Anthropic announced Mythos 5.2 and Fable 5.2 on Monday, releasing new versions of its most advanced models. The company says the new model “features improvements in areas such as coding, reasoning, and vision,” allowing for clearer communication with users.
Notably, Anthropic is lowering the cost of Fable from $10 to $8 per million input tokens. As a result, Fable 5.2 is cheaper than its main competitor, the GPT-6 Astra. Anthropic says the new version of Fable provides 30% faster responses across each effort level. Faster inference services available at twice the price.
Fable 5.2 is also compatible with Zero Data Retention (ZDR) policies through the Enterprise Frontier Safeguard System. This has long been a problem for many enterprise clients.
This is Anthropic’s first model release since the arrival of a third-party report from evaluation group METR that suggests the lab may be moving closer to recursive self-improvement (RSI), the ability for AI models to perform model training independently. METR’s report found that internal models are accelerating Anthropic’s AI progress by about two times, although many uncertainties remain.
Still, Anthropic’s release points to significant ongoing challenges in accurately testing the model’s functionality. “We have expressed skepticism that Claude Mythos 5.2 and Fable 5.2 have each been evaluated more frequently than our recent models, challenging our ability to evaluate how they behave in real-world settings. We expect this challenge to grow even further as these settings expand and improve functionality in future models.”
Anthropic claims that the new version of Mythos does not have the ability to act covertly against users, a ruling detailed in the Mythos 5.2 system card.
“Claude Mythos 5.2 has demonstrated stronger covert capabilities than previous Claude models in several evaluations, especially when its inferences are hidden from monitors,” the system card states. “Nevertheless, we believe the poor absolute performance of the new confidential assessment supports the argument that it relies on detecting harmful behavior in high-stakes situations.”
Fable 5.2 has been available to some stakeholders in pre-release for some time, and Anthropic says the model has already “solved two open problems in theoretical physics, formalized Fermat’s Last Theorem, and supported numerous other research projects.”
Also on Monday, AI startup Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture, a crucial step in setting the theoretical limits of quantum computing. Notably, the institute said the results were achieved by running a “non-public human model” within PSI’s proprietary system. Anthropic would not confirm or deny whether Fable 5.2 was the model in question.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
