The problem it solves
A standard model produces its answer in essentially one forward motion, predicting the response directly. For many tasks that is fine. But for genuinely hard problems, multi-step math, complex logic, careful planning, tricky code, answering immediately tends to produce confident mistakes, because there was no room to work through the steps.
Extended thinking addresses this by letting the model spend more effort at the moment of answering. Given space to reason, break the problem into parts, try an approach, check it, revise, it solves things a single-pass answer would get wrong. The gain on difficult tasks can be large, which is why this became a distinct and important capability rather than a minor tweak.
How it works
The core idea has a name worth knowing: spending more computation at answer time, sometimes called test-time compute.
It thinks before it answers. Instead of going straight to a response, the model first generates intermediate reasoning, a worked-through train of thought, and uses that to reach a better final answer. This mirrors how a person handles a hard question: not blurting the first thing, but reasoning it out.
More thinking helps hard tasks, and wastes effort on easy ones. The extra reasoning meaningfully improves accuracy on complex, multi-step problems. But it also takes longer and costs more, because all that intermediate reasoning consumes tokens. On a simple question, it is pure overhead: you pay in time and money for deliberation the task never needed.
Thinking is becoming a dial. Rather than an all-or-nothing setting, systems increasingly let you control how much a model thinks, more effort for the hard cases, less or none for the easy ones. The exact controls differ across providers and change over time, so for current specifics the authoritative source is the relevant provider’s documentation. What is stable is the tradeoff: reasoning buys accuracy on hard problems at the price of speed and cost.
The practical judgment, then, is knowing when depth is worth paying for. Reaching for a reasoning mode on every task burns money and adds latency; never using it leaves accuracy on the table for the problems that need it.
A concrete example
Ask a model a knotty logic puzzle with several interlocking constraints.
Answering in one pass, a model often commits to a plausible-sounding solution that violates one of the constraints, because it never checked. With extended thinking, the model reasons through the constraints one by one, notices a conflict, backtracks, and arrives at an answer that actually satisfies all of them. For that puzzle, the extra thinking is the difference between right and wrong. Ask the same model “what is the capital of France,” though, and paying for extended reasoning would be a waste, the fast answer is already correct.
How it connects
Extended thinking is a way of using an LLM that trades more tokens, the intermediate reasoning, for better answers. It pairs naturally with an agent, whose multi-step tasks often benefit from deliberate reasoning between actions, and it interacts with the prompt, since how you frame a problem shapes how well the model reasons about it.
For the building roles this is a routine cost-and-quality decision. An AI/LLM Developer chooses when to enable it, and an AI Solutions Architect designs systems that spend reasoning where it pays off and skip it where it does not.