What is p95 latency and why does it matter more than the average?
p95 latency is the time the slowest one in twenty requests takes. Averages hide those: a system that answers in half a second most of the time and eight seconds occasionally has a fine average and an unusable feel. The tail is what users remember, and it is what decides whether they keep the habit.
The slowest 1 in 20. Averages hide exactly what users notice.
Why it matters to you
The number is only meaningful with three things attached: what was being measured, under what load, and on what date. A p95 from a laptop on office wifi with one user is not a number about production.
It also has to be measured rather than felt. OCTYN publishes Mooney's voice time as under five seconds end to end, and says plainly that it is a demonstrated wall-clock time from screen recordings on a Samsung S23 over a live network, not a p95, because per-stage timing is not instrumented in the app. Instrumenting it is the first thing that project's own write-up says it would do differently.
If a vendor quotes you a p95 and cannot say what date it was captured on, it is a feeling with a decimal point.
Where this turns up in something OCTYN runs
Each line carries the date it was recorded and where it came from.
- 2026-09-14
Voice to saved row is under five seconds, demonstrated on a Samsung S23 over a live network rather than instrumented
docs/growth/case-studies/mooney.md, section 4Mooney → - 2026-09-14
p95 voice latency is not instrumented in the app, and is named as a gap rather than estimated
no measurement in repoMooney →
If a proposal used this word and did not explain it, ask.
Book a call →