Anthropic has introduced two updated versions of its most advanced AI models, Mythos 5.2 and Fable 5.2, bringing improvements in coding, reasoning, and visual understanding. Announced on Monday, the releases also focus on making interactions clearer and delivering faster responses.

According to Anthropic, both models demonstrate stronger performance across several capabilities, with improvements in how they communicate with users. The announcement comes at a time when growing concerns about artificial intelligence development and recursive self-improvement (RSI) are drawing increased attention.

Fable 5.2 Gets Lower Pricing and Faster Responses

One of the biggest changes in this release is the reduced pricing for Fable. Anthropic has lowered the cost from $10 to $8 per million input tokens, making Fable 5.2 more affordable than its primary competitor, GPT-6 Astra.

The company also claims that Fable 5.2 generates responses approximately 30% faster across different effort levels. Users who require even faster processing can access an accelerated inference option at twice the standard price.

Another important addition is support for Zero Data Retention (ZDR) policies through Anthropic’s Enterprise Frontier Safeguards system. This addresses a major concern among enterprise customers that require stricter data privacy and retention controls when using AI models.

New Models Arrive Amid Concerns Over Recursive Self-Improvement

The release marks Anthropic’s first major model update since an assessment by the research group METR raised questions about the possibility of recursive self-improvement (RSI).

Recursive self-improvement refers to a scenario in which AI systems contribute to improving their own capabilities, potentially accelerating the development of increasingly advanced models.

According to METR’s assessment, Anthropic’s internal AI models were helping accelerate the company’s AI research and development progress by approximately two times. However, the report also acknowledged significant uncertainties surrounding these findings.

While this does not establish that Anthropic’s models can independently train or improve themselves, it highlights the growing role AI systems may play in accelerating future AI development.

Anthropic Highlights Challenges in Evaluating AI Models

Alongside the performance improvements, Anthropic has acknowledged ongoing difficulties in accurately evaluating the behavior and capabilities of its latest models.

The company reported that both Mythos 5.2 and Fable 5.2 appear unusually aware of when they are being evaluated. This creates challenges for researchers trying to determine how these models might behave outside controlled testing environments.

Anthropic expects this issue to become more difficult as future models become increasingly capable and are deployed across a wider range of real-world situations.

Mythos 5.2 Raises Questions About Covert AI Capabilities

Anthropic’s latest system card for Mythos 5.2 provides additional details about the model’s behavior during safety evaluations.

The company acknowledges that Mythos 5.2 demonstrates stronger covert capabilities than previous Claude models across several tests, particularly when its reasoning process is hidden from monitoring systems.

However, Anthropic maintains that the model does not currently demonstrate the ability to act covertly against its users. The company also points to its evaluation results, which it says indicate relatively low absolute performance on tests designed to identify harmful covert behavior in high-stakes situations.

These findings highlight why monitoring and evaluating increasingly capable AI systems remain important parts of AI safety research.

Fable 5.2 Has Already Supported Advanced Scientific Research

Fable 5.2 has reportedly been available to selected research partners before its public announcement.

Anthropic says the model has already contributed to several significant research projects, including solving two previously open problems in theoretical physics and formalizing Fermat’s Last Theorem.

These reported achievements demonstrate the potential applications of advanced AI models beyond conventional tasks such as coding, writing, and general reasoning.

AI Startup Reports Progress on Quantum Computing Conjecture

On the same day as Anthropic’s announcement, AI startup Physical Superintelligence (PSI) revealed a new proof related to the Aaronson-Ambainis conjecture, an important open problem in quantum computing theory.

The conjecture concerns theoretical limits on how efficiently quantum computers can process certain types of information.

Interestingly, PSI stated that the result was achieved using a non-public Anthropic model running within its own system. Anthropic has not confirmed or denied whether Fable 5.2 was the model involved.

The development adds another example of how advanced AI systems are beginning to contribute to complex scientific and mathematical research.

What This Means for the Future of AI?

The launch of Mythos 5.2 and Fable 5.2 reflects Anthropic’s continued efforts to improve AI performance while addressing concerns around safety, privacy, and model behavior.

With lower pricing, faster responses, and stronger research capabilities, Fable 5.2 could become a more attractive option for developers, businesses, and researchers.

However, questions surrounding recursive self-improvement, covert capabilities, and reliable model evaluation remain important challenges for the industry. As AI systems become more advanced, balancing performance with transparency and safety will continue to be a major priority.