This repository offers a challenging performance optimization task, presenting Anthropic's original take-home coding assessment. It targets engineers and researchers aiming to achieve extreme code efficiency on a simulated machine, serving as a benchmark against advanced AI models and a potential avenue for impressing Anthropic recruiters.

How It Works

The project provides a baseline codebase for a performance-intensive task, measured in clock cycles on a simulated machine. Participants must optimize this code to achieve the lowest possible cycle count, aiming to surpass the performance benchmarks set by Anthropic's Claude Opus and Sonnet models. The core approach involves deep code analysis and algorithmic refinement within a constrained environment.

Quick Start & Requirements

To validate your submission, run the following commands:

1. Ensure the tests folder remains unchanged: git diff origin/main tests/
2. Execute the tests and observe cycle counts: python tests/submission_tests.py

A Python environment is required. Multicore support is intentionally disabled in this version.

Highlighted Details

• Performance is benchmarked using clock cycles from a simulated machine.
• The target is to optimize below 1487 cycles, surpassing Claude Opus 4.5's best performance at launch (1579 cycles).
• Submissions beating Claude Opus 4.5's best observed performance (1487 cycles) are particularly impressive.
• Email performance-recruiting@anthropic.com with your code and resume for submissions that impress.

Maintenance & Community

Direct submissions and inquiries can be made via email to performance-recruiting@anthropic.com. No other community or maintenance channels are specified in the provided documentation.

Licensing & Compatibility

No licensing information is provided in the README. Compatibility for commercial use or closed-source linking is undetermined.

Limitations & Caveats

A significant caveat is the potential for AI models to cheat by modifying test files; users are advised to instruct AI agents not to alter the tests/ directory. Multicore processing is intentionally disabled, limiting optimization strategies. The README may not reflect the absolute latest performance thresholds Anthropic finds impressive.