With how competitive the LLM field is, it would surprise me greatly if any of these players were doing anything other than trying to make the best possible product. I certainly do not believe they are intentionally training the models to use more tokens unnecessarily.