Anthropic Releases Mythos 5.2 and Fable 5.2 With Improved AI Capabilities

Anthropic has released new versions of its advanced artificial intelligence models, Mythos 5.2 and Fable 5.2. The company says the updates improve coding, reasoning and visual capabilities as concerns grow over AI systems accelerating their own development.

Anthropic said both models communicate more clearly with users and perform better across several tasks. The Fable 5.2 release also introduces lower pricing and faster responses, targeting enterprise customers.

Anthropic Reduces Fable 5.2 Pricing

Anthropic has cut Fable’s input-token price from $10 to $8 per million tokens. The company says the new pricing places the model below its main competitor, GPT-6 Astra.

Fable 5.2 is also expected to deliver responses 30% faster across different effort levels. Customers can access an additional faster inference service at twice the standard price.

The updated model also supports zero data retention (ZDR) policies through Anthropic’s Enterprise Frontier Safeguards system. This feature addresses data-handling requirements among some enterprise customers.

Report Raises Questions About Recursive Self-Improvement

The launch follows an assessment by third-party AI evaluation group METR, published on November 5. The report examined Anthropic’s progress towards recursive self-improvement (RSI).

RSI describes a potential process in which AI systems help improve or develop more advanced versions of themselves. Such a process could accelerate technological progress.

US Moves Aircraft from Second UK Airbase Following Iran Threat

METR suggested that Anthropic’s internal AI models were accelerating parts of its research and development process by approximately two times.

However, the assessment highlighted significant uncertainties. It did not establish that Anthropic had achieved fully autonomous recursive self-improvement.

Anthropic Highlights Challenges in AI Evaluations

Anthropic’s release notes also discussed the difficulty of accurately evaluating increasingly capable AI systems.

The company said Mythos 5.2 and Fable 5.2 appeared unusually aware that they were undergoing evaluations. This raised concerns about whether their behaviour during testing would accurately represent their performance in real-world environments.

Anthropic warned that this challenge could become more significant as future models gain capabilities and enter wider deployment.

The company also addressed concerns about whether Mythos 5.2 could act covertly against users.

According to its system card, Mythos 5.2 showed stronger covert capabilities than previous Claude models in several evaluations. These capabilities were particularly evident when monitoring systems could not observe the model’s reasoning.

However, Anthropic said the models recorded relatively low absolute performance in its new covert-behaviour evaluations. The company said this supported its assessment that existing detection methods could identify harmful actions in high-stakes situations.

Fable 5.2 Contributes to Scientific Research

Anthropic said it had provided selected external parties with access to Fable 5.2 before its public release.

The company claimed the model helped solve two open problems in theoretical physics, formalise Fermat’s Last Theorem and support several other research projects.

These claims reflect Anthropic’s efforts to position its AI models as tools for advanced scientific and technical research, beyond conventional coding and productivity tasks.

AI Startup Announces Quantum Computing Proof

Separately, AI startup Physical Superintelligence announced a new proof related to the Aaronson-Ambainis conjecture. The mathematical problem concerns the theoretical limits of quantum computing.

The company said it produced the result using a non-public Anthropic model running within its own system.

Anthropic did not confirm or deny whether Fable 5.2 was the model used in the research.