fix(azure): price streamed Responses relays and keep a full-URL api_base from doubling the native path

litellm/litellm_core_utils/litellm_logging.py

@@ -119,7 +119,6 @@ from litellm.types.utils import (
     CachingDetails,
     CallTypes,
     CostBreakdown,
-    CostResponseTypes,
     CustomPricingLiteLLMParams,
     DynamicPromptManagementParamLiteral,
     EmbeddingResponse,
@@ -201,7 +200,7 @@ if TYPE_CHECKING:
     from mcp.types import EmbeddedResource, ImageContent, TextContent
 
     from litellm.integrations.otel.logger import OpenTelemetryV2
-    from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig
+    from litellm.llms.base_llm.passthrough.transformation import BasePassthroughConfig, LoggedRelayResponse
 try:
     from litellm_enterprise.enterprise_callbacks.callback_controls import (
         EnterpriseCallbackControls,
@@ -2363,7 +2362,7 @@ class Logging(LiteLLMLoggingBaseClass):
         self,
         raw_bytes: list[bytes],
         provider_config: "BasePassthroughConfig",
-    ) -> Optional["CostResponseTypes"]:
+    ) -> Optional["LoggedRelayResponse"]:
         all_chunks: Final = provider_config._convert_raw_bytes_to_str_lines(raw_bytes)
         complete_streaming_response: Final = provider_config.handle_logging_collected_chunks(
             all_chunks=all_chunks,

litellm/llms/azure/passthrough/transformation.py

@@ -23,7 +23,6 @@ if TYPE_CHECKING:
     from httpx import URL
 
     from litellm.llms.base_llm.passthrough.transformation import LoggedRelayResponse
-    from litellm.types.utils import CostResponseTypes
 
 
 class RelayedChatRequest(BaseModel):
@@ -42,13 +41,29 @@ def _relayed_messages(litellm_logging_obj: Logging) -> Sequence[Mapping[str, obj
     return details.request_data.messages if details.request_data else None
 
 
+RESPONSES_RELAY_SHAPE: Final = RelayShape("/responses", CallTypes.aresponses, ResponsesAPIResponse.model_validate)
+
 OPENAI_RELAY_SHAPES: Final = (
     RelayShape("/embeddings", CallTypes.aembedding, EmbeddingResponse.model_validate),
-    RelayShape("/responses", CallTypes.aresponses, ResponsesAPIResponse.model_validate),
+    RESPONSES_RELAY_SHAPE,
     RelayShape("/images/generations", CallTypes.aimage_generation, ImageResponse.model_validate),
 )
 
 
+def logged_responses_stream(all_chunks: Sequence[str], logging_obj: Logging) -> ResponsesAPIResponse | None:
+    from litellm.llms.openai.responses.transformation import OpenAIResponsesAPIConfig
+
+    terminal_response: Final = OpenAIResponsesAPIConfig.parse_terminal_response_from_stream_chunks(
+        all_chunks=list(all_chunks)
+    )
+    if terminal_response is None:
+        return None
+    logging_obj.call_type = (
+        RESPONSES_RELAY_SHAPE.call_type.value
+    )  # rebind-ok: routes cost calculation to the relayed shape's pricing path
+    return terminal_response
+
+
 class AzurePassthroughConfig(BasePassthroughConfig):
     def is_streaming_request(self, endpoint: str, request_data: dict) -> bool:
         return bool(request_data.get("stream"))
@@ -157,11 +172,13 @@ class AzurePassthroughConfig(BasePassthroughConfig):
         model: str,
         custom_llm_provider: str,
         endpoint: str,
-    ) -> Optional["CostResponseTypes"]:
+    ) -> Optional["LoggedRelayResponse"]:
         from litellm.proxy.pass_through_endpoints.llm_provider_handlers.openai_passthrough_logging_handler import (
             OpenAIPassthroughLoggingHandler,
         )
 
