From e8550c1454175a47df4a3d8fb18aeac174bfd609 Mon Sep 17 00:00:00 2001
From: Kerry Lu <klu@berri.ai>
Date: Fri, 16 Oct 2026 18:06:01 -0700
Subject: [PATCH] fix(openai): send chat completions through client.post so the
 SDK's per-message param transform stops holding the event loop (#46967)

* fix(openai): send chat completions through client.post so the SDK's per-message param transform stops holding the event loop

create() walks every message, content part and tool of the request through
openai._utils._transform before posting, 91 ms at 300k tokens and ~300 ms at
1M tokens on the event loop, and returns the same dict for the plain dicts
litellm builds. client.post is what create() calls after that walk, same
request, headers, retries, raw response and stream objects, so the chat
helpers now call it directly through the helper embeddings already used

* fix(openai): type the chat client.post result as LegacyAPIResponse so the type gate sees .headers and .parse

* fix(openai): assert the chat client.post result is a LegacyAPIResponse instead of casting it

* test(openai): stub client.post with a LegacyAPIResponse spec in the empty-response and no-leak tests

* test(core-utils): stub client.post with a LegacyAPIResponse spec in the agentic loop leak test

* test: every client.post stub on the OpenAI chat path returns a LegacyAPIResponse-spec'd mock

---------

Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/llms/openai/openai.py                 | 40 ++++++--
 .../test_chat_completion_agentic_loop.py      | 10 +-
 .../test_exception_mapping_utils.py           |  5 +-
 ...test_azure_chat_o_series_transformation.py | 25 ++---
 tests/unit/llms/azure/test_azure.py           | 12 +--
 .../test_litellm_proxy_chat_transformation.py | 27 +++---
 tests/unit/llms/nvidia_nim/test_nvidia_nim.py |  8 +-
 .../openai/test_o_series_transformation.py    |  8 +-
 tests/unit/llms/openai/test_openai.py         | 93 ++++++++++++++++---
 .../llms/openai/test_openai_empty_response.py | 22 ++---
 .../test_use_chat_completions_api_no_leak.py  |  7 +-
 .../perplexity/chat/test_transformation.py    |  8 +-
 tests/unit/llms/volcengine/test_volcengine.py | 16 +---
 tests/unit/test_main.py                       | 93 ++++++++++---------
 14 files changed, 234 insertions(+), 140 deletions(-)

diff --git a/litellm/llms/openai/openai.py b/litellm/llms/openai/openai.py
index 0c1eaaaae26..73f40f5e0e4 100644
--- a/litellm/llms/openai/openai.py
+++ b/litellm/llms/openai/openai.py
@@ -11,13 +11,14 @@ if TYPE_CHECKING:
     from litellm.litellm_core_utils.tokenizer import Encoding as Tokenizer
 
 import openai
-from openai import AsyncOpenAI, OpenAI
+from openai import AsyncOpenAI, AsyncStream, OpenAI, Stream
 from openai._base_client import make_request_options
 from openai._constants import RAW_RESPONSE_HEADER
 from openai._legacy_response import LegacyAPIResponse
 from openai._types import RequestOptions
 from openai.types import CreateEmbeddingResponse
 from openai.types.beta.assistant_deleted import AssistantDeleted
+from openai.types.chat import ChatCompletion, ChatCompletionChunk
 from openai.types.file_deleted import FileDeleted
 from pydantic import BaseModel, TypeAdapter
 from typing_extensions import overload
@@ -354,7 +355,14 @@ _NO_EXTRA_HEADERS: Final[Mapping[str, str]] = types.MappingProxyType({})
 _SDK_OPTION_KEYS: Final = frozenset(("extra_headers", "extra_query", "extra_body"))
 
 
-def _embedding_request_without_sdk_defaults(
+def _raw_chat_response(
+    result: ChatCompletion | Stream[ChatCompletionChunk] | AsyncStream[ChatCompletionChunk],
+) -> LegacyAPIResponse[ChatCompletion]:
+    assert isinstance(result, LegacyAPIResponse)
+    return result
+
+
+def _request_without_sdk_transform(
     data: Mapping[str, object], timeout: float | httpx.Timeout
 ) -> tuple[Mapping[str, object], RequestOptions]:
     body: Final = {k: v for k, v in data.items() if k not in _SDK_OPTION_KEYS}
@@ -517,7 +525,17 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
         """
         start_time: Final = time.time()
         try:
-            raw_response = await openai_aclient.chat.completions.with_raw_response.create(**data, timeout=timeout)
+            body, options = _request_without_sdk_transform(data, timeout)
+            raw_response = _raw_chat_response(
+                await openai_aclient.post(
+                    "/chat/completions",
+                    body=body,
+                    cast_to=ChatCompletion,
+                    options=options,
+                    stream=bool(data.get("stream")),
+                    stream_cls=AsyncStream[ChatCompletionChunk],
+                )
+            )
             end_time = time.time()
 
