Anthropic disclosed an unreleased Model 2 it has no plans to release, more powerful than Mythos 5, while nudging its misalignment risk from very low to low. Its August Risk Report admits the lab’s AI R&D evaluations have saturated, with Claude now authoring most code merged into its own production repositories.
Watchers did the arithmetic: Model 2 beat Mythos 5 by 12.5 points on a benchmark whose 85% threshold marks researcher replacement. 2027 is the takeoff.