Token costs are not the whole story: measuring latency users feel
Dana Whitfield ·
A request that takes nine seconds but starts streaming at 400ms feels fast. A request that takes three seconds and streams nothing until the end feels broken. Total duration cannot tell those apart.
What we changed
Record time to first token separately from total duration.
Track it at p95, because the average hides exactly the requests people complain about.
Watch it per prompt version — a longer system prompt moves this before it moves cost.
The full walkthrough is in the documentation.