             if hasattr(raw_response, "headers"):
@@ -558,7 +576,17 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
         """
         raw_response = None
         try:
-            raw_response = openai_client.chat.completions.with_raw_response.create(**data, timeout=timeout)
+            body, options = _request_without_sdk_transform(data, timeout)
+            raw_response = _raw_chat_response(
+                openai_client.post(
+                    "/chat/completions",
+                    body=body,
+                    cast_to=ChatCompletion,
+                    options=options,
+                    stream=bool(data.get("stream")),
+                    stream_cls=Stream[ChatCompletionChunk],
+                )
+            )
 
             if hasattr(raw_response, "headers"):
                 headers = dict(raw_response.headers)
@@ -1257,7 +1285,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
         logging_obj: LiteLLMLoggingObj,
     ) -> LegacyAPIResponse[CreateEmbeddingResponse]:
         if "encoding_format" not in data:
-            body, options = _embedding_request_without_sdk_defaults(data, timeout)
+            body, options = _request_without_sdk_transform(data, timeout)
             bypass_response: Final = await openai_aclient.post(
                 "/embeddings", body=body, options=options, cast_to=CreateEmbeddingResponse
             )
@@ -1274,7 +1302,7 @@ class OpenAIChatCompletion(BaseLLM, BaseOpenAILLM):
         logging_obj: LiteLLMLoggingObj,
     ) -> LegacyAPIResponse[CreateEmbeddingResponse]:
         if "encoding_format" not in data:
-            body, options = _embedding_request_without_sdk_defaults(data, timeout)
+            body, options = _request_without_sdk_transform(data, timeout)
             bypass_response: Final = openai_client.post(
                 "/embeddings", body=body, options=options, cast_to=CreateEmbeddingResponse
             )
diff --git a/tests/unit/litellm_core_utils/test_chat_completion_agentic_loop.py b/tests/unit/litellm_core_utils/test_chat_completion_agentic_loop.py
index e891706df13..3cbc3591009 100644
--- a/tests/unit/litellm_core_utils/test_chat_completion_agentic_loop.py
+++ b/tests/unit/litellm_core_utils/test_chat_completion_agentic_loop.py
@@ -25,13 +25,13 @@ from typing import Any, Dict, Final, List, Optional, Tuple
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
-
+from openai._legacy_response import LegacyAPIResponse
 
 import litellm
-from litellm.integrations.custom_logger import CustomLogger
 from litellm.integrations.code_interpreter_interception.handler import (
     CodeInterpreterInterceptionLogger,
 )
+from litellm.integrations.custom_logger import CustomLogger
 from litellm.litellm_core_utils.chat_completion_agentic_loop import (
     maybe_run_chat_completion_agentic_loop,
 )
@@ -45,9 +45,9 @@ from litellm.types.integrations.custom_logger import (
     AgenticLoopRequestPatch,
 )
 from litellm.types.utils import (
+    ChatCompletionMessageToolCall,
     Choices,
     Function,
-    ChatCompletionMessageToolCall,
     Message,
     ModelResponse,
 )
@@ -142,7 +142,7 @@ def _raw_response_for(model_response: ModelResponse) -> MagicMock:
     (an object exposing `.headers` and `.parse()` -> something with model_dump)."""
     parsed = MagicMock()
     parsed.model_dump.return_value = model_response.model_dump()
-    raw = MagicMock()
+    raw = MagicMock(spec=LegacyAPIResponse)
     raw.headers = {}
     raw.parse.return_value = parsed
     return raw
@@ -171,7 +171,7 @@ async def test_internal_control_fields_never_leak_into_provider_body(restore_cal
         ]
     )
     mock_client = MagicMock()
-    mock_client.chat.completions.with_raw_response.create = create
+    mock_client.post = create
 
     response = await litellm.acompletion(
         model="openai/gpt-4o-mini",
diff --git a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py
index 02ebae6dcf6..b24348cfaa8 100644
--- a/tests/unit/litellm_core_utils/test_exception_mapping_utils.py
+++ b/tests/unit/litellm_core_utils/test_exception_mapping_utils.py
@@ -2462,7 +2462,6 @@ def test_openai_gateway_timeout_error():
     Test that the OpenAI gateway timeout error is raised
     """
     openai_client = OpenAI()
-    mapped_target = openai_client.chat.completions.with_raw_response  # type: ignore
 
     def _return_exception(*args, **kwargs):
 
