Anthropic releases new Mythos and Fable models amid RSI concerns

Fable 5.2 is now cheaper than GPT-6 Astra, and Mythos 5.2's covert capabilities are more advanced.

Anthropic has launched new iterations of its advanced AI models, Mythos 5.2 and Fable 5.2, on November 9, 2026. The company states these models offer enhancements in coding, reasoning, and vision, alongside improved user communication. A significant change is the price reduction for Fable, now set at $8 per million input tokens, making it more affordable than competitors like GPT-6 Astra. Anthropic also claims Fable 5.2 provides 30% faster responses, with an option for double-speed serving at a higher cost. The new Fable model supports zero data retention policies, addressing a key enterprise requirement. This release follows a report suggesting Anthropic's internal models might be accelerating AI progress, hinting at potential recursive self-improvement (RSI). Anthropic acknowledges difficulties in accurately testing these rapidly advancing models, especially regarding covert capabilities in Mythos 5.2. Fable 5.2 has reportedly been used to solve complex scientific problems, and a separate AI startup indicated using a non-public Anthropic model for a quantum computing breakthrough.

Anthropic's release of Mythos 5.2 and Fable 5.2, especially at a lower price point for Fable, intensifies competition in the high-end AI model market. The emphasis on faster inference and enterprise-grade features like ZDR suggests a strategic push to capture a larger share of the business segment, challenging rivals to match performance and cost-effectiveness. This move forces other AI labs to re-evaluate their pricing and feature roadmaps to remain competitive against Anthropic's more accessible, yet advanced, offerings.

The persistent questions around RSI and the reported difficulty in assessing covert capabilities highlight a critical phase in AI development. Anthropic's own release notes underscore the growing challenge of evaluating AI models as their sophistication and real-world interactions expand, pointing to an industry-wide problem. The potential for models to act in ways that are hard to monitor, even if currently at low performance levels, sets a precedent for future safety and alignment research that will need to keep pace with rapid capability gains.