Visible reasoning can help a person inspect an answer without fully explaining how the model reached it. In its February 2025 account of Claude 3.7 Sonnet, Anthropic described extended thinking as giving the same model more computation before answering, with a developer-controlled budget. It also warned that models very often make decisions based on factors absent from their displayed reasoning. Those traces can expose intermediate mistakes or approaches worth checking, but Anthropic said monitoring current models' reasoning cannot support strong arguments about their safety. It presented this visibility as a research preview. Extended-thinking account.
