Anthropic Releases Latest Models "Claude Fable/Mythos 5.2"

Anthropic announced new AI models "Claude Fable 5.2" and "Mythos 5.2," which the company positions as its "most capable to date." These two models are offered in two distribution formats based on their intended use cases. One format is released for free public access, while the other is a limited edition available exclusively to organizations that have passed review.

In recent years, Anthropic has positioned the dual pursuit of AI safety research and commercial deployment as the core of its business strategy. This announcement coincides with a period when the company is reportedly considering an IPO. As performance competition in AI models intensifies, this marks an important opportunity for Anthropic to demonstrate the level of its AI capabilities to the broader market.

According to the company's explanation, the new models are equipped with capabilities to tackle unsolved physics problems and enhanced abilities in the cybersecurity domain. The models are also noted for their ability to design and build functioning operating systems. Particularly noteworthy is the achievement in code generation. Over 90% of the changes merged into Anthropic's core codebase are reportedly handled by these models.

Conversely, Anthropic has disclosed that it intentionally imposed limitations on the models' capabilities. Specifically, it intentionally suppressed functionality that would allow AI to accelerate its own development. This measure is framed as a safeguard against what is known as "recursive self-improvement"—the scenario in which AI autonomously continues to make itself progressively more capable.

The situation where AI writes the majority of code can be viewed as a case study demonstrating how the division of roles between humans and AI in software development is shifting. The over 90% code contribution ratio suggests that AI may be beginning to function not as an auxiliary tool but as a primary agent in development.

Attention will focus on how the two distribution formats will function going forward. The distinction between a free public version and a limited version for vetted organizations can be interpreted as a reflection of an approach to managing access to high-capability models in a staged manner. The question of how broadly to open access to AI's advanced capabilities continues to be debated across the industry, and this response represents one possible answer.

The announcement, which simultaneously emphasizes both a commitment to AI safety and the release of high-performance models, remains noteworthy from the perspective of what message Anthropic—on the eve of its potential IPO—is sending to both investors and society at large.