Status

The measured success rate and latency of every engine we sample, over the last 30 days. A tool that sells measurement should show whether its own instruments are working, so this is public. A failed sample is excluded from your scores and counted here instead, never shown as a zero.

A 30-day rate cannot on its own tell you whether an engine is failing now or has already recovered, so the last column names the date of the most recent failure and how many clean samples have run since. A low rate with a stale failure date is a problem we already fixed.

Success rate is completed samples over attempted samples per engine. Latency percentiles are over successful samples only, because a failed request is not a measure of how fast an engine answers. Figures refresh every few minutes.