Latency

How long a user waits, measured as time to first token and total time, set with percentiles rather than averages.