noun · AI concept
Reasoning model
also called thinking model
An AI that thinks before it speaks. It spends extra effort working a problem through before giving you the answer.
A model trained to generate extended step-by-step reasoning before its final answer, trading time and tokens for better results on hard problems.
A model trained to produce a long internal chain of reasoning before its visible reply, often with a setting for how much to spend. Those thinking tokens use real compute, typically count toward output limits and billing, and may be hidden or summarized in the response.
Real-life example
You send the same tricky scheduling puzzle to two models. The first answers instantly and gets it wrong. The second pauses for several seconds, then gets it right. It used a pile of hidden thinking tokens first, and you paid for those too, even though you never saw most of them.
Where the analogy breaks
-
Thinking is not free. The reasoning tokens take time, count toward the output limit and are usually billed like other output.
-
More thinking is not always better. On simple questions it only slows things down and costs more.
-
It does not remove hallucination. A model can reason its way confidently to a wrong answer, so check anything that matters.