Only if computation is the bottleneck. GPT-4.5 shows it’s not.