A tier above Opus leaks and is deleted

The archived draft introduces Mythos as “a new tier of model: larger and more intelligent than our Opus models”, reports dramatically higher scores than Claude Opus 4.6 on software coding, academic reasoning and cybersecurity, and lays out a slow, cyber-defender-first early-access release because the model is large and expensive to serve (archived post, saved by M1Astra). The same paragraph introduces Mirendil, which announced itself as a frontier lab focused solely on systems that excel at AI R&D (launch thread), and the author frames it as betting on the scaffolding-over-weights lesson of the Agentica result in the same issue. It follows Anthropic’s own statement eleven days earlier that recursive self-improvement is a present phenomenon; the leaked tier resurfaces in April with the report that Mythos sped internal research by up to 400x.