One thing that stood out in the Step 5 Preview writeup is how much of it is about actual software work, not just "can it write code." StepFun's examples put a lot of emphasis on bug fixes, feature changes, and refactoring rather than just greenfield generation. That tracks with how most teams actually work: you spend more time reading an existing codebase and making a careful change than you do spinning up something from scratch. Whether that holds up is the interesting part. The useful test isn't another coding leaderboard. It's what happens when people drop it into a real repo and leave it on a long task.