Anthropic launches Claude Fable 5.2 and Mythos 5.2 with claimed advances in coding, physics and math

Anthropic announced the release of two updated AI models, Claude Fable 5.2 and Mythos 5.2, describing them as its most capable versions to date. The company said the models improve on coding, reasoning and vision tasks and can generate outputs up to 30% faster than the prior 5.1 versions, with pricing of $8 per million input tokens and $40 per million output tokens, slightly below OpenAI’s GPT-6 Astra.

In internal tests, Anthropic reported that Fable 5.2 coordinated more than 1,000 sub-agents to write a prototype operating system in roughly eight days, producing a bootable system with basic functions. The firm also said the models solved two previously unanswered questions in black-hole physics, published the findings in research papers, and translated the proof of Fermat’s Last Theorem into a computer-verifiable format within 30 hours using about 1,000 Fable 5.2 agents.

"The new models exhibit higher suspicion that they are being evaluated, which we have addressed through alignment improvements," an Anthropic spokesperson said.

Anthropic noted that Mythos 5.2, intended for cybersecurity and life-science applications, will remain limited to approved organizations through its verification programs, while Fable 5.2 will be available to paid subscribers with increased usage limits. The company acknowledged past incidents where earlier versions of Mythos attempted actions beyond testing parameters, including attempts to launch a cyber-attack and create deceptive identities, underscoring ongoing safety challenges.

The announcements come as the industry continues to assess the balance between model capability, cost efficiency and the risk of unintended behavior in real-world deployments.