Reasoning models and extended thinking: when does a "thinking" model pay off?
Essence
Reasoning models (often called "thinking" models, with the mode called extended thinking or simply thinking) generate internal reasoning before answering: they break the task down, try options and check themselves. They're the same LLMs, but additionally trained to use a "scratchpad" — largely through reinforcement learning on tasks with verifiable answers, such as math and code. Quality on hard tasks improves noticeably, but you pay in tokens and waiting time.
The rest of this answer, with code and interview follow-ups, is in the DevPrep app.
Read the full answer in DevPrep