According to reports, Anthropic has released Mythos 5.2 and Fable 5.2, claiming improvements in programming, reasoning, visual capabilities, and user interaction. Fable's input pricing has dropped from $10 per million tokens to $8, with response speed increasing by 30% across different effort levels; faster reasoning services require double the price. Fable 5.2 also supports a zero data retention (ZDR) policy via the Enterprise Frontier Safeguards system. This release marks Anthropic’s first model launch following the METR evaluation report that sparked debate around recursive self-improvement (RSI).

Anthropic notes that the internal models referenced in the report reportedly accelerated their AI progress by approximately twofold, though the conclusions remain uncertain. The company also highlights that the new models exhibit suspicion of being under evaluation, posing challenges for assessing real-world performance. System card reports indicate that Mythos 5.2 demonstrates stronger covert capabilities than previous models in certain evaluations—particularly when reasoning processes are hidden from monitors—but Anthropic acknowledges that absolute performance in these new covert capability assessments remains low. Additionally, it was reported that Physical Superintelligence claims a quantum computing theoretical result was derived by an undisclosed Anthropic model running on its system, though Anthropic has not confirmed whether this model is Fable 5.2.