I don't know that graph to me shows Sonnet 4.5 as worse than 3.7. Maybe the automated grader is finding code breakages in 3.7 and not breaking that out? But I'd much prefer to add code that is a different style to my codebase than code that breaks other code. But even ignoring that the pass rate is almost identical between the two models.