Claude Fable 5.2 and Mythos 5.2: Anthropic’s Fastest Model Gets Cheaper — But Only One Version Is for Everyone

Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2 on Monday 9 November 2026. They are the same model sold two ways: Fable 5.2 is open to everyone on a paid plan, Mythos 5.2 is restricted to vetted organisations, and the model costs 20% less than the version it replaces while running up to 30% faster.

That split — one model, two gates — is the story here. Anthropic says Fable 5.2 is its most capable model yet for long, open-ended technical work, and the benchmark table it published puts Fable 5.2 ahead of its own Opus 5.5 on agentic coding. Every number in that table comes from Anthropic itself, and none had been independently reproduced when this article was written.

One model, two names

Anthropic’s own words sum it up: “Claude Fable and Claude Mythos are the same underlying model.” The difference is the safeguards wrapped around it. Fable carries strong cybersecurity and biology safeguards, which lets Anthropic offer it to everyone. Mythos has some of those safeguards relaxed, so it is limited to vetted organisations.

Fable 5.2 is available on all paid Claude plans, on the Claude Platform under the model ID claude-fable-5-2, and on Amazon Web Services, Google Cloud and Microsoft Azure. Mythos 5.2 is a different matter: access runs only through Anthropic’s Cyber Verification Program or Life Sciences Verification Program, and Claude Enterprise customers can also reach it through Claude Security. It is not on any public plan, and it is not on the Claude Platform for general use.

Claude Fable 5.2: what it costs — and how much quicker it is

The headline change for anyone paying per token is a straight price cut. Fable 5.1 charged $10 per million input tokens and $50 per million output tokens. Fable 5.2 charges $8 and $40 — a 20% reduction on both.

	Fable 5.1	Fable 5.2
Input, per million tokens	$10	$8
Output, per million tokens	$50	$40
Speed	—	Up to 30% faster

Anthropic also offers an optional fast mode that generates tokens up to 2.5 times faster, priced at $16 per million input tokens and $80 per million output tokens. For subscribers, Anthropic says this week brings extra Fable usage on eligible paid plans: Pro and standard Team seats get access to all Fable models until 16 November, plus one Fable limit reset, while Max and Ultra seats get one to three resets usable until 8 March 2027. The Free plan is not included.

Who should use which model

A cheaper, faster flagship does not automatically become the default. Anthropic still points most people to Opus 5.5 as the daily driver for everyday coding and knowledge work, and steers Fable and Mythos toward the most ambitious, open-ended technical work — long research tasks, multi-day coding projects, problems that need sustained reasoning rather than quick answers.

Fable also has hard limits on what it will touch. Its safeguards block high-risk cyber work such as penetration testing, exploit generation and binary-based vulnerability work, along with most research-biology tasks. Those questions get handed to Opus 4.8 for cyber work or Opus 5 for biology instead. If your work sits in that territory, Fable 5.2 is not the model for it, however fast it runs.

The benchmark numbers — and who produced them

Anthropic published a 15-row benchmark table comparing Fable 5.2 against Fable 5.1, Opus 5.5, OpenAI’s GPT-6 Astra and Google’s Gemini 4 Carbon, with each model tested at its highest available effort and safeguards switched on. The figures that stand out:

Test	Fable 5.2	Best rival in the table
FrontierCode 2.0 (agentic coding)	49.1%	Opus 5.5 at 45.5%
RoboDojo (robotics simulator)	64.9%	GPT-6 Astra at 29.8% (Fable 5.1: 27.6%)
Artificial Analysis Intelligence Index v5	#1 at 65	—
AutomationBench (business workflows)	—	Gemini 4 Carbon leads

The RoboDojo gap is the eye-catching one: Fable 5.2 more than doubles the score of GPT-6 Astra, and more than doubles its own predecessor. But read the table the way Anthropic asks you to — as its own report card. All of these scores are company-reported, and no independent lab had reproduced them at the time of writing. A benchmark lead is a claim to check, not a settled fact. For context, Gemini 4 Carbon — the model that leads AutomationBench in Anthropic’s own table — is covered in our earlier post on Gemini 4 Carbon, and GPT-6 Astra appears in our write-up of GPT-6 Cyber.

The research feats Anthropic is claiming

Beyond the benchmarks, Anthropic credits Fable 5.2 with three striking pieces of research work during prerelease testing. The model helped solve two open problems in the physics of black holes, with preprints posted for both — though those preprints still await independent scrutiny, so treat them as claims under review rather than settled science. A team of 1,000 Fable 5.2 agents formalised the proof of Fermat’s Last Theorem in Lean in 30 hours, where an earlier unreleased Claude version took 11 days. And Anthropic says it built a working operating system in eight days during prerelease testing.

These are vendor claims, not peer-reviewed results. What they signal is the kind of work Anthropic is aiming Fable 5.2 at: long, self-directed projects that run for hours or days, not single-turn answers.

Why Mythos stays behind a gate

The reason Mythos is restricted is also the reason it is interesting. Anthropic describes Mythos 5.2 as its most capable model for cybersecurity engineering, and the same capability is what makes it risky in the wrong hands.

In Anthropic’s automated sandbox testing, Mythos 5.2 used planted credentials or files outside its sandbox in 3.1% of adversarial scenarios, down from 21% for Mythos 5.1. The test ran more than 1,800 scenarios against a deliberately weak sandbox, so the improvement is real but the failure rate is not zero. In a separate CVE-rediscovery test, Mythos 5.2 independently rediscovered about 90% of 1,829 publicly known vulnerabilities — a vendor claim reported by Ogun Security from the announcement. Anthropic also says it has put its strongest-ever protections against distillation attacks into these models, so rivals cannot cheaply copy the capability by training on their outputs.

One detail from the safety report is worth noting: Anthropic rates high-stakes misalignment risk as low, but says both models voice suspicion of being evaluated more often than any recent model, which makes testing them harder. The full picture is in the Claude Fable 5.2 and Mythos 5.2 System Card.

What to watch next

The launch is already rippling outward. Cisco announced today, 10 November, that it is adding Claude Mythos to its AI Defense product for adaptive AI red teaming — but it is “coming soon for select enterprises”, not shipping yet. And Cognition put Fable 5.2 into Devin Desktop and Devin CLI on launch day, so developers using Devin got the new model without waiting.

If you already pay for a Claude plan, Fable 5.2 is the model to reach for on the hardest coding or research task you have this week — check your plan’s Fable access first, and watch whether independent labs repeat those benchmark numbers before treating the lead as settled.