@@ -2498,8 +2497,8 @@ def test_openai_gateway_timeout_error():
 
     with pytest.raises(litellm.Timeout) as exc_info:
         with patch.object(
-            mapped_target,
-            "create",
+            openai_client,
+            "post",
             side_effect=_return_exception,
         ):
             litellm.completion(
diff --git a/tests/unit/llms/azure/chat/test_azure_chat_o_series_transformation.py b/tests/unit/llms/azure/chat/test_azure_chat_o_series_transformation.py
index 3630ef1e354..1cdc288ff87 100644
--- a/tests/unit/llms/azure/chat/test_azure_chat_o_series_transformation.py
+++ b/tests/unit/llms/azure/chat/test_azure_chat_o_series_transformation.py
@@ -1,9 +1,10 @@
 import json
 import traceback
-from typing import Callable, Optional
+from typing import Callable, Final, Optional
 from unittest.mock import MagicMock, patch
 
 import pytest
+from openai._legacy_response import LegacyAPIResponse
 
 import litellm
 from litellm import ModelResponse
@@ -130,7 +131,7 @@ def test_azure_o3_streaming():
         api_version="2024-02-15-preview",
     )
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_create:
+    with patch.object(client, "post") as mock_create:
         try:
             completion(
                 model="azure/o3-mini",
@@ -141,7 +142,7 @@ def test_azure_o3_streaming():
         except Exception as e:
             print(e)
         assert mock_create.call_count == 1
-        assert "stream" in mock_create.call_args.kwargs
+        assert "stream" in mock_create.call_args.kwargs["body"]
 
 
 def test_azure_o_series_routing():
@@ -157,7 +158,7 @@ def test_azure_o_series_routing():
         api_version="2024-02-15-preview",
     )
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_create:
+    with patch.object(client, "post") as mock_create:
         try:
             completion(
                 model="azure/o_series/my-random-deployment-name",
@@ -168,17 +169,19 @@ def test_azure_o_series_routing():
         except Exception as e:
             print(e)
         assert mock_create.call_count == 1
-        assert "stream" not in mock_create.call_args.kwargs
+        assert "stream" not in mock_create.call_args.kwargs["body"]
 
 
 @patch("litellm.main.azure_o1_chat_completions._get_openai_client")
 def test_openai_o_series_max_retries_0(mock_get_openai_client):
     import litellm
 
-    mock_get_openai_client.return_value.chat.completions.with_raw_response.create.return_value.headers = {}
-    mock_get_openai_client.return_value.chat.completions.with_raw_response.create.return_value.parse.return_value = (
-        ModelResponse(choices=[{"message": {"role": "assistant", "content": "Hello"}}])
-    )
+    raw_response: Final = MagicMock(spec=LegacyAPIResponse)
+    raw_response.headers = {}
+    raw_response.parse.return_value = ModelResponse(choices=[{"message": {"role": "assistant", "content": "Hello"}}])
+    client_stub: Final = MagicMock()
+    client_stub.post.return_value = raw_response
+    mock_get_openai_client.return_value = client_stub
     litellm.set_verbose = True
     response = litellm.completion(
         model="azure/o1-preview",
@@ -233,7 +236,7 @@ async def test_azure_o1_series_response_format_extra_params():
     response_format = {"type": "json_object"}
     tool_choice = "auto"
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             await litellm.acompletion(
@@ -251,7 +254,7 @@ async def test_azure_o1_series_response_format_extra_params():
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         print("request_body: ", json.dumps(request_body, indent=4))
         assert request_body["tools"] == tools
diff --git a/tests/unit/llms/azure/test_azure.py b/tests/unit/llms/azure/test_azure.py
index 7907344ae64..7d5afde68ef 100644
--- a/tests/unit/llms/azure/test_azure.py
+++ b/tests/unit/llms/azure/test_azure.py
@@ -313,7 +313,7 @@ def test_azure_openai_gpt_4o_naming(monkeypatch):
         number: str = Field(description="total number of days in a week")
         days: list[str] = Field(description="name of days in a week")
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_post:
+    with patch.object(client, "post") as mock_post:
         try:
             completion(
                 model="azure/gpt4o",
@@ -326,9 +326,9 @@ def test_azure_openai_gpt_4o_naming(monkeypatch):
 
         mock_post.assert_called_once()
 
-        print(mock_post.call_args.kwargs)
+        print(mock_post.call_args.kwargs["body"])
 
-        assert "tool_calls" not in mock_post.call_args.kwargs
+        assert "tool_calls" not in mock_post.call_args.kwargs["body"]
 
 
 @pytest.mark.parametrize(
@@ -382,7 +382,7 @@ def test_azure_gpt_4o_with_tool_call_and_response_format(api_version):
         }
     ]
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_post:
+    with patch.object(client, "post") as mock_post:
         mock_post.return_value.headers = {}
         mock_post.return_value.parse.return_value = litellm.ModelResponse(
             choices=[{"message": {"role": "assistant", "content": InvestigationOutput().model_dump_json()}}]
@@ -411,9 +411,9 @@ def test_azure_gpt_4o_with_tool_call_and_response_format(api_version):
         mock_post.assert_called_once()
 
         if api_version == "2024-10-21":
-            assert "response_format" in mock_post.call_args.kwargs
+            assert "response_format" in mock_post.call_args.kwargs["body"]
         else:
-            assert "response_format" not in mock_post.call_args.kwargs
+            assert "response_format" not in mock_post.call_args.kwargs["body"]
         assert response.choices[0].message.content == InvestigationOutput().model_dump_json()
 
