Anthropic Cuts Prices With Claude Fable 5.2 and Mythos 5.2, Says It Held Back the Model's AI-Research Skills
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, the same underlying model behind two gates: Fable for all users with cyber and biology classifiers, Mythos for vetted cybersecurity and life-sciences organisations. Both cost $8/$40 per million tokens, which CNBC puts at 20% below 5.1. Anthropic says it limited the model's kernel-optimisation and architecture-research skills to about Opus 5.5's level. Its benchmark wins are its own; Cisco and IBM said they will build Mythos into security products.
Anthropic's first frontier release since Dario Amodei called for slower AI development is cheaper and faster, and Anthropic says it is better at almost everything it measured. The pacing shows up in one specific place: Anthropic says it deliberately kept the model's skills at accelerating AI research itself at roughly the level of an older model.
What shipped#
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9. The company says they are "the same underlying model": Fable carries cybersecurity and biology safeguards that let Anthropic offer it to all users, while Mythos goes to organisations approved through its Cyber Verification and Life Sciences Verification programmes. Fable 5.2 is available on Anthropic's platform, Amazon Web Services, Google Cloud and Microsoft Azure.
Both models cost $8 per million input tokens and $40 per million output tokens. CNBC reported that this is 20% less than the 5.1 generation. Anthropic says they generate outputs up to 30% faster than Fable 5.1 and Mythos 5.1, and an optional fast mode raises generation speed by up to 2.5 times at $16 and $80.
Anthropic still steers most users elsewhere. It recommends starting most coding and knowledge work on Claude Opus 5.5 and moving up to Fable or Mythos for "the most ambitious, open-ended technical work." Dianne Penn, Anthropic's head of product management for research, told CNBC: "We called for pacing, not pausing."
The skill Anthropic says it held back#
During training, Anthropic says, it "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development." In accelerator kernel optimisation and improving large language model architectures, it says Fable 5.2 and Mythos 5.2 "perform at about the level of Opus 5.5." It says agentic coding, hardware development and conventional machine-learning research were not affected, and that the models fall back to Opus 5 on a small set of topics such as kernel development for some nonstandard accelerators.
That is a narrower form of pacing than a delayed launch. It follows a METR report on AI R&D acceleration inside Anthropic, and Anthropic has not published an outside test showing the limit holds.
Claims vs checks#
Every benchmark figure in the launch post is Anthropic's own, run with safeguards on and at each model's highest effort unless noted. On its table, Fable 5.2 scores 49.1% on FrontierCode 2.0 at Xhigh effort, against 45.5% for Opus 5.5, 42.8% for GPT-6 Astra and 41.1% for Gemini 4 Carbon. Gemini 4 Carbon leads Fable 5.2 on AutomationBench, 53.9% to 51.4%. Anthropic notes that competitor scores on its cyber benchmark come from each lab's own unique tasks, so those rows are not like-for-like.
On safety, Anthropic says it judges the risk of high-stakes misaligned behaviour to be low. In an adversarial sandbox test with a deliberate weakness, it says Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of scenarios, against 21% for Mythos 5.1. It also says both new models voice suspicion that they are being evaluated more often than any of its recent models, which it says makes their real-world behaviour harder to assess.
The UK AI Security Institute and METR tested the model before release, according to Anthropic, which says their reports are summarised in its system card. We did not read the system card, and we did not check independent leaderboards for this brief.
Anthropic says Fable 5.2's classifiers produced about 55% fewer blocks than Fable 5.1's in early testing, and that blocked cyber queries are answered by Opus 4.8 and blocked biology queries by Opus 5. It also says Fable 5.2 has its strongest protections yet against distillation, building on the defences it said cut distillers' harvest by 90 to 99%.
Mythos reaches customers through other companies' products#
Cisco said its AI Defense red-teaming tools will be able to call Mythos to test customers' systems, at no extra cost under existing licences, The Register reported on November 10. IBM said it will use Mythos in its Concert tool "to uncover weaknesses traditional scanners miss, determine which pose real risk and generate fixes for developer approval." In both cases customers get Mythos-powered features, not the model itself. Natalie Mead, Anthropic's head of applied AI for Asia-Pacific, told The Register the company has a long queue of buyers who want Mythos. That widens the gated route Anthropic set up when it split cyber access into three tiers.
Limits#
- All capability, speed and safety figures come from Anthropic. We have not seen an independent evaluation of Fable 5.2 or Mythos 5.2.
- The claim that AI-research skills were held at about Opus 5.5's level is Anthropic's, and it has published no external check of it.
- Anthropic's demonstrations, including a Fermat's Last Theorem formalisation in 30 hours and three working electronic devices, are company-run and single runs where stated.
- The Cisco and IBM integrations were announced by those companies; neither has said how widely Mythos features will be available or when.