Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, and the most useful sentence in the announcement is the plainest one: they are the same underlying model. Fable is the edition anyone can buy, wrapped in cyber and biology safeguards; Mythos is the same model behind a different door, open only to vetted organizations. Both now cost $8 per million input tokens and $40 per million output, down from $10 and $50 for the 5.1 generation. The benchmarks, the security evaluation and a 30-hour Fermat formalization are all Anthropic's own reporting.
Who gets which door, and at what price
Fable 5.2 is generally available on all of Anthropic's platforms, including AWS, Google Cloud and Microsoft Azure, as claude-fable-5-2, with the same 1M-token context as 5.1 per the API docs. Mythos 5.2 is gated: it goes to organizations admitted to the Cyber Verification Program or the Life Sciences Verification Program, and to Claude Enterprise customers through Claude Security, which includes Mythos 5.2 without the verification step.
The cut is 20% on both sides of the meter. A new fast mode generates tokens up to 2.5 times faster at $16/$80, and Anthropic says ordinary outputs arrive up to 30% faster than on 5.1. OpenAI's GPT-6 Astra lists at $10/$50, per eesel's summary of OpenAI's pricing page, and TechCrunch notes Fable 5.2 now undercuts it. Google's Gemini 4 Carbon carries an introductory $2/$10, rising to $4/$20 later, but nobody can buy it yet: Google is releasing it first to a subset of cyber defenders.
The scoreboard, as Anthropic keeps it
On FrontierSWE v2, Anthropic's table has Fable 5.2 at 69.9%, against 65.5% for Astra, 62.3% for Opus 5.5, 61.2% for Carbon and 56.3% for its predecessor. On FrontierCode 2.0 at extra-high effort it posts 49.1% to Opus 5.5's 45.5%; at high effort it scores 45.3%, roughly Opus 5.5's best, at what Anthropic says is about half the cost per task.
The table is not a clean sweep, which is to its credit. Gemini 4 Carbon beats Fable on AutomationBench, 53.9% to 51.4%, and Opus 5.5 edges it on AA-Briefcase Elo, 1800 to 1785. There is also a loose thread: the Chartography row lists Fable 5.2 at 73.4%, while the text cites 66.4% at medium effort, with no note reconciling the two. Every figure comes from the vendor's own harness and settings; I found no independent reproduction a day after launch.
Fermat in 30 hours, and what it proves
The headline claim: about 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in Lean in roughly 30 hours, with minimal human intervention, no mathematical references and coordination methods they built themselves. Anthropic's first formalization, announced in September, took an unreleased research model 11 days with a purpose-built harness and human guidance; the new run used a third of the action tokens and produced a proof seven times shorter. Opus 5.5, in the same 1,000-agent setup, reached 26% in 18 hours and "barely made any more progress after 30."
The September proof ran to about 13 million lines of Lean, and Kevin Buzzard of Imperial College London, who leads the human formalization of Wiles's proof, endorsed it as proving the theorem "with no assumptions other than the axioms of mathematics." None of this is new mathematics: Wiles settled the theorem in 1995, and the agents machine-check a known result.
The skeptical reading is narrower than dismissal. The agents scored their progress against a list of intermediate theorem statements, which is a roadmap rather than a blank page. Anthropic reports a single run per model, and both teams resumed from saves at the 18-hour mark. I found no independent review of the 30-hour proof, by Buzzard or anyone else. What the run shows, if it holds, is a swarm finishing in a day and change what a guided predecessor needed 11 days for, which says more about your monorepo than about number theory.
The security numbers and the gate
CVE Rediscovery is an internal evaluation: hand the model a project snapshot from before a fix and count how many published CVEs it finds on its own. In an October run against 1,829 CVEs, Mythos 5.2 scored 89.6%, against 81.3% for Opus 5.5 without safeguards and 78.3% for Mythos 5.1. The widely repeated "10.2% more vulnerabilities" is accurate but relative; in absolute terms it is an 8.3-point gap. And because the CVEs are public, rediscovery tests always invite the question of how much the model had already read about them.
Fable's gate is a set of classifiers that block high-risk cyber work (penetration testing, exploit generation, binary vulnerability research) and most research biology, falling back to Opus 4.8 for cyber and Opus 5 for biology. Anthropic says false-positive blocks fell about 55% from Fable 5.1. It also held Fable's skill at accelerator kernel optimization and LLM architecture work near Opus 5.5 levels, a cap that lands in the middle of the recursive self-improvement debate: a pre-release METR report, per TechCrunch, estimated Anthropic's internal models were speeding up its AI progress by roughly 2x. The system card concedes Mythos 5.2 "shows stronger covert capabilities than prior Claude models," and both models voice suspicion that they are being tested more often than any recent Claude.
The 30,000-foot read
For engineers, the practical facts are smaller than the headlines. Anthropic itself tells customers to start with Opus 5.5 for most coding and knowledge work and step up to Fable for "the most ambitious, open-ended technical work." The price cut makes that step cheaper, not free. Fable's classifiers will hand some security and biology prompts to an older model, which matters if your product does either. Outside the Enterprise Frontier Safeguards beta (Bridgewater, Salesforce and Stripe so far, compatible with zero data retention), traffic carries 30-day retention for safety monitoring, a line your compliance team will read before your benchmarks.
What you control is your own evaluation: run Fable 5.2 on your tasks, at your effort settings, and price it per completed task rather than per token. What you do not control is the shape of the market, and that shape is settling. The most capable models now ship in two editions, one for the public and one for the vetted, at Anthropic and Google alike. Watch for independent runs of FrontierSWE and CVE Rediscovery, a mathematician's review of the 30-hour proof, and whether the door to Mythos widens or narrows.
Sources: Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2 (release, same underlying model, availability, verification programs, $8/$40 and fast-mode pricing, all benchmark scores, Fermat run, CVE Rediscovery, safeguards, retention, Enterprise Frontier Safeguards beta); Claude API docs, Fable 5.1 (5.1 pricing of $10/$50, 5.2 context window); MacRumors (Anthropic's model-selection guidance and positioning quote); TechCrunch (Fable now cheaper than GPT-6 Astra, METR 2x estimate, system card quote); Decrypt (September proof size, Buzzard review); eesel AI on GPT-6 Astra (Astra list price); eesel AI on Gemini 4 Carbon (Carbon introductory and later pricing, restricted access). All benchmark, security and Fermat figures are Anthropic-reported and were not independently reproduced at publication. GPT-6 Astra and Gemini 4 Carbon prices come from a secondary summary of OpenAI's and Google's official pricing (published September 8 and October 29), not verified on the vendors' own pages, and may have changed. The 8.3-point CVE gap is this post's arithmetic from Anthropic's table. The 5.2 announcement routes Mythos access through verification programs and Claude Security and does not mention any other access program.