Extended Thinking & Deep Reasoning

When thinking fast isn't enough

You're preparing a board presentation on whether to acquire a competitor. The data is ambiguous — strong revenue growth but declining margins, loyal customers but aging technology, great brand but key employees likely to leave. You need Claude to weigh all these factors, consider second-order effects, and give you a nuanced recommendation. Not a snap judgment — a considered one.

This is where extended thinking changes the game. Instead of generating an immediate response, Claude pauses, reasons through the problem step by step, and produces a significantly more thorough and accurate answer.

What extended thinking actually is

When you enable extended thinking, Claude doesn't just predict the next token immediately. Instead, it generates an internal chain of reasoning — working through the problem, considering alternatives, checking its own logic — before producing its final response.

You can see Claude's thinking process — it shows a summary of its reasoning chain so you can follow its logic. This is both transparent (you can verify its reasoning) and educational (you learn how Claude approaches problems).

When to use extended thinking

Extended thinking isn't always necessary — and using it for simple tasks wastes time and tokens. Here's when it shines:

There Are No Dumb Questions

Does extended thinking cost more?

Yes — the thinking tokens count toward your usage, so extended thinking uses more tokens than a standard response. The trade-off is worth it for complex tasks where accuracy matters more than speed.

Can I see what Claude is thinking?

You see a summary of Claude's reasoning process. On the API, you get the full thinking tokens. This transparency lets you verify Claude's logic and catch any flawed reasoning early.

How long does extended thinking take?

It varies by task complexity — from a few extra seconds to a minute or more for very complex problems. Claude allocates thinking effort proportional to the problem difficulty.

Extended thinking in practice

Example: Code architecture review

Prompt:

I need to redesign our notification system. Currently it's a monolithic
service that handles email, push, SMS, and in-app notifications.
We're hitting scaling issues at 10M notifications/day.

Think deeply about this. Consider:
- Microservice vs event-driven architecture
- Queue systems (SQS, Kafka, RabbitMQ)
- Retry and failure handling
- Template management
- User preference management
- Analytics and tracking

Give me an architecture recommendation with trade-offs.

With extended thinking, Claude will:

1. Analyze the current pain points (monolithic bottlenecks)
2. Compare architectural patterns for notification systems specifically
3. Evaluate each queue system's fit for this use case
4. Consider failure modes and their impact
5. Produce a cohesive recommendation that accounts for all the constraints

The result is a genuinely thoughtful architecture document — not a generic "use microservices" answer.

Example: Strategic decision

We're deciding between three market expansion strategies:
1. Enter the EU market with our existing product
2. Launch a new product line in our current US market
3. Acquire a smaller competitor with EU presence

Our company: $15M ARR, 50 employees, SaaS B2B, 18 months of runway.

Think through each option deeply. Consider financial impact, execution
risk, team capacity, and competitive dynamics. Recommend one and
explain why the others are riskier for a company our size.

How to get the best results from extended thinking

Extended thinking and Claude Code

In Claude Code, extended thinking activates automatically for complex tasks. When you ask Claude Code to:

• Debug a complex multi-file issue
• Design a new feature architecture
• Refactor with multiple constraints
• Solve algorithmic problems

Claude Code allocates additional thinking time proportional to the complexity. You'll see it working through the problem before making changes — which means fewer wrong turns and less backtracking.