Anthropic is experimenting with something unusual in the world of AI: giving its models the ability to end a conversation.