Anthropic releases new Mythos and Fable models amid RSI concerns

Anthropic released new versions of its astir precocious models connected Monday, with the motorboat of Mythos 5.2 and Fable 5.2. The institution says the caller models “show improvements successful areas similar coding, reasoning, and vision,” and pass much intelligibly with users.

Notably, Anthropic is lowering the outgo of Fable from $10 to $8 per cardinal input tokens. As a result, Fable 5.2 is present cheaper than its superior rival GPT-6 Astra. Anthropic says that the caller mentation of Fable volition instrumentality responses 30% faster crossed effort levels. Faster inference serving is disposable for treble the price.

Fable 5.2 is besides compatible with zero information retention (ZDR) policies done the Enterprise Frontier Safeguards system, agelong a sticking constituent for galore endeavor clients.

It’s Anthropic’s archetypal exemplary merchandise since the accomplishment of a third-party study from the appraisal radical METR, which suggested the laboratory mightiness beryllium approaching recursive self-improvement (RSI) — oregon the quality for AI models to transportation connected exemplary grooming independently. The METR study recovered interior models were speeding up Anthropic’s AI advancement by astir 2x, though galore uncertainties remain.

Still, Anthropic’s merchandise notes important ongoing challenges successful accurately investigating exemplary capabilities. “Claude Mythos 5.2 and Fable 5.2 each dependable suspicion that they are being evaluated much often than immoderate of our caller models, which challenges our quality to measure however they volition enactment successful real-world settings. As these settings grow and capabilities summation with aboriginal models, we expect this situation to grow.”

Anthropic maintains that the caller mentation of Mythos does not person the quality to enactment covertly against its user, a verdict that’s elaborate further successful Mythos 5.2’s strategy card.

“Claude Mythos 5.2 shows stronger covert capabilities than anterior Claude models connected respective of our evaluations, peculiarly erstwhile its reasoning is hidden from the monitor,” the strategy paper reads. “That said, we judge that debased levels of implicit show connected our caller covert evaluations enactment claims that trust connected detection of harmful actions successful high-stakes situations.”

Fable 5.2 has been disposable to prime parties successful pre-release for immoderate time, and Anthropic says the exemplary has already “solved 2 unfastened problems successful theoretical physics, formalized Fermat’s Last Theorem, and supported galore different probe projects.”

Also connected Monday, AI startup Physical Superintelligence announced a caller impervious of the Aaronson‑Ambainis conjecture, a pivotal measurement successful mounting theoretical limits for quantum computing. Notably, the laboratory said the effect was achieved by moving “a non-public Anthropic model” wrong PSI’s ain system. Anthropic did not corroborate oregon contradict whether Fable 5.2 was the exemplary successful question.