You may have noticed newer AI models pause and “think” before they answer, sometimes showing a summary of their working-out. These are often called reasoning models, and the pause is a real, deliberate feature, not a loading spinner. Understanding what it does explains why these models are better at hard problems, and why they can be slower and more expensive.
The old way: answer immediately #
A standard language model predicts its answer one chunk at a time and starts producing the final response right away. For simple questions this is perfect. But for a problem that needs several steps of reasoning (a math word problem, a tricky bit of logic, a multi-part coding task) answering immediately is like being forced to blurt out the answer to a hard question with no time to think. Sometimes you get it right, often you do not.
The new way: work it out first #
A reasoning model adds a step. Before writing its final answer, it produces an internal working-out: a kind of scratchpad where it thinks through the problem, considers approaches, catches its own mistakes, and plans. Only after that does it write the answer you see.
It is the difference between blurting and thinking before you speak. Given room to reason step by step first, the model solves problems it would otherwise fumble. This is why reasoning models post big gains on math, logic, and complex coding, and smaller gains on simple tasks where there was not much to think about anyway.
The “thinking” is still the same next-chunk prediction under the hood. The change is that the model is now allowed to spend that prediction on working the problem out privately before committing to a final answer, instead of having to produce the answer directly.
What “effort” controls #
Because thinking takes time and tokens, it is not free, so many reasoning models let you dial how much of it to spend. This is often exposed as an “effort” level, or a thinking budget: low, medium, high, and so on.
- Higher effort means the model thinks longer and harder before answering. On genuinely difficult problems this produces better, more careful results.
- Lower effort means it thinks less and answers faster and cheaper, which is ideal when the task does not need deep reasoning.
The effort setting is really a knob on the three-way tradeoff between quality, speed, and cost. Turn it up when correctness matters most; turn it down when speed or budget matters more.
When it helps, and when it is overkill #
More thinking is not always better. Matching the effort to the task is the actual skill:
- Worth the thinking: multi-step math, debugging subtle code, planning a complex task, careful analysis, anything where a wrong first instinct is likely and the steps matter.
- Overkill: simple lookups, quick rewrites, casual chat, formatting a list. Cranking effort here just makes you wait longer and pay more for an answer that was never going to be hard. A reasoning model set to high effort can even overthink a trivial question, second-guessing an answer that was fine the first time.
The practical move is to reach for high effort on the hard problems and keep it low for routine work, rather than leaving it maxed out and paying the thinking tax on everything.
How this connects to agents #
Reasoning pairs naturally with agents. An agent runs a loop of deciding what to do next, and better reasoning at each decision means smarter choices about which tool to use and how to interpret what came back. A lot of the recent jump in how capable agents feel comes from the underlying models being better at thinking before they act.
The takeaway #
A reasoning model is a language model that works the problem out before answering, and an effort setting is how you decide how much of that working-out to pay for. Use the thinking where the problem earns it, skip it where it does not, and you get the best of both: sharper answers on the hard stuff, and fast cheap answers on everything else.