diff --git a/5_django_backend/apps/gateway/views.py b/5_django_backend/apps/gateway/views.py
index 673aa2e..dbd2b58 100644
--- a/5_django_backend/apps/gateway/views.py
+++ b/5_django_backend/apps/gateway/views.py
@@ -9,6 +9,7 @@
 
 from __future__ import annotations
 
+import asyncio
 import json
 import logging
 from typing import AsyncIterator
@@ -26,6 +27,7 @@ log = logging.getLogger(__name__)
 TRANSPORT: httpx.AsyncBaseTransport | None = None
 # 마지막에 통과한 인증 형식 — 프로세스당 하나. 다음 요청은 이것부터 시도.
 _last_ok: dict[str, Variant | None] = {"variant": None}
+RETRY_PAUSE_S = 1.0  # 마지막 재시도 전 쉬는 시간
 
 
 def _cfg() -> FabrixConfig:
@@ -90,6 +92,9 @@ async def chat_completions(request: HttpRequest) -> HttpResponse:
     kinds = sorted({p.get("type", "?") for m in payload.get("messages") or [] for p in (m.get("content") if isinstance(m.get("content"), list) else [])})
     log.info("ito → model=%s parts=%s stream=%s", headers.get("x-llm-model-id"), kinds or ["text"], bool(body.get("stream")))
     variants = auth_variants(headers, _last_ok["variant"])
+    # FabriX 가 맞는 조합에도 가끔 401 을 뱉음(2026-09-21 고객사 실측: curl 10번 중 1~2번).
+    # 4개 조합 다 돌고 나서 제일 유력한 조합(맨 앞)을 잠깐 쉬고 한 번 더 — 그래도 401 이면 진짜 인증 문제
+    variants = variants + variants[:1]
 
     if body.get("stream"):
 
@@ -103,6 +108,8 @@ async def chat_completions(request: HttpRequest) -> HttpResponse:
                     log.info("ito ← %s (auth %s/%s: %s)", resp.status_code, i + 1, len(variants), vkey)
                     if resp.status_code == 401 and i + 1 < len(variants):
                         await resp.aclose()  # 인증 형식이 안 맞은 것 — 다른 조합으로 한 번 더
+                        if i + 2 == len(variants):
+                            await asyncio.sleep(RETRY_PAUSE_S)
                         continue
                     break
                 assert resp is not None
@@ -126,6 +133,8 @@ async def chat_completions(request: HttpRequest) -> HttpResponse:
             for i, (vkey, vh) in enumerate(variants):
                 resp = await client.post(cfg.url, headers=vh, json=body)
                 if resp.status_code == 401 and i + 1 < len(variants):
+                    if i + 2 == len(variants):
+                        await asyncio.sleep(RETRY_PAUSE_S)
                     continue
                 if resp.status_code < 400:
                     _last_ok["variant"] = vkey
diff --git a/5_django_backend/tests/test_gateway.py b/5_django_backend/tests/test_gateway.py
index a1904c3..a39bff1 100644
--- a/5_django_backend/tests/test_gateway.py
+++ b/5_django_backend/tests/test_gateway.py
@@ -126,6 +126,26 @@ async def test_401_retries_other_auth_shape_and_remembers(upstream):
     assert len(calls) == n + 1 and calls[-1].headers["x-openapi-token"] == "Bearer tok"
 
 
+@override_settings(FABRIX_ENV=FULL_ENV)
+async def test_flaky_401_on_right_auth_is_retried_once(upstream, monkeypatch):
+    """FabriX 가 맞는 조합에도 가끔 401 — 4개 다 돈 뒤 맨 앞 조합 한 번 더."""
+    from apps.gateway import views
+
+    monkeypatch.setattr(views, "RETRY_PAUSE_S", 0)
+    hits = {"n": 0}
+
+    def handler(req: httpx.Request) -> httpx.Response:
+        hits["n"] += 1
+        if hits["n"] <= 4:
+            return httpx.Response(401, content=b"flaky")
+        return httpx.Response(200, json={"choices": []})
+
+    calls = upstream(handler)
+    resp = await _post({"messages": []})
+    assert resp.status_code == 200 and len(calls) == 5
+    assert calls[0].headers.get("x-openapi-token") == calls[4].headers.get("x-openapi-token")
+
+
 @override_settings(FABRIX_ENV=FULL_ENV)
 async def test_stream_upstream_error_becomes_sse_error(upstream):
     upstream(lambda r: _sse(b"boom", 500))