Anthropic said on Nov. 9 it has launched new versions of its top AI models, Mythos 5.2 and Fable 5.2.

Anthropic said Mythos 5.2 and Fable 5.2 improve coding, reasoning and vision performance and communicate more clearly with users.

Fable input token pricing fell to $8 per 1 million from $10. As a result, Fable 5.2 is cheaper than a major rival model, GPT-6 Astra.

Fable 5.2 response speed is 30 percent faster across all reasoning intensities. Users can also use a faster reasoning service by paying double the price. It also supports a zero data retention (ZDR) policy through an enterprise frontier safeguard system. Many corporate customers had cited Anthropic's lack of support for this policy as a problem.

The two models are the first Anthropic has released since an assessment report by the evaluation organisation METR was published. The report suggested Anthropic's AI models could be approaching recursive self-improvement (RSI), in which models continue training themselves.

The report analysed that internal models are increasing Anthropic's AI development speed by about 2 times.

Anthropic said, "Mythos 5.2 and Fable 5.2 more often show suspicion that they are being evaluated than recent models," and added, "It is difficult to judge how the model will behave in real environments." The company said the problem is expected to grow as model capabilities increase.

Anthropic said it judges that Mythos 5.2 does not have the ability to take actions that harm users without their knowledge. According to the system card, Mythos 5.2 showed stronger covert behaviour capabilities than previous models in multiple evaluations. That was especially the case when it hid its reasoning process from monitoring models. Anthropic said it believes it can detect harmful behaviour in high-risk situations because absolute performance in a new covert evaluation was low.

Fable 5.2 was provided to some institutions ahead of its release. Anthropic said the model solved 2 unsolved problems in theoretical physics and formalised Fermat's Last Theorem.