Higher benchmark scores don't mean lower cost. Qwen 3.8-Max and Claude Opus 5 both show it — and cost per successful task is ...