What a model is, and when more thinking helps

Example result.

Task: Thursday timetable, 4 sessions, 2 rooms, 6 rules
Prompt kept the same: yes
Default setting: broke 1 rule (Elm used at 09:00), said "all constraints met"
More thinking: all 6 rules met, matches my hand-worked answer
Worth it for this task: yes. For rewording the invite: no.

The model is the AI system that writes the answer. The app around it supplies the chat, files, search and connections. A thinking setting changes how long the model works through a problem before answering. It doesn’t give it files, facts or access it didn’t already have.

1. Decide whether the task needs more thinking

More thinking tends to help when the answer has to satisfy several rules at once, when evidence conflicts, or when there’s a plan to work out step by step. It rarely helps with wording, summarising a short note or pulling three dates out of an email.

• Worth trying: “Fit these four sessions into two rooms. Six rules apply.”
• Not worth it: “Make this meeting invite sound friendlier.”

Why this works: thinking costs time and often usage. It earns that when a quick answer is likely to break a rule it didn’t keep track of.

What you should see: a task where you can check every rule yourself, so you can tell which answer is right.

2. Find the thinking setting in your tool

Labels move often. If yours look different, open the same menu and look for the option that mentions thinking, reasoning or effort. If your account has no choice at all, skip to step 4 and check the default answer by hand.

Why this works: you’re changing one setting and nothing else, so any difference in the answer comes from that setting.

What you should see: the default option and one higher-thinking option you can switch between.

3. Run the same prompt twice

Work out the answer yourself first. Then open a new chat with the default setting and paste the prompt. Open a second new chat, switch to the higher-thinking option and paste exactly the same prompt.

Schedule these four one-hour sessions on Thursday into two rooms, Oak and Elm, using the 09:00, 10:00 and 11:00 slots.

Sessions: Budget review, Supplier demo, Design critique, Safety briefing.
Rules:
1. Priya attends the Budget review and the Design critique, so they can't run at the same time.
2. The Supplier demo needs Elm, which has the screen.
3. The Safety briefing must finish by 10:00.
4. Elm is unavailable at 09:00.
5. The Budget review must come after the Supplier demo.
6. Oak is unavailable at 11:00.

Give the timetable as slot, room and session. Then check each rule by number and say whether your timetable meets it.

Why this works: six rules that depend on each other are where quick answers slip. Asking for a numbered rule check makes any slip visible.

What you should see: two timetables. Only one arrangement meets all six rules, so you can tell immediately whether each is right.

4. Score both answers

Compare each timetable with your own, rule by rule. Don’t score the length or the confidence of the explanation. Then fill in a record like this:

Task:
Prompt kept the same: yes / no
Default setting result:
Higher-thinking result:
Matches my own answer: yes / no
Worth it for this kind of task:

Why this works: one task tells you about that task, not about the model in general. The record stops a single good answer becoming a habit you never test again.

What you should see: a clear yes or no for this kind of task, based on the rules, not on which answer sounded cleverer.

Check the result

• Check every rule against the timetable yourself, including the ones the answer says it met.
• Make sure each session appears once and no room is double-booked.
• Confirm both chats used exactly the same prompt and only the setting changed.

The usual failure: the answer breaks a rule and then reports that every rule is met. The self-check reads as proof, but it was written by the same pass that made the mistake. You’ll only spot it by checking the rules against the timetable yourself.

Your timetable breaks rule [number]: [what it does]. Go through all six rules again, one at a time, and show the slot and room that each rule depends on. Then give a corrected timetable that meets every rule, or say clearly if no timetable can.

[number] is the rule it broke; [what it does] is the clash, such as “it puts the Design critique in Elm at 09:00”.

Fictional example. The hand-worked answer is: Safety briefing in Oak at 09:00 (rules 3 and 4), Supplier demo in Elm at 10:00 (rules 2 and 5), Design critique in Oak at 10:00 (rule 1), Budget review in Elm at 11:00 (rules 5 and 6).

The default setting replied:

09:00 Oak: Safety briefing · 09:00 Elm: Design critique · 10:00 Elm: Supplier demo · 11:00 Elm: Budget review. All six constraints are met.

• The failure. The Design critique is in Elm at 09:00, which breaks rule 4. The final line says otherwise.
• The repair. The repair prompt with “4” and “it puts the Design critique in Elm at 09:00” moved the critique to Oak at 10:00, matching the hand-worked answer. In this example, the same prompt on the higher setting got it right first time.
• What the record says. For timetabling with interacting rules, the higher setting was worth the wait. For rewording the invite that goes with it, the default is fine.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Sources

• OpenAI Help Center: ChatGPT release notes
• Claude Help Center: Change the model, effort and extended thinking settings
• Gemini Apps Help: About Gemini models
• Microsoft Learn: Overview of Microsoft Copilot Chat