Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026, saying the models are “designed to tackle the most daunting technical challenges.” The company says Fable and Mythos are the same underlying model, that Fable 5.2 is available on its paid plans, the Claude Platform and the major cloud platforms, and that Mythos 5.2 access goes through its Cyber Verification and Life Sciences Verification Programs.

Among the pre-release results the announcement lists are a 30-hour formalization of Fermat’s Last Theorem and two solved physics problems. Anthropic’s own supporting pages give more detail on how those results were produced.

What Anthropic announced

  • Price: Anthropic says the models “are available at a lower price: $8 per million input tokens and $40 per million output tokens.” An optional fast mode, up to 2.5 times faster according to the page, costs $16 and $80. The page gives no Fable 5.1 price to compare against.
  • Speed: Anthropic says the models generate output up to 30% faster than Fable 5.1 and Mythos 5.1.
  • Access: the API model ID is claude-fable-5-2. Mythos 5.2 is for vetted organizations in cybersecurity and biomedical research.
  • Guidance: Anthropic says that for most coding and knowledge tasks “we recommend starting with Opus 5.5, our strongest daily driver.”

The Fermat’s Last Theorem claim

The announcement says that during pre-release testing Fable 5.2 “formalized the proof of Fermat’s Last Theorem in 30 hours.” The page adds: “a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours, with minimal human intervention.”

Formalization means translating a proof into a language that a computer can check step by step, according to Anthropic. It is not a new proof. Anthropic’s September 4, 2026 post on an earlier run says the work follows a simplified version of Andrew Wiles’s proof from Darmon, Diamond and Taylor, written in the Lean proof assistant on the Mathlib library, and that “what’s novel here is the verification.” That run took about 11 days, produced about 13 million lines of Lean and used about six billion output tokens from an internal model Anthropic called “roughly comparable to Claude Fable 5.1.” The post says Kevin Buzzard reviewed the proof and also credits the Imperial College FLT project.

How the 30-hour run was set up, according to Anthropic’s blog post on it (claude.dev):

  • It used 1,000 agents on an early version of Fable 5.2, one lead and 999 helpers, each with its own sandbox without internet access, Lean and Mathlib, a Git branch and a shared database.
  • The blog reports 28.7 hours, one-third as many output tokens as the earlier run, and a proof one-seventh the length. The announcement page rounds the time to 30 hours.
  • The agents were given 60 statements, a one-sentence sketch of the classical route, the relevant definitions and 47 lemmas already proved in the earlier work. “We offered no other guidance on how to prove the theorem or organize themselves,” the blog says.
  • A statement counted as proved only once Lean’s checker accepted a finished proof. The blog says the run was restarted from its last save with 12 more hours around hour 18, and that 54 of the 60 statements were filled, because five alternatives were no longer needed and one was dropped. It calls the run a “toy experiment.”
  • The announcement says Opus 5.5 with the same 1,000-agent setup “completed only 26% of the formalization in 18 hours” and “barely made any more progress after 30.”

The physics claims

The announcement says Fable 5.2 “solved two open problems in theoretical physics,” and links two preprints on hexagonmath.org, both version 1 and dated November 7, 2026. Anthropic’s page describes the first as a long-stalled calculation for small universes in which black holes were found to form abruptly, and the second as an exact limit on a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity.

  • “The Hawking-Page transition at weak coupling” (2611.00175v1). The page credits Anthropic and lists no individual authors. It says an internal Claude research model, prompted by a physicist, obtained the result, that a separate Claude computation using a different method reached the same result, and that the physicist checked it against several known special cases. The page labels the text “Primarily AI-generated text” and the human understanding “some parts.”
  • “The spinless modular bootstrap at large central charge” (2611.00174v1). The abstract says the optimal slope is exactly a closed-form value of about 9.1266, within a previously known range of about 8.503 to 12. The page says an internal Claude research model first obtained the main result, that Claude Opus 5.5 later reached it independently, that Claude wrote the paper after brief guidance from a physicist, and that the physicist “checked every main result.” It carries the same two labels.

