Two ways of answering
A regular language model predicts the most likely next token, word by word (see module 2), and answers instantly as a result. A reasoning model inserts an extra step before the actual answer: it first generates a kind of intermediate deliberation – trying out approaches, checking intermediate results – before composing the final answer.
Why that takes longer
That intermediate step means more computation, and therefore more time – and often higher cost – per answer, typically seconds instead of milliseconds. For simple questions, that's pure waste; for complex, multi-step tasks, it noticeably lowers the error rate.
When the extra effort pays off
For simple tasks – summarizing, rephrasing, standard answers – a reasoning model usually brings no benefit, just higher cost and longer wait times. For complex, multi-step tasks – solving a calculation problem, finding an error in a spreadsheet, drafting a multi-step plan – the intermediate step noticeably reduces errors.
No guarantee against errors
Even a reasoning model can still be wrong – it still has no built-in truth check (see module 4). The difference is a lower error rate on complex tasks, not a guarantee of correct answers.
Why this matters for you as a decision-maker
"Our model reasons" is, at first, just a technical description – not an automatic quality promise. The right question: does the type of task actually fit a reasoning model – or are you paying for intermediate deliberation your actual task doesn't need at all?