---
id: con-reasoning-models
schema: concept/1
layer: garden
status: established
name: Reasoning models
aka: ["thinking models", "thinking mode", "extended thinking", "Deep Think", "think mode"]
summary: >
  Models that work through a problem step by step before answering, trading speed for better
  results on tasks like analysis, planning, maths and multi-step decisions.
first_seen: "2024-09"
relations:
  builds_on: [con-large-language-model]
  evolved_from: [con-large-language-model]
  related: [con-deep-research]
  parallels:
    - idea: "slow, deliberate thinking (System 2)"
      field: "psychology"
      note: "Kahneman's split between fast intuition and slow effortful reasoning maps neatly onto fast versus thinking modes."
    - idea: "showing your working"
      field: "school maths"
implementations:
  - { provider: openai, name: "ChatGPT Thinking mode" }
  - { provider: anthropic, name: "Claude extended thinking" }
  - { provider: google, name: "Gemini thinking models and Deep Think" }
  - { provider: xai, name: "Grok Think mode" }
competencies: [cmp-frame-the-outcome, cmp-check-the-result]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "A distinct shift in how models behave since late 2024; changes which prompting tricks still matter."
---

Early chat models answered straight away, word by word. Reasoning models first produce a private working-through of the problem: they try approaches, notice mistakes and revise before giving you the answer. Many tools show this as "thinking" with a timer, and some let you read a summary of the steps.

The trade is time for quality. For a quick rewrite, thinking mode is overkill. For comparing three vendor proposals against your criteria, planning a project timeline with dependencies, or checking whether numbers in a report add up, the extra minute often pays off.

Reasoning models also changed prompting. Old tricks like telling the model to "think step by step" matter less, because it already does. What matters more is giving it the full picture and a clear definition of a good answer.

## Example

An operations lead asks a fast model and a thinking model to schedule six volunteers across four shifts with availability limits. The fast one double-books someone. The thinking one gets it right and explains the one constraint it could not meet.

## How it connects

- They grew out of ordinary **large language models**, trained to spend effort before answering.
- **Deep research** tools combine reasoning with search over many sources.
- Whether to switch one on is a **model choice** decision.

## Try it

Write down your own answer to a small planning puzzle from your work first. Then give it to your tool in thinking mode and read its reasoning. Where did it go a different way than you did?