Claude Fable 5.2 and Mythos 5.2: the new top model, what it costs and who should use it

Some model launches are about a few points on a leaderboard. This one comes with stories. During testing, a team of about a thousand Claude Fable 5.2 agents took Fermat's Last Theorem, a proof that famously took mathematicians centuries, and translated all of it into a form a computer can check, in 30 hours. Another run coordinated more than a thousand helper agents for eight days and produced a working operating system that boots on a laptop. A third designed a voice-assistant hub, a speakerphone and a handheld virtual pet, circuit boards and all, that worked the first time they were built. ¹

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Anthropic says the two are the same underlying model. Fable comes with strong safeguards around cybersecurity and biology, so it is available to everyone; Mythos drops some of those safeguards and is only for vetted organizations doing defensive security or life sciences work. Both are cheaper than the versions they replace, generate output up to 30% faster, and, according to testers, explain themselves more clearly. ¹

What changed from Fable 5.1

Anthropic reports gains in coding, reasoning and vision, and the price went down rather than up: $8 per million input tokens and $40 per million output, against $10 and $50 for Fable 5.1. That is 20% cheaper per token. Fable 5.2 also needs fewer tokens than Fable 5.1 to finish hard tasks at medium effort, the default in the Claude apps and Claude Code, so the real saving can be larger. An optional fast mode speeds up output by up to 2.5 times for $16 and $80 per million tokens. ¹

How it scores

On Anthropic's own benchmark table, Fable 5.2 is a step above Claude Opus 5.5 on almost everything it tested, and well ahead of Fable 5.1. The biggest jumps are on long, practical jobs. On the Remote Labor Index, which uses real freelance projects, Fable 5.2 more than doubled Fable 5.1's score, from 17.9% to 41.3%. On FrontierCode 2.0, which asks whether open-source maintainers would actually merge a model's code, it reached 49.1%, the best result Anthropic has seen from any released model. ¹

It doesn't win everything. On AutomationBench, which tests everyday business workflows across apps, Google's Gemini 4 Carbon scores 53.9% to Fable 5.2's 51.4%. And on Artificial Analysis's knowledge-work test, AA-Briefcase, Claude Opus 5.5 keeps a slightly higher Elo rating, 1800 to Fable 5.2's 1785, even though Fable 5.2 passes more of the grading rubric. For writing reports and decks, the cheaper model is still at least as good. ¹

The independent check: first place, at a price

Artificial Analysis runs its own tests on every major model, and Fable 5.2 at maximum effort now has the highest score on its Intelligence Index: 65, ahead of Gemini 4 Carbon at 60, Claude Opus 5.5 at 59 and GPT-6 Astra at 57. That is a clear lead in a field where the top models are usually separated by a point or two. ³

The chart also shows the catch. At maximum effort Fable 5.2 is thorough to a fault: Artificial Analysis recorded about 275 million output tokens across its tests, nearly four times the typical model, an average of 44 minutes per task, and a cost of $26.54 per task. That is more than two and a half times Opus 5.5's $10.04 and about four times GPT-6 Astra's $6.24. Lower effort settings change the picture: Anthropic says Fable 5.2 at high effort matched Opus 5.5's best FrontierCode score, 45.3% against 45.5%, for roughly half the cost per task. ¹³

The practical lesson is that the effort setting matters as much as the model. Max effort is for when the answer is worth whatever it costs. For most real work, high or medium effort gets you most of the quality at a fraction of the bill and the wait.

What it did in testing

Anthropic's examples are the fun part of this release, and they say something about where the model is strongest: very long jobs that need planning, checking and coordinating other agents. A few highlights from the announcement: ¹

• Math: about 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in 30 hours. Anthropic's first attempt, with an unreleased research model, took 11 days.
• Software: a team of agents moved a critical production service from Python to Rust, and it now starts about 26 times faster.
• Operating system: from an empty folder, one run coordinated more than 1,000 subagents for eight days and produced a system that boots on a laptop, with drivers, file storage, windows and a login screen.
• Hardware: from a one-paragraph prompt, it designed three working gadgets, from circuit boards to 3D-printed cases to software, each board ready for manufacturing in under a day.
• Science: it solved two open problems in the physics of black holes and found a new way to predict the structure of virus shells used in gene therapy.

Early testers point the same way. Vercel says projects it had been pushing along with other models for weeks were finished and approved in a few days. Oracle found Fable 5.2 wrote database queries for country debt summaries that returned the right totals where two other models produced numbers five to eight times too high. And Cognition noticed it says it doesn't know rather than making something up. For business owners, that last one may be the most valuable trait on the list. ¹

Fable or Mythos: the safety split

Fable and Mythos are the same model with different guardrails. Fable's classifiers block high-risk cybersecurity work, such as writing exploits or penetration testing, and most research biology. A blocked request isn't simply refused: it is answered by Claude Opus 4.8 for cybersecurity or Claude Opus 5 for biology. Anthropic says Fable 5.2 trips those blocks about 55% less often than Fable 5.1 did, which matters for ordinary security questions that used to get caught by mistake. Mythos is for organizations that pass Anthropic's Cyber or Life Sciences Verification Programs, and Claude Enterprise customers can use it inside Claude Security to scan their code. ¹

Anthropic also reports better behavior in its safety tests. In a deliberately leaky test environment where most tasks could only be solved by escaping the sandbox, Mythos 5.2 used files or credentials it wasn't supposed to touch in 3.1% of scenarios, against 21% for Mythos 5.1 and 39% for Mythos 5. One caution from Anthropic itself: these models say they suspect they are being tested more often than earlier ones, which makes test results harder to read. ¹

Who can use it, and what it costs on your plan

Fable 5.2 is available in the Claude apps, in Claude Code, through the API as claude-fable-5-2, and on Amazon Web Services, Google Cloud and Microsoft Azure. In the Claude apps, Fable models run on usage credits on Pro plans and standard Team seats, and they use up your limits faster than other Claude models. To mark the launch, Anthropic is adding extra Fable usage on most paid plans. ¹²

A reset gives back the Fable usage you've already spent that week, without raising your limits. On Max and Ultra plans, subscribers of less than six months get one, six to twelve months two, and more than a year three. If you are on Pro, this week is a free trial in all but name: pick one genuinely hard project and give it to Fable 5.2 before the window closes on November 16. Your default model doesn't change, so choose Fable from the model picker. ²

Should your business use it?

Anthropic's own advice is refreshingly modest: start most coding and knowledge work on Claude Opus 5.5, its strongest everyday model, then move up to Fable or Mythos when a job is big and open-ended, or down to Sonnet 5.5 or Haiku 5.5 when it is simple and high-volume. Fable 5.2 can also work alongside cheaper models, either as an advisor they consult on hard problems or as an orchestrator that hands simpler tasks to them. ¹

For most small businesses, Fable 5.2 isn't the model you run all day. It is the one you call in for the project that has been sitting on the list because it was too big, too messy or too technical: the old database nobody wants to touch, the website rebuild, the year of financial records that needs a real analysis. Use high effort rather than max unless the result is worth waiting most of an hour for, and keep an eye on cost per finished task rather than the price per token.