What Is Extended Thinking
A feature giving AI more "time to think." Normal chat answers instantly, but in Thinking mode the AI builds reasoning step by step internally before answering. OpenAI's o1 / o3, Claude's Extended Thinking, Gemini's Deep Think apply.
When to Use It
| Use | Normal mode | Thinking mode |
|---|---|---|
| Casual question | ◎ | × (too slow) |
| Email draft | ◎ | × |
| Summary | ◎ | △ |
| Hard math/puzzle | △ | ◎ |
| Long-form coding | ○ | ◎ |
| Strategic analysis | △ | ◎ |
| Research review | ○ | ◎ |
How to Turn It ON
Claude
- Select a model with "Extended Thinking" (Opus etc.) in the model menu
- Turn Thinking ON with the switch beside the chat input
- Available on Pro / Team plans or above
ChatGPT
- Select "o1" "o3" in model selection
- These are always Thinking mode
- Available on Plus / Team / Enterprise
Gemini
- Deep Think mode (model menu)
- Available on Gemini Advanced
Prompt Tips to Draw It Out
1. Instruct Stages
Solve the following problem. 1) Organize the premises 2) Consider 3 solution methods 3) The best method and its reason 4) Final answer
2. Include "Self-Criticism"
List 3 problems in your first draft and give a final version fixing each
3. Visualize "Uncertainty"
For each conclusion, state confidence (high/mid/low) and the basis
Limits and Cost
- Slow: takes 30 sec to several minutes. Not for instant answers
- High cost: 5-10x the tokens of a normal model. For long text it can be hundreds of yen
- "Overthinking" problem: over-deliberating even on easy questions, wasting time and money
Pro Use Examples
- Strategy planning: "next year's business strategy in order: market analysis → competitor → own strengths → 4 options → recommendation"
- Code review: "point out all bugs, security, and performance issues in this code"
- Research planning: "based on this paper, 5 unsolved problems and next research-theme ideas"
- Hard-to-read contracts: "list all risk clauses in this contract and analyze each one's actual impact"
Using with Normal Mode
A workflow of tidying Thinking-produced conclusions in normal mode with "restate clearly," "summarize to 3 lines" is effective. Hard thinking in Thinking, communicable text in normal mode.
