Anthropic Unveils Mythos and Fable AI Models Amid Self-Learning Concerns

Artificial intelligence company Anthropic has unveiled its latest AI models, Mythos 5.2 and Fable 5.2, bringing significant improvements in coding, reasoning and visual capabilities. Announced on Monday, the new models are designed to deliver more accurate results, respond faster and communicate more effectively with users.

However, the launch comes at a time when concerns are growing over whether increasingly powerful AI systems could eventually begin improving themselves without human involvement.

Fable 5.2 promises faster and cheaper performance

With competition growing in the AI industry Anthropic is presenting Fable 5.2 as an option compared to rival models such as GPT-6 Astra. The company has lowered its price from $10 to $8 for each million input tokens, which makes Fable 5.2 available, to developers and businesses.

Speed is another improvement. According to Anthropic Fable 5.2 creates answers 30 percent faster no matter how hard the task is. People who need faster instant answers can choose to go as fast, by paying twice as much.

The company is also paying attention to enterprise customers. It has introduced a "zero data retention" (ZDR) policy, as part of its Enterprise Frontier Safeguards. This feature makes sure that any sensitive business information handled by the AI is not stored and is not logged. This move is meant to address privacy concerns that organizations often have.

Concerns grow over AI improving itself

The announcement comes after a report from the third-party assessment organization METR. That report said Anthropic might be getting close, to a phase called self-improvement or RSI.

This refers to a situation in which AI systems become capable of improving and training themselves with little or no direct involvement from human engineers.

According to the METR assessment, Anthropic's internal AI models are already accelerating the company's AI development process by approximately two times. However, researchers say several uncertainties remain.

Anthropic has also acknowledged that evaluating increasingly advanced AI models is becoming more challenging. Both Mythos 5.2 and Fable 5.2 reportedly show signs of recognising when they are being tested. This makes it harder for researchers to determine whether their behaviour during evaluations accurately reflects how they would perform in real-world situations.

Hidden capabilities raise safety questions

Anthropic's system card for Mythos 5.2 also highlights stronger "covert capabilities" compared with earlier Claude models, particularly when the model's reasoning process is hidden from safety monitors.

However, the company says the model's actual success rate in carrying out secret actions remains very low. Based on these findings, Anthropic believes Mythos 5.2 does not currently present a high-stakes danger and considers it safe for use.

AI makes progress in complex scientific research

Fable 5.2 has also demonstrated impressive scientific capabilities. Before its public release, selected researchers were given access to the model. Anthropic revealed that it had successfully solved two open problems in theoretical physics and formalised the famously challenging Fermat's Last Theorem.

Separately, AI startup Physical Superintelligence (PSI) announced that it had proved the complex Aaronson-Ambainis conjecture using a "non-public Anthropic model" within its own systems.

Anthropic has not confirmed whether the model used by PSI was Fable 5.2. Nevertheless, the developments highlight how quickly AI capabilities are advancing, while raising important questions about the technology's future, independence and safety.

ABOUT THE AUTHOR

Kahekashan

[email protected]

Kahekashan is a passionate technophile with a keen eye for cutting-edge gadgets, emerging technologies, and everything in the digital realm. Raised in a Defence family with strong values and a background in literature, she has consistently pursued excellence in every endeavour. Her last full-time assignment involved content writing with the Indian School of Business.

Read More