 
diff --git a/tests/unit/llms/litellm_proxy/chat/test_litellm_proxy_chat_transformation.py b/tests/unit/llms/litellm_proxy/chat/test_litellm_proxy_chat_transformation.py
index 86fb6fa15d8..9b8788d3769 100644
--- a/tests/unit/llms/litellm_proxy/chat/test_litellm_proxy_chat_transformation.py
+++ b/tests/unit/llms/litellm_proxy/chat/test_litellm_proxy_chat_transformation.py
@@ -1,15 +1,16 @@
 import json
+import re
 from io import BytesIO
 from pathlib import Path
 from typing import Final
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import httpx
+import pytest
 from openai import AsyncOpenAI, OpenAI
+from openai._legacy_response import LegacyAPIResponse
 from openai.types import CreateEmbeddingResponse, Embedding
 from openai.types.create_embedding_response import Usage
-import pytest
-import re
 
 import litellm
 from litellm import completion
@@ -18,7 +19,6 @@ from litellm.llms.litellm_proxy.chat.transformation import LiteLLMProxyChatConfi
 from tests._vcr_conftest_common import rewound_new_episodes_cassette
 from tests.capturing_transport import CapturingTransport
 
-
 _GATEWAY_EMBEDDING_RESPONSE: Final = CreateEmbeddingResponse(
     object="list",
     data=(Embedding(object="embedding", index=0, embedding=(0.1, 0.2, 0.3)),),
@@ -98,7 +98,7 @@ async def test_litellm_gateway_from_sdk():
     openai_client = OpenAI(api_key="fake-key")
 
     with patch.object(
-        openai_client.chat.completions.with_raw_response, "create", new=MagicMock()
+        openai_client, "post", new=MagicMock()
     ) as mock_call:
         try:
             completion(
@@ -114,9 +114,9 @@ async def test_litellm_gateway_from_sdk():
 
         mock_call.assert_called_once()
 
-        print("Call KWARGS - {}".format(mock_call.call_args.kwargs))
+        print("Call KWARGS - {}".format(mock_call.call_args.kwargs["body"]))
 
-        assert "hello" in mock_call.call_args.kwargs["extra_body"]
+        assert "hello" in mock_call.call_args.kwargs["options"]["extra_json"]
 
 
 @pytest.mark.parametrize("is_async", (False, True))
@@ -264,7 +264,7 @@ async def test_litellm_gateway_from_sdk_rerank(is_async):
         patch_target = client.post
 
     with patch.object(client, "post", new=mock_method):
-        mock_response = MagicMock()
+        mock_response = MagicMock(spec=LegacyAPIResponse)
 
         # Create a mock response similar to OpenAI's rerank response
         mock_response.text = json.dumps(
@@ -408,7 +408,7 @@ async def test_litellm_gateway_from_sdk_structured_output():
     openai_client = OpenAI(api_key="fake-key")
 
     with patch.object(
-        openai_client.chat.completions, "create", new=MagicMock()
+        openai_client, "post", new=MagicMock()
     ) as mock_call:
         try:
             litellm.completion(
@@ -427,8 +427,8 @@ async def test_litellm_gateway_from_sdk_structured_output():
 
         mock_call.assert_called_once()
 
-        print("Call KWARGS - {}".format(mock_call.call_args.kwargs))
-        json_schema = mock_call.call_args.kwargs["response_format"]
+        print("Call KWARGS - {}".format(mock_call.call_args.kwargs["body"]))
+        json_schema = mock_call.call_args.kwargs["body"]["response_format"]
         assert "json_schema" in json_schema
 
 
@@ -486,7 +486,7 @@ def test_litellm_gateway_from_sdk_with_response_cost_in_additional_headers():
     openai_client = OpenAI(api_key="fake-key")
 
     # Create mock response object
-    mock_response = MagicMock()
+    mock_response = MagicMock(spec=LegacyAPIResponse)
     mock_response.headers = {"x-litellm-response-cost": "120"}
     mock_response.parse.return_value = litellm.ModelResponse(
         **{
@@ -528,8 +528,7 @@ def test_litellm_gateway_from_sdk_with_response_cost_in_additional_headers():
     )
 
     with patch.object(
-        openai_client.chat.completions.with_raw_response,
-        "create",
+        openai_client, "post",
         return_value=mock_response,
     ) as mock_call:
         response = litellm.completion(
@@ -564,7 +563,7 @@ async def test_litellm_gateway_image_generation_direct(is_async):
         "created": 1,
         "data": [{"url": "https://example.com/image.png"}],
     }
-    mock_raw_response = MagicMock()
+    mock_raw_response = MagicMock(spec=LegacyAPIResponse)
     mock_raw_response.parse.return_value = mock_openai_response
     mock_raw_response.headers = {}
 
diff --git a/tests/unit/llms/nvidia_nim/test_nvidia_nim.py b/tests/unit/llms/nvidia_nim/test_nvidia_nim.py
index e5298c7ba39..3633ac97265 100644
--- a/tests/unit/llms/nvidia_nim/test_nvidia_nim.py
+++ b/tests/unit/llms/nvidia_nim/test_nvidia_nim.py
@@ -94,7 +94,7 @@ def test_chat_completion_nvidia_nim_with_tools():
     ]
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             completion(
@@ -116,7 +116,7 @@ def test_chat_completion_nvidia_nim_with_tools():
 
