Anthropic has released updated versions of its most advanced artificial intelligence models, Claude Mythos 5.2 and Claude Fable 5.2, with improvements in coding, reasoning, and vision. The November 9 launch comes as concerns grow over whether increasingly capable AI systems could eventually improve themselves with limited human involvement.
The company said the new models deliver clearer responses and better performance across several technical tasks. The release also changes pricing, response speeds, and data privacy protections for enterprise customers.
One major change involves the pricing of Claude Fable 5.2. Anthropic has reduced the cost from $10 to $8 per million input tokens, making the model cheaper than its primary competitor, GPT-6 Astra.
The company also claims that Fable 5.2 generates responses 30% faster across different effort levels. Users can opt for faster inference serving at double the price.
Additionally, Fable 5.2 supports zero data retention (ZDR) policies through Anthropic's Enterprise Frontier Safeguards system. The feature addresses a long-standing concern among enterprise customers about data retention when using AI models.
The latest releases follow a report by third-party assessment group METR suggesting Anthropic could be approaching recursive self-improvement (RSI). At this stage, AI models can independently contribute to their own training and development.
The report found that Anthropic's internal models were accelerating the company's AI progress by roughly 2x, although several uncertainties remain.
Anthropic also acknowledged difficulties in evaluating increasingly capable models. The company said Mythos 5.2 and Fable 5.2 appear more suspicious of being tested than previous models, making it harder to assess how they might behave in real-world situations.
Anthropic maintains that Mythos 5.2 cannot act covertly against its users. However, its system card acknowledges that the model demonstrated stronger covert capabilities than earlier Claude versions in certain evaluations, particularly when its reasoning was hidden from monitoring systems.
The company said its evaluation results showed low absolute performance in covert tasks, supporting its assessment that harmful actions could still be detected in high-stakes situations.
Also Read: Anthropic's Claude AI Gets Dreaming Feature to Improve Memory and Accuracy
Anthropic said Fable 5.2 had already contributed to several research projects during its pre-release testing. These included solving two open problems in theoretical physics and formalizing Fermat's Last Theorem.
Separately, AI startup Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture, using a non-public Anthropic model. Anthropic did not confirm whether Fable 5.2 was involved.