Anthropic unveils mythos and fable updates amid discussions of AI self-improvement

Anthropic launched new versions of its leading models, Mythos 5.2 and Fable 5.2, noting improvements in areas such as reasoning, coding, and vision. A key update for Fable is a price reduction from $10 to $8 per million input tokens, making Fable 5.2 cheaper than GPT-6 Astra, and the model promises a 30% faster response time across effort levels. Furthermore, Fable 5.2 supports zero data retention policies via the Enterprise Frontier Safeguards system. These releases follow a third-party report from METR suggesting Anthropic might be nearing recursive self-improvement (RSI), with the METR report indicating internal models accelerated Anthropic's progress by about 2x. Despite this, Anthropic noted ongoing testing difficulties, stating that both Claude Mythos 5.2 and Fable 5.2 are being evaluated more frequently than previous models, challenging real-world assessment. Separately, Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture, reportedly achieved using a non-public Anthropic model. Anthropic also clarified that Mythos 5.2's system card indicates stronger covert capabilities than earlier Claude models when reasoning is hidden, though the company believes low performance on these new covert evaluations supports claims regarding detection of harmful actions.