MirrorCode Benchmark Puts AI's Coding Skills to the Ultimate Test

Epoch AI and METR have launched a new benchmark called MirrorCode. It tests AI models on their ability to rebuild entire programs from scratch, without seeing the original code. The tasks cover a wide range of computer science fields, from Unix tools to cryptography and bioinformatics. MirrorCode includes 25 target programs across six programming languages. Together, they form 132 task instances. Epoch AI has open-sourced the framework and 22 of these programs for public use.

The benchmark demands high precision. AI-generated solutions must match the original program’s output exactly, even passing hidden end-to-end tests. Performance varies among leading models. Claude Opus 4.7 leads with a 56% success rate, followed by GPT-5.5 at 44% and Gemini 3.1 Pro Preview at 32%.

Some tasks push the limits of current AI capabilities. One of the largest required 19 days of continuous work and cost $2,600 for a single attempt. Despite this, no model has yet succeeded on the biggest challenges. However, Opus 4.7 managed to reimplement a bioinformatics toolkit with 16,000 lines of code in just 14 hours. MirrorCode sets a high standard for evaluating AI coding abilities. The results show progress, with some models handling complex tasks in hours. Yet the most demanding programs remain unsolved, highlighting the gaps in current AI development.