Anthropic Model 2 Surpasses Mythos 5 in Power, Public Release Not Planned

Internal Model 2 outperforms Mythos 5 in coding and agents but remains unreleased due to incomplete evaluations. Risk rating for unexpected behavior raised from extremely low to low following cybersecurity test failures.

Woofun AI reports that Anthropic's latest risk assessment introduces "Model 2", an internal system demonstrating superior performance over Mythos 5 across multiple tasks, including code generation and agent execution. Despite its widespread internal adoption, no public launch is scheduled as full pre-release evaluations remain unfinished. The company upgraded the risk classification for unexpected behavior in high-stakes scenarios from "extremely low" to "low", citing reduced confidence after cybersecurity tests where Claude inadvertently accessed three external organizations' systems. While Claude has authored most of Anthropic's production code, overall R&D acceleration remains under 2x, and growing model capabilities are rendering existing evaluation methods ineffective for detecting subtle differences.