What happens when it "thinks"?
By default the model starts writing the answer immediately. In reasoning mode it first works something out for itself: it breaks the task down, examines options, checks intermediate results — and only then writes the answer. That working is usually not shown, only its result.
This is not a different model and not a higher intelligence. It is the same model, taking more steps. The principle has been known for a long time: ask a model to think step by step and it answers complex tasks more accurately. Reasoning mode builds that into the system, automatically.
The important change: today this is a dial
A year or two ago this was a product category: there were separate "thinking" models with their own names. That is no longer the case. On most current models it is a setting — providers call it effort, extended thinking or a thinking budget, but all of them regulate the same thing: how much work to do before answering.
If you read somewhere that "reasoning models are better than conventional ones", that text is out of date.
What are you paying for it?
This is where most cost estimates come apart.
Thinking steps count as output tokens. You do not see them, but they are on the bill — and output tokens are the expensive kind. On a hard question the invisible working can easily be ten times longer than the answer itself.
That is how two models with similar list prices can end up an order of magnitude apart in actual cost. There is only one sensible way to compare: run the same twenty real tasks on both and look at the total. A table of token prices will not tell you.
The other price is time. At high effort an answer takes not seconds but ten-odd seconds, sometimes minutes. In a conversational interface that is tolerable. In a customer service chat it is not.
When to turn it up
It is worth turning up when the task has several steps that build on each other:
- mathematics, calculation, unit conversion
- tracking down a bug where you need to find the cause, not the symptom
- analysing legal or contractual text with conditions and exceptions
- preparing a decision that weighs several criteria
- anything where a wrong answer costs more than the wait
When not to
It is worth leaving low when the task is single-step or formal:
- summarising, rewriting, translating
- categorising, tagging, extracting data
- drafting an email, writing a product description
- anything that runs many times a day
On these, high effort does not improve the result — it only adds cost and delay. A summary does not get better summarised because the model brooded over it first.
Two things it does not solve
It does not eliminate hallucination. With nothing to go on, longer thinking produces a better-constructed, more convincing error. That does not make it true. Supplying the material helps more than any effort level — see Why AI Makes Things Up.
It does not fix a bad prompt. If the instruction is missing the information that would make the task unambiguous, the model will simply brood longer over the same gap. A sharper question is worth more than a turned-up dial.
A test that costs five minutes
Pick a typical task from your own work. Run it at low and at high effort with the same prompt.
There are three possible outcomes:
- High is noticeably better → the task is complex, leave the dial up.
- No meaningful difference → run it low from now on. This is the most common outcome.
- High is worse → this happens too: the model overthinks a simple task and explains things you did not ask for.
This five-minute test is the highest-return thing you can do with an AI tool.
What to take away
Reasoning mode is not a "smarter setting", it is more work invested by the same model. Turn it up for complex tasks that run rarely. Leave it down for simple tasks that run often — there it only inflates your bill.