In an unprecedented experiment, 6 frontier AI models were tasked with building 52 different apps from identical prompts, with no editing allowed. The results reveal surprising strengths, weaknesses, and the evolving state of AI-generated code.