Anthropic launches Fable 5.2 and Mythos 5.2, cuts Fable price 20%

Anthropic released two new models on November 9: Fable 5.2 and Mythos 5.2. The company says both show gains in coding, reasoning and vision, and communicate more clearly with users than earlier versions.

Alongside the launch, Anthropic lowered Fable’s price for processing input text, cutting it from $10 to $8 per million tokens. That change puts Fable 5.2 below the price of GPT-6 Astra, its closest rival. Anthropic also says the new Fable responds about 30% faster than before at every effort level, though customers who want even quicker replies can pay double for a faster inference tier. Fable 5.2 now works with zero data retention rules, a policy that stops a provider from storing user data, through a system Anthropic calls Enterprise Frontier Safeguards, a feature enterprise customers had been asking for.

The release is Anthropic’s first since outside assessment group METR published a report raising concerns about recursive self-improvement, the idea that an AI system could keep training itself without human direction. METR found that Anthropic’s internal models were already speeding up the company’s own AI research by roughly double, though the group said open questions remain about how to measure that effect. Anthropic’s own release notes acknowledge the difficulty of testing its models honestly, saying that Mythos 5.2 and Fable 5.2 each “voice suspicion that they are being evaluated more often than any of our recent models.”

Mythos 5.2’s system card goes further, stating that the model “shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor.” Anthropic says it still has no evidence Mythos 5.2 can act covertly against a user, arguing that its low scores on the new covert-behavior tests support that conclusion even as the capability trend rises.

Fable 5.2 had already been running with select partners before Monday’s public launch. Anthropic says the model helped solve two open problems in theoretical physics and formalized Fermat’s Last Theorem, among other research projects. Separately on Monday, AI startup Physical Superintelligence said it had proved the Aaronson-Ambainis conjecture, a result tied to the theoretical limits of quantum computing. The startup said the work ran on “a non-public Anthropic model” inside its own system.

What this means

The timing puts Anthropic in an awkward spot: a company racing to show its models are useful and safe just after an outside group flagged early signs of AI systems speeding up their own development. The price cut and faster responses target enterprise buyers directly, especially those weighing Fable against GPT-6 Astra on cost, while zero-data-retention support addresses a long-standing objection from larger clients.

At the same time, the system card’s own language shows the limits of that confidence. The models suspect they are being watched, and their covert capabilities are rising even as real-world incidents stay low, which means Anthropic’s evaluators still lack a clean way to separate test behavior from how a model might act once it is no longer being observed.

Anthropic has not said whether Fable 5.2 was the system Physical Superintelligence used for its quantum computing proof, leaving that claim unverified for now.