Since 11:57 UTC, when the last of the eighteen probes moved to agent 2.7.0, the probes of a scheduled round no longer start at the same instant. Each begins after the previous one, by the site's own response time (400 ms to 1 second, learned from the site's fastest checks over the last day), and the shadow probes come after the voting ones. A site checked every five minutes used to see seven or eight of our requests in the same second; it now sees one at a time. The number of requests per minute is unchanged, and so is billing.
This also changes the response times we show, and it is worth being plain about why. On a site with a small worker pool, the probes that arrived together waited behind each other, and we reported that wait as the site's response time. One customer's site that answers a single visitor in about 0.6 seconds showed a median of 6 to 7 seconds across the probes; in the first staggered rounds its median was 1.5 seconds, and what is left is distance, the far probes' own round trips. From today the Response column, the 24-hour chart, the hourly figures and the response-times API report the site itself. Numbers stored before today keep the old measurement; we cannot recompute them, because the stored results are the queued times. If your charts drop today, that is the measurement being fixed, not your site getting faster.