Anthropic has released its Mythos and Fable models in an update that arrives only days after an independent evaluation report from the METR group suggested the lab may be approaching so-called recursive self-improvement: the ability of AI models to direct the training of their successors. The news was first reported by TechCrunch on November 9.
According to the report, Fable 5.2 was already available to selected partners in a pre-release program. The company says the model helped solve two open problems in theoretical physics, formalized Fermat's Last Theorem, and supported several other research projects. These are the lab's own claims, and they have not yet been independently verified.
What stands out most is the context of the launch. Until now, discussion of recursive self-improvement has largely stayed within academic papers and theoretical debate. When a lab with frontier models releases new versions at the same moment an outside evaluator raises that risk, the question stops being hypothetical for the market and starts to weigh on adoption decisions.
For teams running systems in production, the practical consequences are less speculative than they might seem. It is worth tracking the terms of use, scope limits, and safety policies that accompany each release, because a more capable model changes the risk profile of automations that depend on code or on decisions made without human review. Trampolim's view is that adoption should come with in-house testing and monitoring, rather than relying solely on the benchmark figures published by the vendor.