The conspiracy theorist in me says that LLM providers do this regularly (or at least, don't bother optimizing for it) beyond some arbitrary "$/task" metric. I am not sure of there is enough SOTA model competition to avoid this.