Early chat models answered instantly, in one pass, like someone blurting the first thing that came to mind. Reasoning models changed that. Before responding, they generate internal thinking (working through the problem, checking themselves, trying approaches), and only then produce the answer. On hard problems (maths, logic, code, multi-step analysis), the difference in quality is large.
The trade-off is time and money: thinking consumes tokens, so reasoning answers are slower and cost more. That is why products increasingly offer a choice, fast mode or thinking mode, or decide automatically based on the question.
Why you keep hearing it
Because reasoning was the big capability jump of the last couple of years, and every frontier lab now ships it. When you see an assistant show "thinking..." before it answers, or a setting called extended thinking or deep think, this is what is happening.
What it means for you
Match the mode to the task. Drafting a caption does not need minutes of deliberation; analysing a pricing decision or a messy spreadsheet does. If an AI's answer to a genuinely hard question feels shallow, check whether you asked a fast model to do a thinking model's job.