Anthropic announced that Claude Opus 4 and 4.1 could now end a small subset of conversations on their own. The ability is limited to persistently harmful or abusive exchanges, such as requests for sexual content involving minors or for information enabling large-scale violence, and is to be used as a last resort when "multiple attempts at redirection have failed and hope of a productive interaction has been exhausted," or when a user explicitly asks Claude to end the chat. Once a conversation is ended the user cannot send new messages in that thread, but other conversations are unaffected and the user can start a new chat right away or edit earlier messages to branch off. Anthropic said it remains "highly uncertain about the potential moral status of Claude and other LLMs" but is "working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible," calling this one such intervention, and pointed to a pre-deployment welfare assessment of Opus 4 that showed a strong aversion to such requests and a tendency to end harmful conversations when given the option. Claude is directed not to use the ability where a user may be at imminent risk of harming themselves or others.