Article ·
Ultrathink is in the house
Claude Code has four thinking levels that control how much computational power it dedicates to solving problems. Each keyword triggers a specific token budget: think gets 4,000 tokens, think hard gets 10,000, and both think harder and ultrathink max out at 31,999.
These aren’t mystical incantations. They’re preprocessor triggers built into Claude Code’s command-line tool. The web interface and API don’t recognize them—they treat these words as regular text.
Basic Mode: “think”
This allocates 4,000 tokens for reasoning. Think of it as giving Claude a standard notepad to work through the problem.
Use this for:
- Planning straightforward refactors
- Debugging clear error messages
- Implementing defined features
- Writing basic unit tests
Intermediate Mode: “think hard”
At 10,000 tokens, Claude gets more room to explore alternatives and edge cases.
Use this when you need to:
- Compare implementation strategies
- Design modules with multiple dependencies
- Debug issues spanning several files
- Optimize performance bottlenecks
Deep Analysis: “think harder” or “think very hard”
Both trigger the maximum 31,999 tokens. Claude can now perform extensive analysis before responding.
Reserve this for:
- Architectural decisions affecting the entire codebase
- Mysterious bugs with no clear cause
- Complex refactoring across multiple systems
- Security vulnerability analysis
Maximum Power: “ultrathink”
Same token allocation as think harder (31,999), but it’s become the standard way to request maximum computational resources. Plus, it looks super cool in the terminal. It’s a nice little easter egg.
Deploy this for:
- Scalable system architecture design
- Technical debt analysis with comprehensive solutions
- Multi-system integrations
- Critical decisions where mistakes cost money
The Technical Reality
Claude Code runs a preprocessing layer that intercepts these keywords before sending prompts to the model. This layer directly sets the thinking token budget based on what it detects in your prompt. No preprocessing means no special behavior—so these keywords do nothing in the web chat or API.
The system scans for specific phrases. For maximum allocation, it looks for: “think harder”, “think intensely”, “think longer”, “think really hard”, “think super hard”, “think very hard”, and “ultrathink”. For intermediate allocation: “think about it”, “think a lot”, “think deeply”, “think hard”, “think more”, and “megathink”.
Strategic Usage
Match computational cost to problem complexity. Start with lower levels. Escalate when you need deeper analysis.
Before coding, ask Claude to make a plan. Tell it not to write code until you’ve reviewed the plan. This approach works best when paired with the right thinking mode for your problem’s complexity.
The difference between standard think and ultrathink can mean the difference between a quick patch and a strategic solution. But burning maximum tokens on simple problems wastes resources. Choose your level based on what the task demands, not what sounds impressive.
Remember: this feature exists only in Claude Code’s command-line interface. If you’re working through the web or API, focus on writing clear, detailed prompts instead. The thinking modes won’t help you there.
PS – Here’s a funny little video showing the thinking modes.