diff --git a/.changelog/692.bugfix.md b/.changelog/692.bugfix.md
new file mode 100644
index 000000000..7f5fd3be5
--- /dev/null
+++ b/.changelog/692.bugfix.md
@@ -0,0 +1 @@
+metrics: dont create new 'api_request_latencies_buckets' for 404s
diff --git a/api/middleware.go b/api/middleware.go
index 759dc2641..d63c66e03 100644
--- a/api/middleware.go
+++ b/api/middleware.go
@@ -65,7 +65,6 @@ func MetricsMiddleware(m metrics.RequestMetrics, logger log.Logger) func(next ht
 			)
 			t := time.Now()
 			metricName := normalizeEndpoint(r.URL.Path)
-			timer := m.RequestLatencies(metricName)
 
 			// Serve the request.
 			next.ServeHTTP(w, r.WithContext(
@@ -74,14 +73,6 @@ func MetricsMiddleware(m metrics.RequestMetrics, logger log.Logger) func(next ht
 
 			// Observe results and log/record them.
 			httpStatus := reflect.ValueOf(w).Elem().FieldByName("status").Int()
-			if httpStatus < 400 || httpStatus >= 500 {
-				// Only observe the request timing if it's not going to be
-				// ignored (see below).
-				// Timers are reported to Prometheus only after they're
-				// observed, so it's OK to create it above and then ignore it
-				// like we're doing here if we don't want to use it.
-				timer.ObserveDuration()
-			}
 			latency := time.Since(t)
 			logger.Info("ending request",
 				"query_path", r.URL.Path,
@@ -103,6 +94,7 @@ func MetricsMiddleware(m metrics.RequestMetrics, logger log.Logger) func(next ht
 				metricName = "ignored"
 			}
 			m.RequestCounts(metricName, statusTxt).Inc()
+			m.RequestLatencies(metricName).Observe(latency.Seconds())
 		})
 	}
 }
diff --git a/metrics/requests.go b/metrics/requests.go
index dbd404293..e79f3c480 100644
--- a/metrics/requests.go
+++ b/metrics/requests.go
@@ -48,7 +48,9 @@ func (m *RequestMetrics) RequestCounts(endpoint, status string) prometheus.Count
 	return m.requestCounts.WithLabelValues(endpoint, status)
 }
 
-// RequestLatencies creates a new latency timer for the provided request endpoint.
-func (m *RequestMetrics) RequestLatencies(endpoint string) *prometheus.Timer {
-	return prometheus.NewTimer(m.requestLatencies.WithLabelValues(endpoint))
+// RequestLatencies fetches the Histogram Observer for the given endpoint
+// If it doesn't exist, a new one is created.
+// This creates 12 (with DefBuckets) empty tables in the prometheus database.
+func (m *RequestMetrics) RequestLatencies(endpoint string) prometheus.Observer {
+	return m.requestLatencies.WithLabelValues(endpoint)
 }