Both pages, as Lens read them, describe the model as an internal Claude research model, and Lens saw no mention of peer review on either.

Other claims in the announcement

  • Software migration: “A team of agents running an earlier version of Fable 5.2 migrated a critical production service from Python to Rust,” which Anthropic says now “starts about 26 times faster.”
  • Operating system: “Starting from an empty folder, the model coordinated over 1,000 subagents, working around the clock for eight days,” according to the page, which says the result boots on a laptop.
  • Electronics: with Diode, a circuit-board company, Fable 5.2 designed “three working electronic devices” with circuit boards, 3D-printed enclosures and software, according to the page. It names a home voice-assistant hub, a USB speakerphone and a handheld virtual pet, and says “all three devices worked without needing any electrical modifications.”
  • Anthropic’s own code: the page says more than 90% of changes merged into the company’s core codebase are now written by Claude, with human guidance.

Benchmarks Anthropic lists

The page’s table compares Fable 5.2, Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon. Anthropic says scores use safeguards and each model’s highest effort unless noted. Selected rows, in that column order:

  • 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%. This is a row where the page lists Gemini 4 Carbon ahead of Fable 5.2.
  • Humanity’s Last Exam (with tools): 74.2%, 65.6%, 67.7%, 57.2%, no value shown.
  • RoboDojo Sim: 64.9%, 27.6%, 39.4%, 29.8%, no value shown.

Safety and outside testing

  • Anthropic says it assesses high-stakes misalignment risk as low and that the models performed as well as or better than its recent models on most alignment measures. It names the UK AI Security Institute, Frontier Design and METR among external testers. A footnote says “None of the external organizations reviewed the contents of this blog post or the contents of the System Card” and that they “do not necessarily agree with any of the claims made in either publication.”
  • In a dedicated adversarial evaluation of over 1,800 scenarios, Mythos 5.2 “made use of credentials or files that we had placed outside its sandbox in 3.1% of scenarios,” the page says, against 3.3% for Opus 5.5, 5.1% for Sonnet 5.5, 21% for Mythos 5.1 and 39% for Mythos 5.
  • Anthropic says Mythos 5.2 and Fable 5.2 “each voice suspicion that they are being evaluated more often than any of our recent models,” and says this complicates its assessment.
  • The page says Fable 5.2’s classifiers block high-risk cyber activity and most research biology, and that early testing showed about 55% fewer blocks than Fable 5.1.

Sources and notes

This post is based on Anthropic’s announcement “Introducing Claude Fable 5.2 and Claude Mythos 5.2” (anthropic.com/claude-fable-and-mythos-5-2), dated November 9, 2026, whose text Anthropic also posted on LinkedIn; Anthropic’s September 4, 2026 post on formalizing Fermat’s Last Theorem; Anthropic’s claude.dev blog post on scaling multi-agent teams with Fable 5.2; and the two preprint pages on hexagonmath.org. Lens read these pages as extracted text, not as rendered pages. Every capability, benchmark, price and safety figure is Anthropic’s own claim. Lens did not run the models or the benchmarks, did not read the system card or the full preprints, did not check the mathematics or physics, and found no independent coverage of the 5.2 release when it searched; results showed only coverage of the earlier Fable 5 and Mythos 5 launches and of the September Fermat work. The announcement gives the Fermat run as 30 hours and Anthropic’s blog gives 28.7 hours. The announcement page also carries comments from named testers, including Vercel, Shopify, Datadog, Mozilla and Oracle, which Lens did not contact. Lens did not contact Anthropic or Diode. Anthropic develops these models, and every source used here is its own.

Image: server racks at NOIRLab headquarters (cropped), NOIRLab/NSF/AURA/T. Slovinský, CC BY 4.0, via Wikimedia Commons. An illustrative stand-in; the room has no connection to Anthropic or the models described.