diff --git a/src/interfaces/chat_app/app.py b/src/interfaces/chat_app/app.py
index 555eac96..14c9d924 100644
--- a/src/interfaces/chat_app/app.py
+++ b/src/interfaces/chat_app/app.py
@@ -1647,8 +1647,18 @@ def stream_ab_comparison(
         # PipelineEventFormatter yielding *accumulated* content (not deltas);
         # the last write per arm is therefore the complete response text.
         arm_results = {
-            "a": {"final_text": "", "error": None, "final_emitted": False},
-            "b": {"final_text": "", "error": None, "final_emitted": False},
+            "a": {
+                "final_text": "",
+                "error": None,
+                "final_emitted": False,
+                "duration_ms": None,
+            },
+            "b": {
+                "final_text": "",
+                "error": None,
+                "final_emitted": False,
+                "duration_ms": None,
+            },
         }
         arm_model_used = {
             "a": f"{arm_a_variant.provider or ''}/{arm_a_variant.model or ''}".strip("/"),
@@ -1695,6 +1705,8 @@ def _stream_arm(arm_archi, arm_label):
                         final_text = getattr(output, "answer", "") or formatter.last_text or arm_results[arm_label]["final_text"]
                         arm_results[arm_label]["final_text"] = final_text
                         arm_results[arm_label]["final_emitted"] = True
+                        duration_ms = int((_time.monotonic() - t0) * 1000)
+                        arm_results[arm_label]["duration_ms"] = duration_ms
                         event_queue.put({
                             "type": "final",
                             "arm": arm_label,
@@ -1702,7 +1714,7 @@ def _stream_arm(arm_archi, arm_label):
                             "usage": output_meta.get("usage"),
                             "model": output_meta.get("model"),
                             "model_used": arm_model_used[arm_label],
-                            "duration_ms": int((_time.monotonic() - t0) * 1000),
+                            "duration_ms": duration_ms,
                         })
             except Exception as exc:
                 arm_results[arm_label]["error"] = str(exc)
@@ -1745,6 +1757,8 @@ def _stream_arm(arm_archi, arm_label):
         arm_b_error = arm_results["b"]["error"]
         arm_a_final_text = arm_results["a"]["final_text"]
         arm_b_final_text = arm_results["b"]["final_text"]
+        arm_a_duration_ms = arm_results["a"]["duration_ms"]
+        arm_b_duration_ms = arm_results["b"]["duration_ms"]
 
         if arm_a_error and arm_b_error:
             yield {"type": "error", "message": "Both A/B arms failed",
@@ -1790,6 +1804,10 @@ def _stream_arm(arm_archi, arm_label):
             pipeline_used=self.current_pipeline_used,
         )
 
+        # Persist per-arm latency for analysis by reusing the timing table keyed by message_id.
+        self._persist_ab_arm_timing(arm_a_mid, arm_a_duration_ms)
+        self._persist_ab_arm_timing(arm_b_mid, arm_b_duration_ms)
+
         # Get user prompt message ID if not already stored above
         if not user_prompt_mid:
             user_prompt_mid = self._get_last_user_message_id(context.conversation_id)
@@ -1829,6 +1847,8 @@ def _stream_arm(arm_archi, arm_label):
             "is_champion_first": is_champion_first,
             "arm_a_message_id": arm_a_mid,
             "arm_b_message_id": arm_b_mid,
+            "arm_a_duration_ms": arm_a_duration_ms,
+            "arm_b_duration_ms": arm_b_duration_ms,
             "variant_label_mode": self.ab_pool.variant_label_mode,
         }
 
@@ -1851,6 +1871,33 @@ def _store_assistant_message(self, conversation_id, content, model_used=None, pi
             logger.error("Failed to store assistant message: %s", exc)
             return None
 
+    def _persist_ab_arm_timing(self, message_id: Optional[int], duration_ms: Optional[int]) -> None:
+        """Persist A/B arm latency into the timing table for post-hoc analysis."""
+        if not message_id or duration_ms is None:
+            return
+
+        safe_duration_ms = max(int(duration_ms), 0)
+        end_ts = datetime.now(timezone.utc)
+        start_ts = end_ts - timedelta(milliseconds=safe_duration_ms)
+
+        synthetic_timestamps = {
+            "client_sent_msg_ts": start_ts,
+            "server_received_msg_ts": start_ts,
+            "lock_acquisition_ts": start_ts,
+            "vectorstore_update_ts": start_ts,
+            "query_convo_history_ts": start_ts,
+            "chain_finished_ts": end_ts,
+            "archi_message_ts": end_ts,
+            "insert_convo_ts": end_ts,
+            "finish_call_ts": end_ts,
+            "server_response_msg_ts": end_ts,
+        }
+
+        try:
+            self.insert_timing(message_id, synthetic_timestamps)
+        except Exception as exc:
+            logger.warning("Failed to persist A/B timing for message %s: %s", message_id, exc)
+
     def _get_last_user_message_id(self, conversation_id):
         """Get the most recent user message_id for a conversation."""
         try: