But if the final result is of high enough quality, who cares about reasoning? It’s a trick to get the quality higher, at the cost of tokens and latency.