From e441e43d8672b784602359653f72ae43e10f8a5a Mon Sep 17 00:00:00 2001
From: libaojiang <2461443384@qq.com>
Date: Tue, 13 Oct 2026 15:17:00 +0800
Subject: [PATCH] fix(streaming): keep the served model on openrouter routed
 stream chunks

Every SSE chunk OpenRouter sends names the model it actually served, but
chunk_creator only re-read the chunk model for azure and azure_ai, so a
request through a router alias like openrouter/auto surfaced, priced, and
logged every chunk under the alias, which has no price entry, and
stream_chunk_builder inherited the wrong name too. Read the chunk model
for openrouter the same way.

Fixes #46095
---
 .../litellm_core_utils/streaming_handler.py   |  4 +-
 .../test_streaming_handler.py                 | 63 +++++++++++++++++++
 2 files changed, 66 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/streaming_handler.py b/litellm/litellm_core_utils/streaming_handler.py
index a78e6f17cc9..c60ef35cc35 100644
--- a/litellm/litellm_core_utils/streaming_handler.py
+++ b/litellm/litellm_core_utils/streaming_handler.py
@@ -1365,9 +1365,11 @@ class CustomStreamWrapper:
             if self.custom_llm_provider in [
                 LlmProviders.AZURE.value,
                 LlmProviders.AZURE_AI.value,
+                LlmProviders.OPENROUTER.value,
             ]:
                 if isinstance(chunk, BaseModel) and hasattr(chunk, "model"):
-                    # for azure, we need to pass the model from the original chunk
+                    # these providers route to a served model whose name differs
+                    # from the requested one, so pass the chunk's model through
                     self.model = getattr(chunk, "model", self.model)
             response_obj = self.handle_openai_chat_completion_chunk(chunk)
             if response_obj is None:
diff --git a/tests/unit/litellm_core_utils/test_streaming_handler.py b/tests/unit/litellm_core_utils/test_streaming_handler.py
index 01fae0958e6..9dd7036c4f3 100644
--- a/tests/unit/litellm_core_utils/test_streaming_handler.py
+++ b/tests/unit/litellm_core_utils/test_streaming_handler.py
@@ -3228,6 +3228,69 @@ async def test_azure_streaming_role_preserved_with_include_usage(sync_mode: bool
     )
 
 
+def test_openrouter_streaming_chunks_carry_served_model():
+    """
+    Regression test for https://github.com/BerriAI/litellm/issues/46095
+
+    Every raw SSE chunk OpenRouter sends names the model it actually served,
+    e.g. openrouter/auto resolving to deepseek/deepseek-v4.1-flash. The
+    wrapper must keep that per-chunk model like it does for azure, so routed
+    streams are priced and logged under the model that answered instead of
+    the requested router alias, which has no price entry.
+    """
+    served_model = "deepseek/deepseek-v4.1-flash"
+    openrouter_chunks = [
+        ModelResponseStream(
+            id="chatcmpl-or-1",
+            created=1742056047,
+            model=served_model,
+            object="chat.completion.chunk",
+            choices=[
+                StreamingChoices(
+                    finish_reason=None,
+                    index=0,
+                    delta=Delta(content="hi", role="assistant"),
+                )
+            ],
+        ),
+        ModelResponseStream(
+            id="chatcmpl-or-1",
+            created=1742056047,
+            model=served_model,
+            object="chat.completion.chunk",
+            choices=[
+                StreamingChoices(
+                    finish_reason="stop",
+                    index=0,
+                    delta=Delta(content="", role=None),
+                )
+            ],
+        ),
+    ]
+
+    response = CustomStreamWrapper(
+        completion_stream=ModelResponseListIterator(model_responses=openrouter_chunks),
+        model="openrouter/auto",
+        custom_llm_provider="openrouter",
+        logging_obj=Logging(
+            model="openrouter/auto",
+            messages=[{"role": "user", "content": "Say hi"}],
+            stream=True,
+            call_type="completion",
+            start_time=time.time(),
+            litellm_call_id="12345",
+            function_id="1245",
+        ),
+    )
+
+    chunks = [chunk for chunk in response]
+
+    assert len(chunks) > 0
+    assert all(
+        chunk.model == served_model for chunk in chunks
+    ), f"Expected every chunk to carry the served model {served_model}, got {[c.model for c in chunks]}"
+
+
 def test_gemini_legacy_vertex_stop_finish_reason_normalised():
     """
     The legacy vertex_ai SDK streaming path sets finish_reason from a proto enum