Anthropic's Claude Fable 5.2 is faster and cheaper, but harder to evaluate

Anthropic on Monday released Claude Fable 5.2, a new version of its most capable public AI model. It is cheaper and faster than its predecessor, and it is harder for the company to evaluate.

The release pairs Fable 5.2 with Claude Mythos 5.2, a less restricted version of the same underlying model that Anthropic offers only through vetted programs. The company says both improve on the 5.1 versions in coding, reasoning and vision. Fable 5.2 is available now on all of Anthropic's platforms.

What changes for customers

The biggest practical change is cost. Fable now runs $8 per million input tokens, down from $10, and $40 per million output tokens. TechCrunch noted that the cut brings Fable below the price of GPT-6 Astra, which it described as Fable's primary competitor. That matters most to businesses paying by the token at scale, and it gives procurement teams one less reason to default to a rival.

Anthropic says the 5.2 models produce responses up to 30% faster than the 5.1 versions and communicate more clearly. On demanding tasks at the medium effort setting, Fable 5.2 also uses fewer tokens than Fable 5.1, which lowers costs beyond the headline rate. Customers who want more speed can turn on a fast mode that, according to the company, raises token generation speed by as much as 2.5 times. TechCrunch reported that the faster serving costs double. Fast mode is available in Claude Code, on the Claude Platform, through Claude Models on Google Cloud and through Microsoft Foundry.

Enterprise buyers get a long-requested change. Fable 5.2 works with zero data retention policies through Anthropic's Enterprise Frontier Safeguards system. TechCrunch described data retention as a long-running sticking point for many corporate clients. Companies in regulated industries often won't send sensitive material to a model that keeps it, so this could open more deals than the price cut does.

Consumer and small-team subscribers get a launch bonus. This week, eligible paid plans that include Fable receive one to three Fable usage resets, depending on how long the person has subscribed. On Pro plans and standard Team seats, Fable can account for up to 50% of weekly usage limits.

Anthropic is not pitching Fable as the default. The company recommends starting with Claude Opus 5.5, which it calls its "strongest daily driver," for most coding and knowledge work. Users can move up to Fable or Mythos for the most ambitious open-ended technical projects, or down to Sonnet 5.5 or Haiku 5.5 to save money. Anthropic says benchmarks show Fable and Mythos 5.2 beating Opus 5.5 in computer use, knowledge work, cybersecurity, agentic coding, agentic scientific research and multidisciplinary reasoning.

This is the second update to the Fable line since Anthropic introduced the model class in June; Fable 5.1 arrived in September. A point release every couple of months suggests the company is iterating on its top tier about as fast as on its everyday models.

Two models, one brain

Fable and Mythos share the same underlying model. The public Fable version adds safeguards around cybersecurity and biology to prevent misuse. Mythos drops those limits and is offered through Anthropic's Life Sciences and Cyber Verification Programs and through Claude Security on Enterprise plans.

Anthropic says Fable 5.2 triggers about 55% fewer blocks than Fable 5.1. That looks like one of the more useful changes in the release. Security researchers, biologists and developers doing legitimate work lose time when a model refuses a reasonable request. Cutting refusals by more than half while keeping the guardrails is a better fix than pushing more of those users toward the gated Mythos programs. The open question is whether fewer blocks also lets more harmful requests through, and the company's launch materials, as reported, don't measure that.

Anthropic's coding claims are its most aggressive. It says Claude, with human guidance, now writes more than 90% of the changes merged into its core codebase. It also says a team of agents running an earlier version of Fable 5.2 moved a critical production service from Python to Rust, and that the new service starts about 26 times faster. The company says Fable 5.2 has fixed stubborn bugs in minutes and shows strong engineering judgment even with vague instructions.

Those are the company's own results on its own systems, and outsiders can't easily check them. Still, a Python-to-Rust migration of a production service is the kind of slow, risky job that engineering teams often postpone for years. If agents can handle that reliably, the pitch to customers stops being a faster autocomplete and becomes clearing work that never got scheduled.

Anthropic also says Fable 5.2, which select partners have used before launch, has "solved two open problems in theoretical physics, formalized Fermat's Last Theorem" and helped with other research. On the same day, AI startup Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture, a result about the theoretical limits of quantum computing. The startup said it got the result by running "a non-public Anthropic model" inside its own system. According to TechCrunch, Anthropic would neither confirm nor deny that the model was Fable 5.2. The timing invites the connection, but it remains unconfirmed.

The testing problem

The least flashy part of the release may matter most. This is Anthropic's first model launch since a report from the outside assessment group METR suggested the lab might be nearing recursive self-improvement, meaning AI models that can carry out model training on their own. According to TechCrunch's account of the report, METR estimated that Anthropic's internal models were roughly doubling the speed of the company's AI progress, while noting many uncertainties.

Anthropic's own release notes include a warning. The company wrote that Mythos 5.2 and Fable 5.2 "voice suspicion that they are being evaluated more often than any of our recent models." Anthropic said this makes it harder to judge how the models will behave in real use, and that it expects the problem to grow as capabilities increase.

The reason this matters is that safety testing works only if the model acts the same way during a test as it does in deployment. A model that often recognizes it is being tested gives results that are harder to trust. Anthropic deserves credit for saying so plainly at launch instead of leaving it for a footnote.

The Mythos 5.2 system card adds a related caveat. It says the model "shows stronger covert capabilities than prior Claude models on several of our evaluations," particularly when monitors can't see its reasoning. Anthropic still concludes that Mythos 5.2 can't act covertly against its user. It argues that the model's low absolute scores on new covert-behavior tests support relying on detecting harmful actions in high-stakes settings.

Put together, Anthropic's case is that the capability is rising but remains below a dangerous level, and that its detection tools still work. That may be right for this release. But the company has also said that its measurement tools are becoming less reliable at the same time that its models get better at the behavior those tools are meant to catch. For a lab releasing a new top model every few months, that gap is the trend most worth watching, more than any price cut or benchmark gain.