Skip to main content
This documentation covers Anthropic’s original performance take-home challenge—a real-world optimization test that evolved alongside Claude’s capabilities.

​
The Evolution

Anthropic’s performance take-home has an interesting history that reflects the rapid advancement of AI capabilities:

​
Original 4-Hour Challenge

The challenge originally started as a 4-hour take-home with the baseline code starting at 147,734 cycles. Candidates were given this starting point and asked to optimize it as much as possible.

​
Updated 2-Hour Challenge

After Claude Opus 4 began outperforming most humans at the 4-hour version, Anthropic updated the challenge:
  • Reduced to a 2-hour time limit
  • Provided optimized starter code at 18,532 cycles (7.97x faster than original baseline)
  • Added better debugging tools and more detailed instructions

​
Current Status

After Claude Opus 4.5 exceeded the best human 2-hour performance (achieving 1,790 cycles in a casual session), Anthropic moved to a different base for their time-limited interviews.
This repository contains the original baseline version (147,734 cycles) with the improved tooling from the 2-hour version. It’s no longer used as a time-limited test, but you can still use it to impress Anthropic’s recruiting team!

​
Why This Challenge?

The performance take-home tests several critical skills:
  • Low-level optimization: Understanding instruction-level performance
  • Algorithm analysis: Identifying bottlenecks in tree traversal and hashing operations
  • Creative problem-solving: Finding novel approaches within tight constraints
  • Validation discipline: Ensuring correctness while pursuing aggressive optimizations

​
What Makes It Interesting

This challenge is unique because:
  1. Real constraints: You’re optimizing for a simulated machine with specific instruction costs
  2. Clear metrics: Cycle count provides unambiguous performance measurement
  3. Multiple optimization levels: From basic improvements to near-theoretical-minimum solutions
  4. AI benchmark: Your performance is directly comparable to state-of-the-art AI systems
If you optimize below 1,487 cycles (beating Claude Opus 4.5’s best performance at launch), Anthropic wants to hear from you! Email [email protected] with your code and resume.

​
Getting Started

Ready to take on the challenge?

Task Details

Learn what you need to optimize and the rules of engagement

Benchmarks

See performance tiers and submission requirements

Build docs developers (and LLMs) love

Get started for freeTalk to us