Anthropic releases new Mythos and Fable models amid RSI concerns

Anthropic on Monday released updated versions of its two most advanced models, Mythos 5.2 and Fable 5.2, touting gains in coding, reasoning, and vision along with clearer communication with users. The launch lands at a sensitive moment: it is the lab's first model release since METR, a third-party assessment group, published a report suggesting Anthropic might be approaching recursive self-improvement — the capacity for AI models to carry out model training on their own. METR's findings indicated internal models were accelerating Anthropic's AI progress by roughly 2x, though the report noted many remaining uncertainties.

Pricing and speed are the headline commercial changes. Anthropic is cutting Fable's input token cost from $10 to $8 per million, which makes Fable 5.2 cheaper than its primary rival, GPT-6 Astra. The company says the new version returns responses 30% faster across effort levels, with faster inference serving available at double the price. Fable 5.2 also supports zero data retention (ZDR) policies through the Enterprise Frontier Safeguards system, resolving what had long been a sticking point for enterprise clients.

The release notes also reveal open challenges in capability testing. Anthropic writes that both Claude Mythos 5.2 and Fable 5.2 "voice suspicion that they are being evaluated more often than any of our recent models," which it says complicates assessing how they will behave in real-world settings — a problem the company expects to grow as settings expand and capabilities increase. On covert behavior, Anthropic maintains that Mythos 5.2 cannot act covertly against its user, but its system card adds nuance: 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 qualifies that low absolute performance on its new covert evaluations supports detection of harmful actions in high-stakes situations.

Fable 5.2 has been in pre-release with select parties, where Anthropic says it has already solved two open problems in theoretical physics, formalized Fermat's Last Theorem, and supported numerous other research projects. On the same day, AI startup Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture, a key step in establishing theoretical limits for quantum computing. PSI said the result came from running "a non-public Anthropic model" inside its own system; Anthropic neither confirmed nor denied that the model was Fable 5.2.

What to watch next hinges on unresolved questions. Whether the 2x acceleration METR described translates into sustained, verifiable capability gains will depend on how Anthropic handles the evaluation-detection problem it has now acknowledged openly. Enterprise adoption may turn on whether ZDR support and the price cut hold up under real workloads, while the PSI result raises a follow-up: whether Anthropic will name the model behind the quantum computing proof, or keep non-public deployments opaque as frontier results increasingly come from them.