Explore AI
What "reasoning" actually means
A real advance, badly named, and often the wrong tool for the job.
Some current models are described as reasoning models, and the framing invites a misunderstanding it is worth clearing up.
They are not thinking. They are spending more computation before answering.
The mechanism
An ordinary model answers immediately: your text goes in, a response starts coming out.
A reasoning model produces intermediate text first — working, backtracking, checking — and only then gives you an answer. The visible result is usually a longer wait, a higher bill, and often a better answer on hard problems.
That is the entire innovation. It is a real one, and the name oversells it.
Where they are genuinely better
Problems with a verifiable answer. Mathematics, logic puzzles, code that either runs or does not. This is where the extra computation pays for itself most clearly, because the model can check its own work against something objective.
Multi-step tasks where an early mistake compounds. Planning, debugging, anything where the first step determines the rest.
Anything where being wrong is expensive. The extra cost is trivial next to a wrong decision.
Where they are worse
Simple requests. Asking one to summarise an email means paying for deliberation you did not need, and waiting for it.
Creative and stylistic work. Rewriting, tone, drafting. There is no right answer to reason toward, and the extra machinery often makes the output more stilted.
Fast conversation. The delay is real. For chat, it is frequently annoying.
Cost-sensitive, high-volume work. Classification and extraction do not need deliberation, and you would be multiplying a large bill for no gain.
The honest caveat
When a reasoning model shows you its working, that text is generated just like any other text. It is not a readout of an internal process. It is a plausible-looking explanation produced by the same mechanism that produces everything else.
Which means: the visible reasoning is not guaranteed to be the reasoning that determined the answer, and it is not a proof that the answer is right. A confident chain of steps ending in a wrong conclusion is exactly as easy to produce as a correct one.
This is not a scandal. It is a reason to keep checking rather than trusting the working.
When to use one
A short test that works well:
Could someone check my answer against something objective?
If yes — maths, code, logic — reach for a reasoning model. If the answer is a matter of judgement, taste or wording, an ordinary model is usually faster, cheaper and better.