Claude Fable 5.2: cheaper, faster, and twice as good at real remote work, says Anthropic
Anthropic's new top model costs a fifth less per output token than the last one and doubles its score on a test of paid freelance jobs. It also ran a team of 1,000 agents that formalised Fermat's Last Theorem in about 30 hours. The figures are Anthropic's own, and for most work it still recommends a cheaper model.

Key points
- Anthropic released Claude Fable 5.2 on 9 November, on its own apps, its API and on AWS, Google Cloud and Azure. Claude Mythos 5.2 is the same model with fewer cyber and biology safeguards, for vetted organisations only.
- The price is $8 per million input tokens and $40 per million output tokens; output cost $50 for Fable 5.1. An optional fast mode generates up to 2.5 times faster at double the price.
- The biggest jump is on the Remote Labor Index, a test of real paid freelance projects: 41.3%, against 17.9% for Fable 5.1 and about 21% for Opus 5.5 and OpenAI's GPT-6 Astra.
- Anthropic says 1,000 Fable 5.2 agents reproduced a formal proof of Fermat's Last Theorem in 28.7 hours. They were given 60 target statements and 47 proved lemmas from an earlier run.
- Fable 5.2 still blocks most hacking and research-biology tasks, but Anthropic says it blocks about 55% fewer harmless requests than 5.1.
- For most coding and office work, Anthropic itself says to start with Claude Opus 5.5 and move up to Fable only for the hardest, open-ended jobs.
Disclosure: AI Beat uses Claude, among other tools, to help research and check its articles, including this one. All benchmark figures below come from Anthropic.
Anthropic has a new top model. Claude Fable 5.2, released on 9 November, is cheaper and faster than the version it replaces, and Anthropic says it is a clear step up on almost every test it ran. The most telling number is not about coding or exams. It is about whether a model can do a paid freelance job well enough that a client would accept it.
Two names, one model
Fable 5.2 and Mythos 5.2 are the same model with different safety settings. Fable is for everyone: it is on Claude’s apps, on Claude Code, on the API as claude-fable-5-2, and on Amazon Web Services, Google Cloud and Microsoft Azure. It blocks penetration testing, exploit writing and most research-biology tasks; when it does, a smaller Claude model answers instead.
Mythos 5.2 drops those blocks and goes only to organisations vetted through Anthropic’s Cyber and Life Sciences Verification Programs. Mozilla, which has used Claude models to harden Firefox all year, says Mythos 5.2 roughly doubled its bug-finding numbers compared with Mythos 5.1.
Anthropic says Fable 5.2 now blocks about 55% fewer harmless requests than Fable 5.1: about 60% fewer in cybersecurity and 40% fewer in biology.
The price
Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Fable 5.1 cost $50 per million output tokens when we last checked, on 30 September, so output is a fifth cheaper. Anthropic also says the new model writes up to 30% faster.
A new fast mode speeds generation up by as much as 2.5 times, for double the price: $16 and $80. It is on Claude Code, the Claude API, Google Cloud and Microsoft Foundry.
Paid subscribers get extra Fable time this week. Plans that include Fable get one to three usage resets, depending on how long they have subscribed, and on Pro and standard Team seats Fable can take up to half of the weekly usage limit.
Paid work, not just puzzles
The Remote Labor Index measures something most benchmarks don’t: whether a model can complete real projects of the kind people post on freelance sites, to a standard a client would accept. Fable 5.2 more than doubled its predecessor’s score.
Remote Labor Index: share of real freelance projects done to an acceptable standard
Fable 5.2 at High effort; other models at their highest available effort. Gemini 4 Carbon was not tested.
Source: Anthropic, 9 November 2026
Elsewhere the gains are smaller but broad. On FrontierSWE, a coding test, Fable 5.2 scores 69.9% against 65.5% for GPT-6 Astra. On Humanity’s Last Exam with tools it scores 74.2%. In robot control, on the RoboDojo simulator, it reaches 64.9%, more than double Fable 5.1 (27.6%). It does not win everywhere: Google’s Gemini 4 Carbon leads on AutomationBench, a test of business workflows, with 53.9% against 51.4%. And on one Elo rating of office work, Opus 5.5 still edges ahead.
These are the company’s own results, run on its own harnesses. Anthropic says outside testers, including the UK AI Security Institute and METR, examined the model before release, but none of them reviewed the announcement.
A thousand agents and a famous proof
The headline demonstration is mathematical. In September Anthropic formalised Wiles’s proof of Fermat’s Last Theorem, writing it in Lean, a language in which a computer checks every step. That took a researcher and dozens of agents 11 days.
This time a team of 1,000 Fable 5.2 agents did it in 28.7 hours, with one-third of the output tokens, and wrote a proof one-seventh as long. Each agent ran in its own sandbox with no internet, and a classifier checked every command before it ran.
The detail matters. The agents were reproducing a proof that had already been formalised once. They got 60 target statements, a one-sentence sketch of the classical route and 47 lemmas already proved in the earlier run. Anthropic calls it a “toy experiment”. It shows that a large team of agents can organise itself on a long job. It is not new mathematics.
Anthropic gives other examples. One engineer asked Fable 5.2 to design an operating system from an empty folder; it coordinated more than 1,000 sub-agents for eight days and produced one that boots on a laptop. It also designed three electronic devices, from circuit boards to cases, that worked without electrical changes, and it solved two open problems in the physics of black holes, now in preprints.
Who should use it
Anthropic’s own advice is modest. For most coding and knowledge work it recommends starting with Claude Opus 5.5, which is cheaper, and moving up to Fable only for “the most ambitious, open-ended technical work”. Fable 5.2 can also act as an adviser that smaller models consult, or as an orchestrator that hands simple tasks down to them.
Why it matters to you: if you pay for a Claude plan, this is a good week to try Fable on the hardest task you have, while the extra usage lasts. If you build on the API, the lower price makes Fable worth a fresh test against Opus 5.5 on your own work. Keep in mind that every number here is the maker’s own. The independent tests will tell more.
Comments
What do you think? No account needed. Please be kind; comments that attack people are removed.
No comments yet. Be the first.