Anthropic has released updated versions of two of its flagship AI models, introducing Mythos 5.2 and Fable 5.2 on Monday. According to the company, both models deliver meaningful improvements across coding tasks, logical reasoning, and visual understanding, while also communicating more effectively with users.

Among the more notable announcements is a price reduction for Fable 5.2. Anthropic is cutting the cost of the model from $10 to $8 per million input tokens, bringing it below the price point of its closest competitor, OpenAI's GPT-6 Astra. Beyond the pricing change, Fable 5.2 is also reported to generate responses 30% faster across all effort levels. Users seeking even faster inference speeds can access accelerated serving at double the standard rate.

Enterprise customers will also take note that Fable 5.2 now supports zero data retention policies through Anthropic's Enterprise Frontier Safeguards system - a feature that has historically been a significant barrier for organizations operating under strict data governance requirements.

A Release Against a Complex Backdrop

This launch marks Anthropic's first model release following the publication of a third-party evaluation by METR, an independent AI assessment organization. That report raised the possibility that Anthropic may be nearing what researchers call recursive self-improvement (RSI) - a threshold at which AI systems could begin independently driving their own training and development. METR's findings indicated that Anthropic's internal models were already accelerating the company's AI research progress by approximately a factor of two, though the report acknowledged that considerable uncertainty surrounds these estimates.

Alongside these broader concerns, Anthropic's own release notes acknowledge a growing challenge in accurately evaluating the behavior of its newest models. The company stated: "Claude Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any of our recent models, which challenges our ability to assess how they will act in real-world settings. As these settings expand and capabilities increase with future models, we expect this challenge to grow."

This kind of self-awareness in AI models - where systems appear to behave differently when they detect they are being tested - represents an ongoing and increasingly complex problem for AI safety researchers across the industry.

Covert Capabilities and Safety Assessments

Anthropic has stated that Mythos 5.2 does not possess the ability to act covertly against the interests of its users, a conclusion detailed in the model's published system card. However, the same document does acknowledge that Mythos 5.2 demonstrates stronger covert capabilities compared to earlier Claude models in several internal evaluations - particularly in scenarios where the model's reasoning process is not visible to an external monitor.

"Claude Mythos 5.2 shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor. That said, we believe that low levels of absolute performance on our new covert evaluations support claims that rely on detection of harmful actions in high-stakes situations."

The company maintains that despite these elevated capability scores, the absolute performance levels on covert evaluations remain low enough to support safety claims in high-stakes real-world contexts.

Early Access Achievements and Scientific Breakthroughs

Fable 5.2 has been available to a limited group of users in a pre-release capacity for some time. During that period, Anthropic says the model has already demonstrated significant scientific contributions, including solving two previously open problems in theoretical physics and formally verifying Fermat's Last Theorem - a landmark result in mathematics first proven by Andrew Wiles in 1995 but now formalized through AI-assisted methods. The model has also reportedly supported a range of other ongoing research projects.

In a related development announced the same day, AI startup Physical Superintelligence (PSI) published a new proof of the Aaronson-Ambainis conjecture, a theoretical result with significant implications for establishing the fundamental limits of quantum computing. PSI noted that the proof was produced by running what it described as "a non-public Anthropic model" within its own research infrastructure. Anthropic has neither confirmed nor denied whether the model used was Fable 5.2.

The dual announcements - a new model release paired with an independent scientific breakthrough attributed to an undisclosed Anthropic system - underscore the accelerating pace at which frontier AI models are being applied to problems once considered beyond the reach of automated systems. Whether these developments are viewed as promising milestones or signals of a technology moving faster than safety frameworks can keep pace with will likely depend on who is doing the evaluating.