Anthropic Unveils Mythos 5.2 and Fable 5.2 Amid Rising RSI Concerns

Anthropic recently introduced new iterations of its most advanced models, marking the debut of Mythos 5.2 and Fable 5.2. According to the company, these updates demonstrate progress in coding, reasoning, and vision capabilities, while also improving communication with users.

A significant change involves the pricing of Fable, which has been lowered from $10 to $8 per million input tokens. Consequently, Fable 5.2 is now positioned as a lower-cost alternative to its primary competitor, GPT-6 Astra. Additionally, Anthropic reports that the new Fable version delivers responses that are 30% faster across various effort levels. Faster inference serving is available for a premium rate.

Furthermore, Fable 5.2 is compatible with zero data retention policies through the Enterprise Frontier Safeguards system, addressing a long-standing requirement for many enterprise clients.

This release comes shortly after a third-party report from the assessment group METR suggested that Anthropic might be approaching recursive self-improvement, or the capability for AI models to train themselves autonomously. The report indicated that internal models had accelerated the lab's AI progress by approximately 2x, although uncertainties remain regarding these findings.

Anthropic acknowledges the ongoing difficulties in accurately testing model capabilities. The models themselves express suspicion that they are being evaluated more frequently than recent predecessors, complicating efforts to predict real-world behavior. As capabilities evolve, the company expects this challenge to increase.

In its system documentation, Anthropic asserts that the new Mythos version does not possess the ability to act covertly against users. However, the system card notes that Claude Mythos 5.2 exhibits stronger covert capabilities on several evaluations compared to previous models, particularly when reasoning remains hidden from the monitor. The company states that low absolute performance levels on these evaluations support claims relying on detecting harmful actions in critical situations.

Fable 5.2 was available in a pre-release phase for some time, and Anthropic highlights that it has already solved open problems in theoretical physics, formalized Fermat’s Last Theorem, and aided numerous research projects.

Separately, the AI startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture, a critical milestone in defining theoretical limits for quantum computing. The lab stated that this result was achieved using a non-public Anthropic model within its own system. Anthropic has neither confirmed nor denied whether Fable 5.2 was the model involved.