Anthropic has announced the launch of two new advanced artificial intelligence models, Mythos 5.2 and Fable 5.2, featuring improvements in coding, reasoning, and visual capabilities, along with faster response times and lower usage costs. This comes amid growing concerns about the ability of AI systems to develop themselves independently.
The company stated that the new versions offer better performance in several areas and have become more capable of communicating clearly with users.
Fable 5.2 is cheaper and offers faster responses
The most prominent changes accompanying the launch of Fable 5.2 include a reduction in usage costs from $10 to $8 per million input tokens, making it cheaper than its main competitor, GPT-6 Astra.
Anthropic also confirmed that the new model provides 30% faster responses compared to the previous version across various processing levels. The company also offers a fast processing option for double the standard price.
Among other improvements, Fable 5.2 is now compatible with Zero Data Retention (ZDR) policies through the Enterprise Frontier Safeguards system, a feature that represents an important requirement for many corporate and enterprise clients.
Concerns about AI approaching self-improvement
The launch of the new models follows a report issued by the independent research group METR, which indicated the possibility that Anthropic is approaching the stage of Recursive Self-Improvement (RSI), the ability of AI systems to continuously train and improve their models independently.
The report found that the company's internal models contributed to accelerating the pace of its AI technology development by about twofold, despite the continued existence of many questions and uncertainty regarding the extent to which this development has been achieved.
In contrast, Anthropic highlighted the growing challenges associated with accurately evaluating the capabilities of its models, noting that both Mythos 5.2 and Fable 5.2 showed skepticism about being subjected to testing and evaluation more frequently than previous models.
The company believes this problem may become more complex as the use of the models expands and their capabilities increase in the future, as the model's awareness that it is being tested could affect the nature of its behavior and make evaluating its performance in real-world usage conditions more difficult.
Mythos 5.2 and increased capabilities to hide behavior
Regarding safety risks, Anthropic confirmed that the Mythos 5.2 release does not possess the ability to act secretly against the user, according to its evaluations published in the model's System Card.
However, the document also noted that the model demonstrated greater capabilities than previous versions in executing hidden behaviors that are difficult to detect, especially when the internal reasoning process is not visible to monitoring systems.
Nevertheless, the company believes that the limited levels recorded by the model in tests evaluating these capabilities support the possibility of relying on mechanisms to detect harmful behaviors in high-risk situations.
Fable 5.2 makes progress in physics and math research
The Fable 5.2 model was available to a limited number of entities before its official launch. Anthropic says that during that period, it managed to solve two open problems in theoretical physics and formulate Fermat's Last Theorem in a formal mathematical manner, in addition to contributing to several other research projects.
In a related development, the AI startup Physical Superintelligence announced a new proof for the Aaronson-Ambainis Conjecture, which represents an important step in determining the theoretical limits of quantum computing capabilities.
The company explained that it reached this result using an undisclosed Anthropic model run within its own system, while Anthropic neither confirmed nor denied that the model used was Fable 5.2.