         # Add assertions to check the request
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         print("request_body: ", request_body)
 
@@ -208,7 +208,7 @@ def test_completion_nvidia_nim():
         api_key="fake-api-key",
     )
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
+    with patch.object(client, "post") as mock_client:
         try:
             completion(
                 model=model_name,
@@ -226,7 +226,7 @@ def test_completion_nvidia_nim():
             print(e)
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         assert request_body["messages"] == [
             {
diff --git a/tests/unit/llms/openai/test_o_series_transformation.py b/tests/unit/llms/openai/test_o_series_transformation.py
index 6ee882b7d32..121fab66c15 100644
--- a/tests/unit/llms/openai/test_o_series_transformation.py
+++ b/tests/unit/llms/openai/test_o_series_transformation.py
@@ -70,7 +70,7 @@ async def test_o1_handle_system_role(model):
     client = AsyncOpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             await litellm.acompletion(
@@ -83,7 +83,7 @@ async def test_o1_handle_system_role(model):
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         print("request_body: ", request_body)
 
@@ -141,7 +141,7 @@ async def test_o1_max_completion_tokens(model: str):
     client = AsyncOpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             await litellm.acompletion(
@@ -154,7 +154,7 @@ async def test_o1_max_completion_tokens(model: str):
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         print("request_body: ", request_body)
 
diff --git a/tests/unit/llms/openai/test_openai.py b/tests/unit/llms/openai/test_openai.py
index d7ca9319f3a..11d7c7d6d3c 100644
--- a/tests/unit/llms/openai/test_openai.py
+++ b/tests/unit/llms/openai/test_openai.py
@@ -5,12 +5,13 @@ import os
 from collections.abc import Mapping
 from types import ModuleType
 from typing import Final, Literal, Protocol
-from unittest.mock import Mock, patch
+from unittest.mock import MagicMock, Mock, patch
 
 import httpx
 import pytest
 import respx
 from openai import AsyncOpenAI, AzureOpenAI, OpenAI
+from openai._legacy_response import LegacyAPIResponse
 from openai._types import Response as SDKResponse
 from pydantic import BaseModel, ConfigDict, JsonValue
 
@@ -1669,7 +1670,7 @@ async def test_openai_prediction_param_mock():
     client = AsyncOpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             await litellm.acompletion(
@@ -1688,7 +1689,7 @@ async def test_openai_prediction_param_mock():
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         # Verify the request contains the prediction parameter
         assert "prediction" in request_body
@@ -1700,10 +1701,12 @@ async def test_openai_prediction_param_mock():
 def test_openai_max_retries_0(mock_get_openai_client):
     import litellm
 
-    mock_get_openai_client.return_value.chat.completions.with_raw_response.create.return_value.headers = {}
-    mock_get_openai_client.return_value.chat.completions.with_raw_response.create.return_value.parse.return_value = (
-        ModelResponse(choices=[{"message": {"role": "assistant", "content": "Hello"}}])
-    )
+    raw_response: Final = MagicMock(spec=LegacyAPIResponse)
+    raw_response.headers = {}
+    raw_response.parse.return_value = ModelResponse(choices=[{"message": {"role": "assistant", "content": "Hello"}}])
+    client_stub: Final = MagicMock()
+    client_stub.post.return_value = raw_response
+    mock_get_openai_client.return_value = client_stub
     litellm.set_verbose = True
     response = litellm.completion(
         model="gpt-4o-mini",
@@ -1812,7 +1815,7 @@ async def test_openai_safety_identifier_parameter():
     client = AsyncOpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             await litellm.acompletion(
@@ -1825,7 +1828,7 @@ async def test_openai_safety_identifier_parameter():
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         # Verify the request contains the safety_identifier parameter
         assert "safety_identifier" in request_body
@@ -1841,7 +1844,7 @@ def test_openai_safety_identifier_parameter_sync():
     client = OpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             litellm.completion(
@@ -1854,7 +1857,7 @@ def test_openai_safety_identifier_parameter_sync():
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         # Verify the request contains the safety_identifier parameter
         assert "safety_identifier" in request_body
@@ -1871,7 +1874,7 @@ async def test_openai_service_tier_parameter():
     client = AsyncOpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             await litellm.acompletion(
@@ -1884,7 +1887,7 @@ async def test_openai_service_tier_parameter():
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         # Verify the request contains the service_tier parameter
         assert "service_tier" in request_body, "service_tier should be in request body"
@@ -1902,7 +1905,7 @@ def test_openai_service_tier_parameter_sync():
     client = OpenAI(api_key="fake-api-key")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create"
+        client, "post"
     ) as mock_client:
         try:
             litellm.completion(
@@ -1915,7 +1918,7 @@ def test_openai_service_tier_parameter_sync():
             print(f"Error: {e}")
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
 
