I keep seeing and trying to work with newer models described as better because they’re faster, stronger at reasoning, better at coding, or better at benchmarks. None of that matters much if they cannot reliably follow instructions and use material they were…