A reasoning model is trained to work through a problem at length before committing to an answer.
The extra thinking happens at inference time, so quality can be bought with time rather than size.
Reinforcement learning on problems with checkable answers is what teaches the habit.
They help most on maths, code and multi-step logic, and much less on writing or recall.
You pay for every thinking token, so they are the wrong default for simple, high-volume work.
Chain-of-thought prompting showed that letting a model write its steps improves its answers.
Reasoning models take that finding and train the behaviour in rather than asking for it each time.
The model produces a long internal working-out, then a final answer, and only the answer is usually shown.
The training signal comes from reinforcement learning on tasks where correctness can be checked automatically.
Mathematics with a known result and code with a passing test suite are the obvious sources.
The model explores different approaches and is rewarded for the ones that reach the verified answer.
What it learns is not a single fixed procedure but the habit of checking, backtracking and trying again.
This changed how capability is bought: previously more capability meant a larger model.
Here the same weights get better results by being given more tokens to think with, at inference time.
That is what test-time compute means, and it scales along its own curve.
The gains are concentrated in domains with verifiable structure: competition maths, programming, formal logic.
On summarisation, drafting or straightforward retrieval the extra thinking buys very little.
The costs are real: latency of tens of seconds and token bills several times higher than a normal reply.
Hidden reasoning also complicates debugging, since the visible answer no longer shows the path taken.
The sensible pattern is a fast model by default and a reasoning model reserved for the hard minority of requests.
One question on this card. Every option explains itself once you pick, right or wrong.
22 concepts in this module, then a graded test with every answer explained.