Google for Developers
13 minutes
Chain-of-thought is now something a model does rather than something you ask for: frontier models produce a reasoning trace on their own, and that trace is real, inspectable output you can read to see where an answer went wrong. What became obsolete was the typed instruction 'think step by step', not the phenomenon. The reasoning trace is therefore a monitoring surface: reading it lets you catch a flawed assumption, a skipped constraint, or a hallucinated fact before the final answer is used, and vendor tools expose it for exactly that purpose.