You can read the paper here: https://labs.scale.com/papers/swe_bench_pro

TL;DR its very effective as it directly tests model on REAL codebases: "The benchmark is constructed from GPL-style copyleft repositories and private proprietary codebases". The use case is very real.