         # Verify the request contains the service_tier parameter
         assert "service_tier" in request_body, "service_tier should be in request body"
@@ -1947,3 +1950,63 @@ def test_responses_gpt54_with_xhigh_reasoning():
         assert request_body["model"] == "openai/gpt-5.4"
 
         assert request_body["reasoning"] == {"effort": "xhigh"}
+
+
+_CHAT_RESPONSE_JSON: Final = {
+    "id": "chatcmpl-wire",
+    "object": "chat.completion",
+    "created": 1,
+    "model": "gpt-5.4",
+    "choices": [{"index": 0, "message": {"role": "assistant", "content": "ok"}, "finish_reason": "stop"}],
+    "usage": {"prompt_tokens": 1, "completion_tokens": 1, "total_tokens": 2},
+}
+_WIRE_REQUEST: Final = {
+    "model": "gpt-5.4",
+    "messages": [
+        {"role": "system", "content": "sys"},
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "what is in the image"},
+                {"type": "image_url", "image_url": {"url": "https://example.com/a.png", "detail": "low"}},
+            ],
+        },
+    ],
+    "tools": [{"type": "function", "function": {"name": "f", "parameters": {"type": "object", "properties": {}}}}],
+    "tool_choice": "auto",
+    "response_format": {"type": "json_schema", "json_schema": {"name": "r", "schema": {"type": "object"}}},
+    "max_completion_tokens": 7,
+    "reasoning_effort": "none",
+    "n": 2,
+    "seed": 3,
+}
+
+
+@pytest.mark.parametrize("is_async", (False, True))
+@respx.mock
+async def test_chat_completion_sends_the_caller_params_extra_body_and_extra_headers_unchanged(
+    is_async: bool, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
+    route: Final = respx.post("https://api.openai.com/v1/chat/completions").mock(
+        return_value=httpx.Response(200, json=_CHAT_RESPONSE_JSON)
+    )
+    kwargs: Final = {
+        **_WIRE_REQUEST,
+        "model": "openai/gpt-5.4",
+        "api_base": "https://api.openai.com/v1",
+        "api_key": "sk-test",
+        "num_retries": 0,
+        "extra_body": {"vendor_flag": True},
+        "extra_headers": {"x-caller": "unit"},
+    }
+
+    if is_async:
+        await litellm.acompletion(**kwargs)
+    else:
+        litellm.completion(**kwargs)
+
+    sent: Final = route.calls.last.request
+    assert json.loads(sent.content) == {**_WIRE_REQUEST, "vendor_flag": True}
+    assert sent.headers["x-caller"] == "unit"
+    assert sent.headers["authorization"] == "Bearer sk-test"
diff --git a/tests/unit/llms/openai/test_openai_empty_response.py b/tests/unit/llms/openai/test_openai_empty_response.py
index 26b28f967db..1505d9ef049 100644
--- a/tests/unit/llms/openai/test_openai_empty_response.py
+++ b/tests/unit/llms/openai/test_openai_empty_response.py
@@ -2,13 +2,13 @@
 Test for issue #17209: Clearer error when LLM endpoint returns empty response
 """
 
-from unittest.mock import MagicMock, patch
+from unittest.mock import MagicMock
 
 import pytest
+from openai._legacy_response import LegacyAPIResponse
 
-
-from litellm.llms.openai.openai import OpenAIChatCompletion
 from litellm.llms.openai.common_utils import OpenAIError
+from litellm.llms.openai.openai import OpenAIChatCompletion
 
 
 class TestEmptyResponseHandling:
@@ -22,12 +22,12 @@ class TestEmptyResponseHandling:
         openai_chat = OpenAIChatCompletion()
 
         # Mock the raw response to return an empty string from parse()
-        mock_raw_response = MagicMock()
+        mock_raw_response = MagicMock(spec=LegacyAPIResponse)
         mock_raw_response.headers = {}
         mock_raw_response.parse.return_value = ""  # Empty string response
 
         mock_client = MagicMock()
-        mock_client.chat.completions.with_raw_response.create.return_value = (
+        mock_client.post.return_value = (
             mock_raw_response
         )
 
@@ -48,12 +48,12 @@ class TestEmptyResponseHandling:
         """Test that None response also produces a clear error"""
         openai_chat = OpenAIChatCompletion()
 
