How long until the answer arrives, in human-felt time. Often stated as "p95 under 3 seconds" — meaning 95% of requests beat 3 seconds, a promise about the worst typical case, not the average.
Why it matters: latency is a product feature and a testable requirement. Agent chains multiply it — five model calls in a row is five waits stacked — which is a design constraint, not a surprise.
More in “Keeping it running in the real world”