+        if f"/{endpoint.strip('/')}".endswith(RESPONSES_RELAY_SHAPE.path_suffix):
+            return logged_responses_stream(all_chunks, litellm_logging_obj)
         if "chat/completions" not in endpoint:
             return None
 

litellm/llms/azure_ai/passthrough/transformation.py

@@ -30,7 +30,6 @@ if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging
     from litellm.llms.base_llm.ocr.transformation import BaseOCRConfig, OCRResponse
     from litellm.llms.base_llm.passthrough.transformation import LoggedRelayResponse
-    from litellm.types.utils import CostResponseTypes
 
 
 EMPTY_QUERY: Final[Mapping[str, object]] = MappingProxyType({})
@@ -111,8 +110,8 @@ class AzureAIPassthroughConfig(AzureFoundryModelInfo, BasePassthroughConfig):
         if base_target_url is None:
             raise ValueError("Azure AI api base not found: set `api_base` on the deployment or AZURE_AI_API_BASE")
 
-        root: Final = foundry_root(base_target_url)
         native_endpoint: Final = strip_leading_model_segment(endpoint, (model, model_group_from(litellm_params)))
+        root: Final = foundry_root(base_target_url).removesuffix(f"/{native_endpoint.strip('/')}")
         query_params: Final = relay_query_params(
             request_query_params, api_version_from(litellm_params), base_target_url
         )
@@ -200,7 +199,7 @@ class AzureAIPassthroughConfig(AzureFoundryModelInfo, BasePassthroughConfig):
         model: str,
         custom_llm_provider: str,
         endpoint: str,
-    ) -> CostResponseTypes | None:
+    ) -> LoggedRelayResponse | None:
         from litellm.llms.azure.passthrough.transformation import AzurePassthroughConfig
 
         return AzurePassthroughConfig().handle_logging_collected_chunks(

litellm/llms/base_llm/passthrough/transformation.py

@@ -179,7 +179,7 @@ class BasePassthroughConfig(BaseLLMModelInfo):
         model: str,
         custom_llm_provider: str,
         endpoint: str,
-    ) -> CostResponseTypes | None:
+    ) -> LoggedRelayResponse | None:
         return None
 
     def _convert_raw_bytes_to_str_lines(self, raw_bytes: list[bytes]) -> list[str]:

tests/test_litellm/llms/azure/passthrough/test_azure_passthrough_transformation.py

@@ -283,6 +283,53 @@ def test_azure_passthrough_streaming_chunks_for_unknown_endpoint_return_none():
     assert response is None
 
 
+def _azure_responses_stream_chunks(terminal_event: str | None = "response.completed") -> list[str]:
+    in_progress = {**RESPONSES_BODY, "status": "in_progress", "output": [], "usage": None}
+    events = [
+        ("response.created", {"type": "response.created", "sequence_number": 0, "response": in_progress}),
+        (
+            "response.output_text.delta",
+            {"type": "response.output_text.delta", "sequence_number": 1, "item_id": "msg_1", "delta": "hi"},
+        ),
+    ] + ([(terminal_event, {"type": terminal_event, "sequence_number": 2, "response": RESPONSES_BODY})] if terminal_event else [])
+    return [line for name, payload in events for line in (f"event: {name}", _sse_line(payload))]
+
+
+def test_azure_passthrough_streaming_responses_chunks_are_costed_per_token():
+    logging_obj = _relay_logging_obj("gpt-4.1-mini")
+
+    response = AzurePassthroughConfig().handle_logging_collected_chunks(
+        all_chunks=_azure_responses_stream_chunks(),
+        litellm_logging_obj=logging_obj,
+        model="gpt-4.1-mini",
+        custom_llm_provider="azure",
+        endpoint="openai/responses",
+    )
+    info = litellm.get_model_info("azure/gpt-4.1-mini")
+
+    assert isinstance(response, ResponsesAPIResponse)
+    assert response.usage.input_tokens == 1000
+    assert logging_obj.call_type == "aresponses"
+    assert logging_obj._response_cost_calculator(result=response) == pytest.approx(
+        1000 * info["input_cost_per_token"] + 100 * info["output_cost_per_token"]
+    )
+
+
+def test_azure_passthrough_streaming_responses_without_a_terminal_event_are_not_costed():
+    logging_obj = _relay_logging_obj("gpt-4.1-mini")
+
+    response = AzurePassthroughConfig().handle_logging_collected_chunks(
+        all_chunks=_azure_responses_stream_chunks(terminal_event=None),
+        litellm_logging_obj=logging_obj,
+        model="gpt-4.1-mini",
+        custom_llm_provider="azure",
+        endpoint="openai/responses",
+    )
+
+    assert response is None
+    assert logging_obj.call_type == "allm_passthrough_route"
+
+
 def _complete_url(request_query_params: dict, litellm_params: dict) -> httpx.URL:
     url, _ = AzurePassthroughConfig().get_complete_url(
         api_base="https://my-resource.openai.azure.com",