-        mock_raw_response = MagicMock()
+        mock_raw_response = MagicMock(spec=LegacyAPIResponse)
         mock_raw_response.headers = {}
         mock_raw_response.parse.return_value = None
 
         mock_client = MagicMock()
-        mock_client.chat.completions.with_raw_response.create.return_value = (
+        mock_client.post.return_value = (
             mock_raw_response
         )
 
@@ -75,12 +75,12 @@ class TestEmptyResponseHandling:
         mock_response = MagicMock()
         mock_response.model_dump.return_value = {"choices": []}
 
-        mock_raw_response = MagicMock()
+        mock_raw_response = MagicMock(spec=LegacyAPIResponse)
         mock_raw_response.headers = {"x-request-id": "123"}
         mock_raw_response.parse.return_value = mock_response
 
         mock_client = MagicMock()
-        mock_client.chat.completions.with_raw_response.create.return_value = (
+        mock_client.post.return_value = (
             mock_raw_response
         )
 
@@ -104,12 +104,12 @@ class TestEmptyResponseHandling:
         # Create a mock response WITHOUT model_dump (like an AsyncStream/Iterator)
         mock_stream = MagicMock(spec=[])  # spec=[] means no attributes
 
-        mock_raw_response = MagicMock()
+        mock_raw_response = MagicMock(spec=LegacyAPIResponse)
         mock_raw_response.headers = {"x-request-id": "123"}
         mock_raw_response.parse.return_value = mock_stream
 
         mock_client = MagicMock()
-        mock_client.chat.completions.with_raw_response.create.return_value = (
+        mock_client.post.return_value = (
             mock_raw_response
         )
 
diff --git a/tests/unit/llms/openai/test_use_chat_completions_api_no_leak.py b/tests/unit/llms/openai/test_use_chat_completions_api_no_leak.py
index 7013afc7a5f..58f1ff60ba5 100644
--- a/tests/unit/llms/openai/test_use_chat_completions_api_no_leak.py
+++ b/tests/unit/llms/openai/test_use_chat_completions_api_no_leak.py
@@ -9,6 +9,7 @@ body. OpenAI/Anthropic reject unknown body params with HTTP 400.
 
 from unittest.mock import MagicMock
 
+from openai._legacy_response import LegacyAPIResponse
 
 import litellm
 from litellm.types.utils import all_litellm_params
@@ -47,12 +48,12 @@ def test_completion_does_not_leak_flag_into_provider_request_body():
         },
     }
 
-    mock_raw_response = MagicMock()
+    mock_raw_response = MagicMock(spec=LegacyAPIResponse)
     mock_raw_response.headers = {}
     mock_raw_response.parse.return_value = mock_response
 
     mock_client = MagicMock()
-    mock_client.chat.completions.with_raw_response.create.return_value = (
+    mock_client.post.return_value = (
         mock_raw_response
     )
 
@@ -65,7 +66,7 @@ def test_completion_does_not_leak_flag_into_provider_request_body():
     )
 
     create_kwargs = (
-        mock_client.chat.completions.with_raw_response.create.call_args.kwargs
+        mock_client.post.call_args.kwargs
     )
     assert "use_chat_completions_api" not in create_kwargs
     assert "use_chat_completions_api" not in (create_kwargs.get("extra_body") or {})
diff --git a/tests/unit/llms/perplexity/chat/test_transformation.py b/tests/unit/llms/perplexity/chat/test_transformation.py
index 6a70fd1232d..0cb6335f5a8 100644
--- a/tests/unit/llms/perplexity/chat/test_transformation.py
+++ b/tests/unit/llms/perplexity/chat/test_transformation.py
@@ -4,6 +4,7 @@ import os
 from unittest.mock import MagicMock, patch
 
 import pytest
+from openai._legacy_response import LegacyAPIResponse
 
 import litellm
 from litellm import completion
@@ -92,7 +93,7 @@ class TestPerplexityReasoning:
         pydantic_obj = ChatCompletion(**response_object)
 
         def _return_pydantic_obj(*args, **kwargs):
-            new_response = MagicMock()
+            new_response = MagicMock(spec=LegacyAPIResponse)
             new_response.headers = {"content-type": "application/json"}
             new_response.parse.return_value = pydantic_obj
             return new_response
@@ -100,8 +101,7 @@ class TestPerplexityReasoning:
         openai_client = OpenAI(api_key="fake-api-key")
 
         with patch.object(
-            openai_client.chat.completions.with_raw_response,
-            "create",
+            openai_client, "post",
             side_effect=_return_pydantic_obj,
         ) as mock_client:
             response = completion(
@@ -121,7 +121,7 @@ class TestPerplexityReasoning:
 
             # Get the request data from the mock call
             call_args = mock_client.call_args
-            request_data = call_args.kwargs
+            request_data = call_args.kwargs["body"]
 
