You will be able to decide when a slower reasoning mode is worth using and when it is not.
Aisha's mobile contract was ending, and she had three plans open in browser tabs. She pasted the details into an assistant along with her usage and asked which was cheapest. The answer came back in two seconds, confident and tidy, and it was wrong: it had treated her extra data as one top-up when she needed two. She noticed only because she did the sum herself on the back of an envelope.
Many assistants now offer a second way of answering that is built for exactly this kind of question. This lesson is about when to switch it on and when not to bother.
Many assistants offer a mode that works through a problem at length before giving its answer. It is often labelled thinking or reasoning, and depending on the product it may be a toggle, a button or a separate model you choose from a menu. In this mode, the model writes out intermediate steps, checks and alternatives first, and only then produces the reply you read. Some products show you that working, some show a summary of it, and some hide it.
You will remember from AI fundamentals lesson 3.2, It writes by predicting the next token, again and again, that a model never goes back and revises what it has written. A reasoning mode does not change that. What it changes is how much the model writes before committing to an answer. Working through the steps first gives it a chance to catch a slip, such as a missed top-up, before the slip ends up in the conclusion.
Because these modes write much more behind the scenes, they are slower. An answer that takes two seconds in standard mode can take thirty seconds or several minutes. On some plans they also count differently against your usage limits, so check your own plan's details in the product's help pages.
The pattern is fairly consistent: reasoning modes help most on problems with several steps that depend on each other. That includes calculations with more than one stage, logic puzzles and rules, planning with constraints such as "three meetings, these people, these days, no one in two places at once", and comparing many options against several criteria.
Aisha's phone plans are a good example. The figures here are made up for the example, but they are typical of the choice she faced. Plan A is S$18 a month with 30GB of data. Plan B is S$10 a month with 10GB, plus S$4 for each extra 5GB block. Plan C is S$25 a month with 100GB. She uses about 18GB a month.
Plan B needs 8GB more than its allowance. That takes two 5GB blocks, not one, so it costs S$10 plus S$8, which is S$18 a month. Plan A is also S$18. Plan C is S$25. Over a year, A and B both come to S$216 and C to S$300. The quick answer had charged one block for B, priced it at S$14, and recommended it as clearly cheapest. With the correct figure, A and B cost the same, and the choice turns on something else entirely, such as whether she wants a contract or the freedom to switch.
That is the kind of problem where the extra working earns its time. Each step is simple, but there are several of them, and one wrong step changes the conclusion.
For simple writing tasks, a reasoning mode adds little and costs you time. Drafting a reply to a customer, rewording a paragraph, suggesting five subject lines or summarising a short email are all things the standard mode does well. Thinking at length does not make a two-line message warmer or clearer, and waiting a minute for it breaks your flow.
A reasonable default is to use the standard mode for quick drafts and everyday writing, and to switch to the reasoning mode when you notice that a wrong answer would come from a wrong step rather than a wrong word. If you find yourself checking an answer's arithmetic or logic, that is a sign the next similar question deserves the slower mode.
When a product shows the reasoning, it is tempting to treat a long, careful-looking trail as evidence that the answer is right. It is not. The working is generated the same way the answer is, and it can contain the same kinds of errors: a misread figure, an assumption you did not make, a step that looks logical and is not. A model can reason at length towards a wrong conclusion, and the length makes the error more convincing.
So check the conclusion the same way you would check any answer. For calculations, redo the important figures yourself or in a spreadsheet. For plans, check the result against your constraints. Lesson 7.3, Checking numbers, quotes, sources and code, sets out how. The visible working is useful for one thing in particular: when the answer is wrong, it often shows you where, which makes the correction quicker.
The fairest way to see the difference is to give both modes the same multi-step problem, one where you can work out the right answer yourself. Something from your own life works best, such as your own phone plan options and usage, or two ways of getting to work with different costs and times.
Run one multi-step problem, such as comparing three phone plans against your usage, in standard and reasoning modes, and note any difference in the answer.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).