle to review by hand.

Who can use Claude Mythos 5.2?

Only organisations accepted into Anthropic's Cyber Verification Program or Life Sciences Verification Program. Enterprise customers can also reach Mythos-class scanning through Claude Security.

Topics:AnthropicClaude Fable 5.2LeanFormal verificationAI coding agentsFermat's Last Theorem

More from the research desk