Since 09:12 UTC the API has GET /monitors/{id}/response-times: the average, median (p50) and 95th percentile (p95) of a monitor's response time, per hour from the rollups kept for a year (up to 90 days per call) or per 15 minutes from the raw checks of the last 14 days (up to 7 days per call). Every monitor object carries response_ms, the median over the voting probes of the latest round that came back up, and the monitor detail adds avg_ms for the last 24 hours, 7 days and 30 days, next to the uptime windows. Out-of-range input answers 422 with the field named, like every other validation error.
At the same time, response times everywhere (the chart on the monitor page, the hourly rollups and the API) are now measured by the voting probes only. The three shadow probes, which check monitors at one minute and slower but do not vote on outages, used to be included, so a European site measured from Mumbai or Sydney pulled its p95 up. Stored hours since 30 October were recomputed with the same rule. Checks, failures, uptime and credits are unchanged: they count every probe, as before.