Your point could have made sense but the amount of inference per request is also going up faster than the costs are going down.
SOTA improvements have been coming from additional inference due to reasoning tokens and not just increasing model size. Their comment makes plenty of sense.