diff --git a/tests/unit/test_gpu_encoder.py b/tests/unit/test_gpu_encoder.py
index 1c72190e1..f3895a009 100644
--- a/tests/unit/test_gpu_encoder.py
+++ b/tests/unit/test_gpu_encoder.py
@@ -3,7 +3,6 @@
 from __future__ import annotations
 
 import json
-import os
 import tempfile
 import urllib.error
 import urllib.request
@@ -34,6 +33,7 @@ def client():
     return GpuEncoderClient(
         endpoint="http://gpu.example.com:8900",
         relay_base_url="http://api.example.com",
+        relay_internal_base_url="http://api-internal:8000",
         sync_timeout=60,
         health_timeout=2,
         relay_secret="test-secret",
@@ -41,15 +41,20 @@ def client():
 
 
 def _fake_response(status: int = 200, body: dict | bytes | None = None, headers=None):
+    """Fake HTTPResponse that supports chunked read(size) used by _download_to_file."""
     if isinstance(body, dict):
         data = json.dumps(body).encode("utf-8")
     elif body is None:
         data = b""
     else:
         data = body
+
+    # Use a real BytesIO so read(size) works for chunked downloads
+    bio = BytesIO(data)
+
     resp = mock.MagicMock(spec=HTTPResponse)
     resp.status = status
-    resp.read.return_value = data
+    resp.read.side_effect = lambda n=-1: bio.read(n)
     resp.__enter__ = mock.MagicMock(return_value=resp)
     resp.__exit__ = mock.MagicMock(return_value=False)
     return resp
@@ -111,36 +116,65 @@ class TestPostSync:
             )
         assert res["status"] == "completed"
         assert res["ffmpeg_rc"] == 0
-        # verify request sent to sync endpoint
         req = m.call_args[0][0]
         assert req.full_url == "http://gpu.example.com:8900/api/render/sync"
 
     def test_ffmpeg_failure_raises(self, client):
         body = {"status": "failed", "ffmpeg_rc": 1, "message": "Invalid data found"}
         with mock.patch("urllib.request.urlopen", return_value=_fake_response(body=body)):
-            with pytest.raises(GpuEncodeError, match="ffmpeg_rc=1"):
+            with pytest.raises(GpuEncodeError, match="rc=1"):
                 client._post_sync(
-                    {"inputs": {}, "ffmpeg_args": [], "output_url": "", "timeout": 10}, mezzanine_path=Path("/tmp/x")
+                    {"inputs": {}, "ffmpeg_args": [], "output_url": "", "timeout": 10},
+                    mezzanine_path=Path("/tmp/x"),
                 )
 
     def test_http_4xx_raises(self, client):
         err = urllib.error.HTTPError(
-            url="http://gpu/render/sync", code=422, msg="Unprocessable", hdrs={}, fp=BytesIO(b"bad request")
+            url="http://gpu/render/sync",
+            code=422,
+            msg="Unprocessable",
+            hdrs={},
+            fp=BytesIO(b"bad request"),
         )
         with mock.patch("urllib.request.urlopen", side_effect=err):
             with pytest.raises(GpuEncodeError, match="HTTP 422"):
                 client._post_sync(
-                    {"inputs": {}, "ffmpeg_args": [], "output_url": "", "timeout": 10}, mezzanine_path=Path("/tmp/x")
+                    {"inputs": {}, "ffmpeg_args": [], "output_url": "", "timeout": 10},
+                    mezzanine_path=Path("/tmp/x"),
                 )
 
 
 class TestRelayUrl:
-    def test_url_contains_token_and_key(self, client):
-        url = client._relay_url("abc123", "secret!")
+    def test_put_url_uses_external_base(self, client):
+        url = client._relay_put_url("abc123", "secret!")
         assert "abc123" in url
         assert "token=secret%21" in url  # urlencoded
         assert url.startswith("http://api.example.com/api/v1/internal/gpu-relay/")
 
+    def test_internal_url_uses_internal_base(self, client):
+        url = client._relay_internal_url("abc123", "s")
+        assert url.startswith("http://api-internal:8000/api/v1/internal/gpu-relay/abc123")
+
+    def test_internal_url_falls_back_to_external_when_not_set(self):
+        c = GpuEncoderClient(
+            endpoint="http://gpu",
+            relay_base_url="http://api.example.com",
+            relay_secret="s",
+        )
+        put = c._relay_put_url("k", "s")
+        internal = c._relay_internal_url("k", "s")
+        assert put.startswith("http://api.example.com/")
+        # When internal not set, internal_url falls back to external base
+        assert internal == put
+
+    def test_encode_uses_different_put_and_get_urls(self, client):
+        """encode_mezzanine_to_output should use external URL for PUT and internal for GET/DELETE."""
+        put_url = client._relay_put_url("k", "test-secret")
+        get_url = client._relay_internal_url("k", "test-secret")
+        assert "api.example.com" in put_url
+        assert "api-internal:8000" in get_url
+        assert put_url != get_url
+
 
 class TestGetRelaySecret:
     def test_explicit_secret_used(self, client):
@@ -175,9 +209,3 @@ class TestDownloadToFile:
             with pytest.raises(GpuEncodeError, match="empty file"):
                 client._download_to_file("http://relay/k", out)
         assert not out.exists()
-
-
-class TestFfmpegOutputToMezzanineIntegration:
-    """_ffmpeg_output_to_mezzanine is on UnifiedRenderService; unit-tested there via mocks."""
-
-    pass