tests/test_litellm/llms/azure_ai/passthrough/test_azure_ai_passthrough_transformation.py

@@ -86,6 +86,22 @@ def test_api_base_that_already_ends_in_models_is_cut_back_to_the_foundry_root():
     assert base == FOUNDRY_BASE
 
 
+def test_full_url_api_base_that_already_ends_with_the_native_path_is_not_doubled():
+    model_router_url = "https://my-resource.cognitiveservices.azure.com/openai/deployments/model-router/chat/completions"
+
+    url, base = AzureAIPassthroughConfig().get_complete_url(
+        api_base=f"{model_router_url}?api-version=2025-01-01-preview",
+        api_key="key",
+        model="model_router/model-router",
+        endpoint="model-router/chat/completions",
+        request_query_params=None,
+        litellm_params={"litellm_metadata": {"model_group": "model-router"}},
+    )
+
+    assert str(url) == f"{model_router_url}?api-version=2025-01-01-preview"
+    assert base == "https://my-resource.cognitiveservices.azure.com/openai/deployments/model-router"
+
+
 def test_parse_relay_under_a_models_api_base_targets_the_foundry_root():
     url, _ = AzureAIPassthroughConfig().get_complete_url(
         api_base=f"{FOUNDRY_BASE}/models",
@@ -490,3 +506,44 @@ def test_streaming_chat_completion_chunks_are_costed_like_azure():
     assert isinstance(response, ModelResponse)
     assert response.choices[0].message.content == "hi"
     assert response.usage.total_tokens == 4
+
+
+def test_streaming_responses_chunks_through_a_router_relay_are_costed_like_azure():
+    completed = {
+        "type": "response.completed",
+        "sequence_number": 2,
+        "response": {
+            "id": "resp_1",
+            "object": "response",
+            "created_at": 1,
+            "status": "completed",
+            "model": "gpt-5.4-mini",
+            "output": [
+                {
+                    "type": "message",
+                    "id": "msg_1",
+                    "role": "assistant",
+                    "status": "completed",
+                    "content": [{"type": "output_text", "text": "hi", "annotations": []}],
+                }
+            ],
+            "usage": {"input_tokens": 1000, "output_tokens": 100, "total_tokens": 1100},
+        },
+    }
+    logging_obj = _relay_logging_obj("gpt-5.4-mini", FOUNDRY_BASE)
+
+    response = AzureAIPassthroughConfig().handle_logging_collected_chunks(
+        all_chunks=["event: response.completed", "data: " + json.dumps(completed)],
+        litellm_logging_obj=logging_obj,
+        model="gpt-5.4-mini",
+        custom_llm_provider="azure_ai",
+        endpoint="gpt/openai/responses",
+    )
+    info = litellm.get_model_info("azure_ai/gpt-5.4-mini")
+
+    assert response is not None
+    assert response.usage.output_tokens == 100
+    assert logging_obj.call_type == "aresponses"
+    assert logging_obj._response_cost_calculator(result=response) == pytest.approx(
+        1000 * info["input_cost_per_token"] + 100 * info["output_cost_per_token"]
+    )