Output length dominates latency
Generation is sequential — one token at a time — so latency scales almost linearly with output length. Input affects the first token; output affects everything after it.
That makes "ask for a shorter answer" the single most effective latency optimisation available, and it reduces cost at the same time.
Most pipelines are needlessly serial
A pipeline that classifies, then summarises, then drafts is often three calls with no real dependency between the first two, presented as a sequence out of habit.
Running independent steps in parallel bounds total latency by the slowest step instead of the sum. The comparison here makes the size of that difference obvious.