Using Think Harder to Unlock OpenAI Codex's Hidden Reasoning
Overview
One extra word can unlock OpenAI Codex's hidden reasoning budget and give developers building with AI the planning superpowers. This recipe shows how to use special trigger words to allocate more compute for complex tasks.
What You'll Learn
How to use trigger words like "think harder" and "ultrathink" to get OpenAI Codex to allocate more computational resources to your requests, resulting in more thorough analysis and better solutions.
The Reasoning Tiers
According to Anthropic, OpenAI Codex allocates compute tiers based on trigger words:
- think - Basic enhanced reasoning
- think hard - More computational resources
- think harder - Significant reasoning budget
- ultrathink - Maximum reasoning allocation
How to Use
Simply prepend any of these phrases at the start of your prompt:
think harder: Design a scalable microservices architecture for an e-commerce platform
ultrathink: Debug this complex race condition in my multi-threaded application
When to Use
This technique is particularly effective for:
- Architecture discussions
- API design decisions
- Complex debugging scenarios
- Edge-case analysis
- Performance optimization problems
- Algorithm design challenges
Example Usage
Basic Request
How do I implement user authentication?
Enhanced Request
think harder: How do I implement a secure, scalable user authentication system with JWT tokens, refresh token rotation, and rate limiting?
The enhanced version will trigger more thorough analysis, considering security implications, edge cases, and implementation details.
Compatibility
This technique works across multiple platforms:
- OpenAI Codex CLI
- OpenAI Codex web interface (codex.ai)
- Cursor IDE
- Windsurf IDE
- Any interface using OpenAI Codex's API
No special flags or configuration needed!
Tips & Best Practices
- Use for complex problems that benefit from deeper analysis
- Not necessary for simple, straightforward tasks
- Combine with clear, detailed problem descriptions
- Use "ultrathink" sparingly for the most challenging problems
Real-World Applications
- Debugging Complex Issues: When standard models fail to solve a problem
- System Design: For comprehensive architecture planning
- Code Reviews: To catch subtle bugs and optimization opportunities
- Refactoring: When planning major code restructuring