Extended thinking is Claude’s most powerful reasoning mode — and the one most people never use correctly. This guide explains what extended thinking does, when it genuinely improves outputs, how to enable it, and when you’re better off with a standard prompt.
What Is Extended Thinking?
Extended thinking gives Claude a dedicated reasoning phase before generating its final response. Claude works through a problem on “scratch paper” before writing its answer — exploring multiple approaches, identifying errors in its own reasoning, and building a more deliberate chain of thought. In Claude 4.6 models, this is called adaptive extended thinking — Claude dynamically adjusts how much thinking it does based on problem complexity.
When Extended Thinking Genuinely Helps
- Complex math and logic problems requiring step-by-step reasoning
- Multi-step coding tasks with many interdependent components
- Strategic analysis requiring weighing many variables
- Difficult research synthesis where accuracy matters most
- Any task where “think step by step” would help — extended thinking does this automatically
When Extended Thinking Is Overkill
- Simple factual questions with clear answers
- Routine writing tasks (emails, summaries, short copy)
- Format conversion or data transformation
- Tasks where speed matters more than depth
How to Enable Extended Thinking
In Claude.ai: Look for the thinking toggle before sending your message. Available on Max tiers and higher.
Via API: Pass "thinking": {"type": "enabled", "budget_tokens": 10000} in your request. Higher budget_tokens allows more thorough reasoning but increases latency and cost.
What You See During Extended Thinking
Claude shows a collapsed “thinking” section before its response. Expand it to see the reasoning chain — useful for verifying logic or understanding how Claude approached a problem. The thinking section is exploratory and may contain dead ends; this is normal.
Frequently Asked Questions
Does extended thinking always give better answers?
No. It improves accuracy on complex reasoning tasks but adds latency. For simple tasks, standard mode is faster and just as accurate.
Leave a Reply