diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2f0c57d30b..1b84e3f867 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -36,6 +36,7 @@ We use *breaking :warning:* to mark changes that are not backward compatible (re
 - [#6330](https://github.com/thanos-io/thanos/pull/6330) Store: Fix inconsistent error for series limits.
 - [#6342](https://github.com/thanos-io/thanos/pull/6342) Cache/Redis: Upgrade `rueidis` to v1.0.2 to to improve error handling while shrinking a redis cluster.
 - [#6325](https://github.com/thanos-io/thanos/pull/6325) Store: return gRPC resource exhausted error for byte limiter.
+- [#6399](https://github.com/thanos-io/thanos/pull/6399) *: Fix double-counting bug in http_request_duration metric
 
 ### Changed
 - [#6168](https://github.com/thanos-io/thanos/pull/6168) Receiver: Make ketama hashring fail early when configured with number of nodes lower than the replication factor.
diff --git a/pkg/extprom/http/instrument_server.go b/pkg/extprom/http/instrument_server.go
index 6c0e626267..745a5c8af7 100644
--- a/pkg/extprom/http/instrument_server.go
+++ b/pkg/extprom/http/instrument_server.go
@@ -76,7 +76,7 @@ func httpInstrumentationHandler(baseLabels prometheus.Labels, metrics *defaultMe
 
 						requestLabels := prometheus.Labels{"code": wd.Status(), "method": strings.ToLower(r.Method)}
 						observer := metrics.requestDuration.MustCurryWith(baseLabels).With(requestLabels)
-						observer.Observe(time.Since(now).Seconds())
+						requestDuration := time.Since(now).Seconds()
 
 						// If we find a tracingID we'll expose it as Exemplar.
 						var (
@@ -104,11 +104,13 @@ func httpInstrumentationHandler(baseLabels prometheus.Labels, metrics *defaultMe
 
 						if traceID != "" {
 							observer.(prometheus.ExemplarObserver).ObserveWithExemplar(
-								time.Since(now).Seconds(),
+								requestDuration,
 								prometheus.Labels{
 									"traceID": traceID,
 								},
 							)
+						} else {
+							observer.Observe(requestDuration)
 						}
 					}),
 				),