How do you know this? I'm not trying to attack your statement, I am genuinely curious how anyone knows anything about model performance outside of benchmarks that are already in the training set.