Claude Ultrathink: The Secret Thinking Levels Most Users Don't Know Exist

There’s a feature in Claude that most users don’t know exists.

It’s not hidden exactly—Anthropic has documented it. But unless you’re deep in the Claude Code ecosystem, you’ve probably never heard of it.

It’s called extended thinking. And at its highest level, there’s something almost mythical-sounding: ultrathink.

I spent a while digging into what these actually do, when they matter, and—importantly—when they don’t work at all. Here’s what I found.

What Extended Thinking Actually Is

When you ask Claude a question, it normally responds pretty quickly. It reads your prompt, generates an answer, done.

But some problems need more… thinking.

Extended thinking gives Claude permission to pause before answering. Instead of jumping straight to a response, it works through the problem step by step, explores different approaches, and considers edge cases—all before you see a single word.

Think of it like this: normal Claude is your colleague giving a quick answer off the top of their head. Extended thinking Claude is that same colleague saying, “Let me think about this for a minute” and actually doing that.

For complex problems—architecture decisions, tricky debugging, multi-step analysis—this extra thinking time often produces dramatically better results.

The Thinking Hierarchy: From “Think” to “Ultrathink”

Here’s where it gets interesting.

Claude Code (Anthropic’s command-line coding tool) has built-in keywords that trigger different levels of thinking. Each level allocates more “thinking budget”—essentially, how long Claude can reason before responding.

The trigger phrases are flexible. “Think about it,” “think deeply,” or “think more” all activate the megathink level. For maximum thinking, you can say “think harder,” “think really hard,” “think very hard,” or simply “ultrathink.”

What Happens When You Ultrathink

When you trigger ultrathink, Claude gets roughly 32,000 tokens to reason through the problem before responding.

That’s a lot of thinking space.

In practice, this means Claude can:

• Consider multiple solution approaches before committing
• Work through complex logic step by step
• Catch edge cases it might miss with faster responses
• Plan multi-file changes coherently
• Debug intricate issues by methodically tracing through possibilities

For a gnarly production bug or a “should I refactor this entire system” question, ultrathink can be the difference between a helpful answer and a genuinely insightful one.

The Important Caveat: This Only Works in Claude Code

Here’s what tripped me up initially, and I don’t want you to make the same mistake.

Ultrathink and extended thinking keywords only work in Claude Code—the terminal/CLI tool.

If you type “ultrathink” into Claude.ai’s web interface, nothing special happens. It’s just another word in your prompt. The same goes for the API—unless you explicitly configure thinking parameters in your API request, the keywords don’t trigger anything.

This matters because I’ve seen people recommending “just add ultrathink to your prompt” as a general tip. That advice only applies if you’re using Claude Code specifically.

For everyone else using the web chat or API without thinking parameters, focus on other prompting techniques. The keywords won’t help you.

When to Use Each Level

Not every task needs ultrathink. In fact, overusing it wastes time and—if you’re paying per token—money.

Here’s my rough guide:

Use “think” (4K tokens) when:

• Fixing a simple bug
• Writing a straightforward function
• Routine code refactoring
• Quick questions with clear answers

Use “megathink” (10K tokens) when:

• Designing an API structure
• Planning database schemas
• Performance optimization
• Reviewing code architecture
• Multi-step problem solving

Use “ultrathink” (32K tokens) when:

• Redesigning system architecture
• Debugging critical production issues
• Complex migrations affecting multiple systems
• Problems where getting it wrong has serious consequences
• You’ve tried regular approaches and need deeper analysis

The key is matching thinking level to problem complexity. Simple problems don’t get better answers from more thinking—they just take longer.

The Cost Reality

Extended thinking isn’t free.

Based on current Claude pricing:

• Basic thinking (~4K tokens): ~$0.06 per task
• Enhanced thinking (~10K tokens): ~$0.15 per task
• Maximum ultrathink (~32K tokens): ~$0.48 per task

For personal projects, these costs are usually negligible. But if you’re running Claude Code for a team or on lots of tasks, systematic ultrathink usage adds up fast.

Use ultrathink for problems that deserve it, not as a default.

How to Actually Use This

If you’re using Claude Code and want to try extended thinking:

For moderate complexity:

Think hard about how to refactor this authentication module.
Consider security implications and backward compatibility.

For maximum thinking:

Ultrathink about this architecture. We're migrating from a
monolith to microservices. I need you to identify service
boundaries, data flow, and potential failure points.

The keyword can go anywhere in your prompt—beginning, end, or middle. Claude Code’s preprocessing will catch it and allocate the appropriate thinking budget.

What About Claude 4 and Interleaved Thinking?

The latest Claude 4 models introduced something called “interleaved thinking”—Claude can now think between tool calls, not just at the start.

This is particularly powerful for agentic tasks where Claude is doing multiple steps: reading files, running commands, making changes. Instead of planning everything upfront and hoping for the best, Claude can reason about results as they come in and adjust its approach.

For complex coding tasks that involve multiple file operations, this makes extended thinking even more valuable.

The Bottom Line

Extended thinking is a real, useful feature—but only in the right context.

If you use Claude Code for complex development work, learning to trigger appropriate thinking levels can genuinely improve your results. Start with “think” for routine tasks, escalate to “megathink” for design work, and save “ultrathink” for the hard problems that really need it.

If you’re using Claude through the web interface or API, these keywords won’t help you. Focus instead on clear prompting, good context, and structured requests.

And either way, remember: more thinking isn’t always better thinking. Match the tool to the task.