If they want to reduce token cost, just use a smaller model instead of dumbing down a more expensive.