---
id: thinking-modes
title: Thinking and reasoning modes
level: 2
modes: [day-pack, rucksack]
moment: "The person is choosing between a fast and a slower thinking or reasoning setting, is asking the AI to plan something with many steps, or is about to run a slow setting for a quick reword."
keywords: [thinking mode, reasoning, extended thinking, think deeper, effort, slower, fast mode, allowance, think harder, reasoning steps]
source: "research/2026-10-20-concepts/thinking-modes.md; research/2026-10-23-essentials/batch-6-facts.md; mapping (product/learning-system/MAPPING-1-2.md): FAQ 1.07 (home), 1.06; MISTAKES 1.6 (home); tips T084, T086, T087, T166; strand U"
---

## What it is
Some AI apps offer a slower setting that works through a problem in steps before it answers, and a quick setting that answers straight away. The slower one helps with jobs that have steps, such as planning a group trip around everyone's dates, doing maths or comparing options. It takes longer and in many apps uses more of your allowance, which is the amount of use your plan gives you. A quick setting suits simple jobs, such as rewording a message.

## Tips
- Keep the slower thinking setting for jobs with steps, like plans, maths and comparisons, because working through them first helps, though it takes longer and often uses more allowance.
- Check the answer itself, not the steps an app shows, because a long chain of steps is no proof the answer is right.

## Questions people ask
- Q: Why are some models smarter than others? A: A bigger AI with more training tends to do better, and apps release newer versions often, usually with a basic version and a more advanced one. A thinking setting can lift results on jobs with steps, because the AI works through them first.
- Q: Why is thinking mode slower? A: It works through a problem in steps, trying approaches and checking results, before it gives the answer. That takes longer and, in many apps, uses more of your allowance.
- Q: Is a thinking mode more accurate? A: On problems with several steps, often. On simple jobs like rewording, a quick setting can do as well, and neither setting guarantees the answer is right.
- Q: Can I trust the steps it shows me? A: They can leave out what drove the answer, and in some apps what you see is only a summary. Check the answer itself, especially the one fact that matters.
- Q: Does telling it to think harder work? A: Typing it may nudge one reply, but in some apps it does not change the thinking level. Where your app has a thinking or effort setting, use that instead.

## For your AI
Facts (checked 23 October 2026; fact check research/2026-10-23-essentials/batch-6-facts.md; research note
research/2026-10-20-concepts/thinking-modes.md).
- The settings: ChatGPT's slider (Instant to Pro; Free and Go get Instant and Think), where asking it to "think harder"
  does not switch the level on Plus and Pro (https://help.openai.com/en/articles/20001354-gpt-6-and-other-models-in-chatgpt);
  Claude's effort setting and thinking toggle, where higher effort takes longer and reaches usage limits faster
  (https://support.claude.com/en/articles/8664678-change-the-model-effort-and-thinking-settings); Gemini's Low,
  Medium and High, where higher levels use more of the allowance; Copilot's Think Deeper, slower but not limited on
  personal accounts.
- When it helps: on complex, multi-step problems; simple questions and general writing do not need it, and at low
  complexity standard models can do better (Apple, June 2025, https://machinelearning.apple.com/research/illusion-of-thinking).
- The steps shown are a summary in Claude's app, and in OpenAI's and Google's models through their developer tools, not
  the raw working; ChatGPT's and Copilot's apps do not say. Models mentioned a
  planted hint in their steps only 25 to 39 percent of the time (Anthropic, 3 April 2025,
  https://www.anthropic.com/research/reasoning-models-dont-say-think), so steps are not proof.
- Why some models are smarter: capability grows with the size of the model, the data and the computing power used to
  train it (Kaplan et al., 2020, https://arxiv.org/abs/2001.08361), then with training after that; companies offer
  tiers from cheaper and faster to more capable, and newer versions arrive often.

Teach it like this: suggest the slower setting for a plan, maths or a comparison, the quick one for a reword, and check
the answer, not the steps. Before saying it uses more allowance, know which app they are in: in Copilot's personal app
the slower setting is not limited.