MirrorCode Benchmark Reveals AI's Growing—but Still Limited—Coding Mastery

Epoch AI and METR have launched a new benchmark called MirrorCode to test advanced AI coding abilities. The challenge requires models to rebuild entire programs from scratch across various computer science fields. Unlike simpler tests from a year ago, MirrorCode pushes AI to handle complex, real-world software tasks without seeing the original code. MirrorCode includes 25 target programs covering areas such as Unix utilities, cryptography, and bioinformatics. These span six programming languages and total 132 task instances. Epoch AI has open-sourced the framework and 22 of the programs for public use.

The benchmark demands exact reproduction of a program’s output, including passing hidden end-to-end tests the model never sees during development. Even when models fail to fully reimplement a program, they often pass over 90% of its tests.

Performance results show significant progress in AI coding. The Claude Opus 4.7 model leads with a 56% success rate, followed by GPT-5.5 at 44% and Gemini 3.1 Pro Preview at 32%. Opus 4.7 also rebuilt a 16,000-line bioinformatics toolkit in 14 hours. However, no model has yet succeeded with the largest tasks, some of which took 19 days and cost $2,600 per run to attempt. MirrorCode demonstrates how far AI coding has advanced in a year. Leading models now tackle complex software projects with growing accuracy. Yet the benchmark also highlights remaining gaps, as even the best performers struggle with the most demanding tasks.