MirrorCode Benchmark Exposes AI's Struggle to Rebuild Complex Programs from Scratch

Epoch AI and METR have released a new benchmark called MirrorCode to test AI models’ ability to rebuild entire programs from scratch. The benchmark includes 25 target programs across six languages, covering areas like bioinformatics, cryptography, and Unix utilities. Models must independently recreate these programs without seeing the original code, matching the exact output and passing hidden tests. MirrorCode presents a significant challenge for current AI systems. None of the models have yet succeeded with the largest tasks, classified as LARGE. These include complex programs such as a bioinformatics toolkit with 16,000 lines of Go code, which Claude Opus 4.7 reimplemented in 14 hours at a cost of $251. A human engineer would typically need 2 to 17 weeks to complete the same task without AI support.

Running one of the largest tasks cost $2,600 for a single attempt, with the AI working non-stop for 19 days. Despite failures on full reimplementation, models often pass 90% or more of the individual tests. Epoch AI has open-sourced the scaffold and 22 of the 25 target programs, providing 132 task instances for evaluation.

Claude Opus 4.7 currently leads the benchmark with a 56% success rate. GPT-5.5 follows at 44%, and Gemini 3.1 Pro Preview achieves 32%. Leading models from a year ago would have scored around 30% and struggled beyond simple programs like a calendar utility. The MirrorCode benchmark highlights the progress and limitations of AI in software development. Models can handle many tests but still fail to fully recreate the most complex programs. The results show a clear gap between current capabilities and the demands of large-scale, independent software reimplementation.