Anthropic released a detailed postmortem identifying three overlapping changes—a reasoning effort downgrade, a prompt verbosity limit, and a caching bug—that degraded performance. The technical investigation found that a caching optimization accidentally cleared model memory during long sessions, causing Claude to lose track of its logic. All issues are resolved as of April 20, and usage limits have been reset for affected subscribers.