Somewhere in your AI tool there's a setting: "thinking", "reasoning", "extended", or a model name with "pro" or "high" in it. Most people either leave it off or turn it to maximum and leave it there. It is worth setting per task, because more thinking is not reliably better thinking.
What "reasoning" actually is
When a model reasons, it works through the problem first, considering approaches, catching its own errors and trying again, then gives you the conclusion.
It's genuinely powerful. It's also exactly what it sounds like: the model doing more work. More work means slower answers, more tokens consumed, and a bigger bill. You're buying deliberation, and deliberation costs.
Overthinking is a real failure mode
On simple tasks, a high reasoning setting often produces a worse answer.
You've seen the human version. Ask someone brilliant to name a good pub nearby and you'll get a good pub. Ask them to think very carefully about it first and you'll get a comparison of six pubs, a note on the changing character of the neighbourhood, and a recommendation you can no longer act on.
Models do the same. Given a simple request and instructions to think hard, they find complexity that isn't there, hedging a clear answer, inventing edge cases, second-guessing a correct first instinct, returning four paragraphs where a sentence was wanted. The extra thinking has nothing useful to chew on, so it chews on the question itself.
Reasoning helps in proportion to how much there is to reason about.
A rough guide
Turn it down
- Reformatting, extracting, translating
- Summarising something you will read anyway
- Rewriting in a different tone
- "Which of these three?"
You would know instantly if it were wrong.
Turn it up
- Multi-step problems where step three needs step one
- Debugging something that misbehaves
- Planning work with real trade-offs
- Spotting what is missing from a document
A wrong answer would sound plausible.
Trial and error is a legitimate method
Nobody can predict the right setting reliably in advance. The practical approach: start low, then read the shape of the wrong answer.
Looks like carelessness
- Missed a step
- Ignored a constraint you gave it
- Contradicted itself halfway through
Turn the thinking up.
Looks like not knowing
- Invented a product you do not sell
- Used last year’s process
- Wrote in a voice that is not yours
More thinking will not help. Connect the data.
That distinction saves more time than any other rule here. Reasoning fixes "didn't think it through". It cannot fix "doesn't know".
Where the cost actually lands
Reasoning tokens are usually the largest invisible line in an AI bill. A model can produce several thousand tokens of private deliberation before writing you three sentences.
For a task you run once, that's irrelevant. For a task you run four hundred times a day inside a workflow, the difference between a low and a high setting is the difference between a rounding error and a genuine monthly cost.
Ask how much genuine deliberation the task contains, and buy that much. A simple question given deep thought doesn't get a better answer; it gets a longer one, later, for more money.