fix(router): resolve realtime session model to routed deployment

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/router.py

@@ -29,6 +29,7 @@ import anyio
 import httpx
 import openai
 from openai import AsyncOpenAI
+from pydantic import TypeAdapter, ValidationError
 from typing_extensions import overload
 
 import litellm
@@ -342,6 +343,26 @@ def _stream_chunks_have_generated_content(chunks: Sequence[ModelResponseStream])
     return False
 
 
+_NO_SESSION_KWARGS: Final[Mapping[str, Mapping[str, object]]] = MappingProxyType({})
+_SESSION_ADAPTER: Final = TypeAdapter(Mapping[str, object])
+
+
+def _with_router_resolved_session_model(session: object, model_name: str) -> Mapping[str, Mapping[str, object]]:
+    """
+    Realtime client-secret requests carry the model inside ``session`` as well, and the caller's copy of it still
+    holds the pre-routing model group name, so it has to follow the deployment the router just picked.
+
+    Returns kwargs to merge into the downstream call, empty when there is no session model to resolve.
+    """
+    try:
+        typed_session: Final = _SESSION_ADAPTER.validate_python(session)
+    except ValidationError:
+        return _NO_SESSION_KWARGS
+    if "model" not in typed_session:
+        return _NO_SESSION_KWARGS
+    return MappingProxyType({"session": {**typed_session, "model": model_name}})
+
+
 class RoutingArgs(enum.Enum):
     ttl = 60  # 1min (RPM/TPM expire key)
 
@@ -4685,6 +4706,7 @@ class Router:
                 "caching": self.cache_responses,
                 **kwargs,
                 "model": model_name,
+                **_with_router_resolved_session_model(kwargs.get("session"), model_name),
             }
             # Only set custom_llm_provider if it's not None
             if custom_llm_provider is not None:

tests/test_litellm/test_router.py

@@ -1298,6 +1298,82 @@ async def test_ageneric_api_call_deployment_model_overrides_alias():
     ), f"Expected deployment model 'vertex_ai/gemini-2.5-flash', got '{captured['model']}'"
 
 
+@pytest.mark.asyncio
+async def test_ageneric_api_call_resolves_realtime_session_model():
+    """
+    Regression for #36742: realtime client secret requests carry the model inside `session` too, and the proxy
+    fills it with the pre-routing model group name. The underlying litellm function reads session.model first,
+    so it must see the resolved deployment, while a caller's nested transcription model stays untouched.
+    """
+    captured: dict = {}
+
+    async def capture_kwargs(**kwargs):
+        captured.update(kwargs)
+        return {"result": "ok"}
+
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "my-realtime-group",
+                "litellm_params": {
+                    "model": "openai/gpt-realtime-2.1-mini",
+                    "api_key": "fake-key",
+                },
+                "model_info": {"mode": "realtime"},
+            }
+        ]
+    )
+
+    await router._ageneric_api_call_with_fallbacks(
+        model="my-realtime-group",
+        original_function=capture_kwargs,
+        session={
+            "type": "realtime",
+            "model": "my-realtime-group",
+            "audio": {"input": {"transcription": {"model": "gpt-4o-transcribe"}}},
+        },
+    )
+
+    assert captured["model"] == "openai/gpt-realtime-2.1-mini"
+    assert captured["session"]["model"] == "openai/gpt-realtime-2.1-mini"
+    assert captured["session"]["audio"]["input"]["transcription"]["model"] == "gpt-4o-transcribe"
+
+
+@pytest.mark.asyncio
+async def test_ageneric_api_call_does_not_add_session_model():
+    """
+    A session that never carried a model must not gain one from routing: the underlying function then falls back
+    to the resolved `model` kwarg itself, and the outgoing session body keeps the caller's shape.
+    """
+    captured: dict = {}
+
+    async def capture_kwargs(**kwargs):
+        captured.update(kwargs)
+        return {"result": "ok"}
+
+    router = litellm.Router(
+        model_list=[
+            {
+                "model_name": "my-realtime-group",
+                "litellm_params": {
+                    "model": "openai/gpt-realtime-2.1-mini",
+                    "api_key": "fake-key",
+                },
+                "model_info": {"mode": "realtime"},
+            }
+        ]
+    )
+
+    await router._ageneric_api_call_with_fallbacks(
+        model="my-realtime-group",
+        original_function=capture_kwargs,
+        session={"type": "realtime"},
+    )
+
+    assert captured["model"] == "openai/gpt-realtime-2.1-mini"
+    assert captured["session"] == {"type": "realtime"}
+
+
 def test_router_get_model_access_groups_team_only_models():
     """
     Test that Router.get_model_access_groups returns the correct response for team-only models