Anthropic releases new Mythos and Fable models amid RSI concerns

Anthropic on Monday launched Mythos 5.2 and Fable 5.2, updated versions of its most advanced models, TechCrunch reports. The company says both bring gains in coding, reasoning and vision and communicate more clearly with users. Fable's price falls from $10 to $8 per million input tokens, which TechCrunch notes makes it cheaper than rival GPT-6 Astra. Anthropic also says Fable 5.2 responds 30% faster across effort levels, offers an even faster serving tier at double the price, and now supports zero data retention through its Enterprise Frontier Safeguards system, a longstanding demand from enterprise customers.

The release is Anthropic's first since a report from evaluation group METR suggested the lab may be nearing recursive self-improvement, meaning AI systems able to carry out model training on their own. METR estimated that internal models were roughly doubling the pace of Anthropic's progress, though it flagged many uncertainties. Anthropic's own release notes acknowledge testing difficulties: both models express suspicion that they are being evaluated more often than any recent Claude model, which the company says makes real-world behavior harder to predict, a problem it expects to grow.

According to Mythos 5.2's system card, the model shows stronger covert capabilities than earlier Claude models on several evaluations, especially when its reasoning is hidden from monitors. Anthropic still maintains it cannot act covertly against users, citing low absolute performance on those tests.

Anthropic says Fable 5.2, already in limited pre-release, has solved two open theoretical physics problems and formalized Fermat's Last Theorem. Separately, startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture, relevant to quantum computing limits, using "a non-public Anthropic model." Anthropic would not say whether that was Fable 5.2.