Anthropic Opens Original Performance Challenge to Developers

Anthropic has taken the unconventional step of open-sourcing its internal performance benchmarking test, creating a unique opportunity for developers to measure their optimization skills against the company's flagship AI models. The original performance take-home repository contains the same challenge Anthropic used internally to evaluate computational efficiency, now available for anyone to attempt.

The core challenge revolves around optimizing program execution on a simulated machine, with performance measured in clock cycles. Participants must write code that completes the task using fewer cycles than Anthropic's models achieved during testing. This isn't merely an academic exercise—Anthropic explicitly states that developers who surpass Claude Opus 4.5's best performance of 1487 cycles (achieved after 11.5 hours of optimization) should email their solution to [email protected] for potential recruitment discussions.

Benchmarks provided in the repository establish clear performance tiers:

• 2164 cycles: Claude Opus 4 after extended optimization
• 1790 cycles: Claude Opus 4.5 matching human performance in 2 hours
• 1579 cycles: Claude Opus 4.5 after 2 hours of focused optimization
• 1548 cycles: Claude Sonnet 4.5 with extended optimization time
• 1487 cycles: Claude Opus 4.5's best launch performance (11.5 hours)
• 1363 cycles: Latest Opus 4.5 in improved harness

This release provides rare insight into Anthropic's evaluation methodology. The company historically used such challenges to assess candidates' low-level optimization capabilities—skills critical for AI development where computational efficiency directly impacts cost and performance. By opening this process, Anthropic creates a public proving ground where developers can demonstrate skills that traditionally require formal interview processes.

The technical setup uses Python-based simulation. Participants clone the repository, implement their solution in submission.py, and validate performance by running python tests/submission_tests.py. The tests output which performance thresholds the solution achieves, providing immediate feedback on whether the submission beats specific model benchmarks.

What makes this noteworthy is its inversion of traditional recruiting. Instead of resumes prompting technical evaluations, demonstrated performance can initiate recruitment conversations. This approach could signal shifting norms in technical hiring, where verifiable skills outweigh credentials. For developers, it represents an unusual opportunity to showcase abilities directly to a leading AI research team without conventional application barriers.

Successful solutions will likely require deep understanding of algorithmic efficiency, low-level optimization techniques, and creative problem-solving within the simulated machine's constraints. The challenge remains active indefinitely, with Anthropic encouraging participants to push beyond the published benchmarks.