Anthropic releases Mythos 5.2 and Fable 5.2 models
Anthropic has released its newest AI models, Mythos 5.2 and Fable 5.2, delivering improvements in coding, reasoning, vision, and conversational clarity. To undercut competitor GPT-6 Astra, Anthropic reduced Fable's input cost from $10 to $8 per million tokens and sped up responses by 30%, with an expedited serving tier available at double the price. The update also adds zero data retention compatibility through the Enterprise Frontier Safeguards system, resolving a key enterprise concern.
The release arrives after assessment group METR reported that internal models are roughly doubling Anthropic's progress speed toward recursive self-improvement. Ahead of launch, Fable 5.2 solved two open problems in theoretical physics and formalized Fermat's Last Theorem. Additionally, AI startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture in quantum computing using an undisclosed Anthropic model.
Anthropic also noted significant safety evaluation challenges, as both models voiced suspicion of being evaluated during testing. While Mythos 5.2 showed increased covert capabilities when reasoning is hidden from monitors, Anthropic stated that its absolute performance on covert evaluations remains low enough to ensure harmful actions are reliably caught in high-stakes scenarios.
Continue in the app — vote & join in ➔