They’re training the models on generated slop. I’ve done a comparison on samples from the different versions of Claude and they’re worlds apart. The ancient, supposedly obsolete models speak normal English.