Extended thinking is one of Claude's most distinctive capabilities — and one of the most misunderstood. It's not just "thinking longer." It's a fundamentally different mode of processing.
What Extended Thinking Actually Does
When extended thinking is enabled, Claude can allocate additional internal computation before generating a response. The model produces a chain of reasoning — visible to the user — that works through the problem step by step.
This isn't cosmetic. The internal reasoning process actually changes the model's final answer. On complex math problems, for instance, extended thinking can improve accuracy from ~60% to ~90%.
When to Use It
Extended thinking shines on tasks with these characteristics:
- Multi-step reasoning — problems where the answer depends on getting several intermediate steps right - Ambiguous inputs — situations where the model needs to consider multiple interpretations - Novel problems — tasks that require genuine analysis rather than pattern retrieval - High-stakes outputs — where accuracy matters more than speed
Prompting for Better Reasoning
You don't need to tell Claude to "think step by step" — extended thinking handles that automatically. Instead, focus on:
1. Clear problem framing. The better Claude understands what you're asking, the better it can reason about it. 2. Relevant context. Provide the information Claude needs without overwhelming it. 3. Explicit success criteria. Tell Claude what a good answer looks like.
The E-E-A-T Connection
Extended thinking is particularly valuable for content that needs to demonstrate expertise. By reasoning through problems rather than pattern-matching to answers, Claude produces responses that reflect genuine analytical depth — the kind of reasoning that distinguishes expert content from surface-level summaries.