Multi-step problems need multiple steps
Hard problems in math, coding, and logic typically can't be solved correctly by pattern-matching straight to an answer. They require working through a sequence of dependent steps, where getting an early step wrong derails everything that follows. A model that answers immediately has to get every one of those steps right implicitly, in a single pass, with no opportunity to notice or correct a mistake partway through.
Extended reasoning gives the model room to make that structure explicit. By breaking a problem into smaller sub-steps and working through them in sequence, a model is more likely to arrive at a correct final answer than it would by trying to produce the whole solution in one shot, because each step is a simpler, more tractable prediction than the full problem at once.