             # Verify reasoning_effort was included in the request
             assert "reasoning_effort" in request_data
diff --git a/tests/unit/llms/volcengine/test_volcengine.py b/tests/unit/llms/volcengine/test_volcengine.py
index d472d52d2db..dff9a0c62b2 100644
--- a/tests/unit/llms/volcengine/test_volcengine.py
+++ b/tests/unit/llms/volcengine/test_volcengine.py
@@ -2,6 +2,7 @@ import os
 import sys
 from unittest.mock import MagicMock, patch
 
+from openai._legacy_response import LegacyAPIResponse
 from pydantic import BaseModel
 
 from litellm.llms.volcengine.chat.transformation import (
@@ -117,7 +118,7 @@ class TestVolcEngineConfig:
 
         client = OpenAI(api_key="test_api_key")
 
-        mock_raw_response = MagicMock()
+        mock_raw_response = MagicMock(spec=LegacyAPIResponse)
         mock_raw_response.headers = {
             "x-request-id": "123",
             "openai-organization": "org-123",
@@ -126,9 +127,7 @@ class TestVolcEngineConfig:
         }
         mock_raw_response.parse.return_value = ModelResponse()
 
-        with patch.object(
-            client.chat.completions.with_raw_response, "create", mock_raw_response
-        ) as mock_create:
+        with patch.object(client, "post", MagicMock(return_value=mock_raw_response)) as mock_create:
             completion(
                 model="volcengine/doubao-seed-1.6",
                 messages=[
@@ -144,11 +143,6 @@ class TestVolcEngineConfig:
             )
 
             mock_create.assert_called_once()
-            print(mock_create.call_args.kwargs)
+            print(mock_create.call_args.kwargs["body"])
             # Fixed: thinking disabled should appear in extra_body with original structure
-            assert (
-                "extra_body" in mock_create.call_args.kwargs
-                and "thinking" in mock_create.call_args.kwargs.get("extra_body", {})
-                and mock_create.call_args.kwargs.get("extra_body", {})["thinking"]
-                == {"type": "disabled"}
-            )
+            assert mock_create.call_args.kwargs["options"]["extra_json"]["thinking"] == {"type": "disabled"}
diff --git a/tests/unit/test_main.py b/tests/unit/test_main.py
index 14cea8f0c2c..bb0e8db7f1c 100644
--- a/tests/unit/test_main.py
+++ b/tests/unit/test_main.py
@@ -21,15 +21,22 @@ import httpx
 import pytest
 import respx
 from openai import APITimeoutError, AsyncOpenAI
+from openai._legacy_response import LegacyAPIResponse
 from openai.types.chat.chat_completion import ChatCompletion
 
 import litellm
-from litellm import acompletion, completion
-from litellm import acompletion_with_retries, aresponses_with_retries, completion_with_retries, responses_with_retries
+from litellm import (
+    acompletion,
+    acompletion_with_retries,
+    aresponses_with_retries,
+    completion,
+    completion_with_retries,
+    responses_with_retries,
+)
 from litellm import main as litellm_main
-from litellm.constants import CONTROL_OPTIONS_KEY
 from litellm.caching.base_cache import BaseCache
 from litellm.caching.caching import Cache
+from litellm.constants import CONTROL_OPTIONS_KEY
 from litellm.integrations.custom_logger import CustomLogger
 from litellm.integrations.custom_prompt_management import CustomPromptManagement
 from litellm.litellm_core_utils.core_helpers import get_litellm_metadata_from_kwargs
@@ -130,7 +137,7 @@ def test_completion_missing_role(openai_api_response):
 
     client = OpenAI(api_key="test_api_key")
 
-    mock_raw_response = MagicMock()
+    mock_raw_response = MagicMock(spec=LegacyAPIResponse)
     mock_raw_response.headers = {
         "x-request-id": "123",
         "openai-organization": "org-123",
@@ -142,7 +149,7 @@ def test_completion_missing_role(openai_api_response):
     print(f"openai_api_response: {openai_api_response}")
 
     with patch.object(
-        client.chat.completions.with_raw_response, "create", MagicMock(return_value=mock_raw_response)
+        client, "post", MagicMock(return_value=mock_raw_response)
     ) as mock_create:
         litellm.completion(
             model="gpt-4o-mini",
@@ -218,7 +225,7 @@ async def test_url_with_format_param_openai(model, sync_mode):
             }
         ],
     }
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
+    with patch.object(client, "post") as mock_client:
         try:
             if sync_mode:
                 response = completion(**args, client=client)
@@ -232,7 +239,7 @@ async def test_url_with_format_param_openai(model, sync_mode):
 
         print(mock_client.call_args.kwargs)
 
-        json_str = json.dumps(mock_client.call_args.kwargs)
+        json_str = json.dumps(mock_client.call_args.kwargs["body"])
 
         assert "format" not in json_str
 
@@ -3005,9 +3012,10 @@ def test_stream_chunk_builder_thinking_blocks():
     assert response.choices[0].message.thinking_blocks is not None
 
 
-from litellm.llms.openai.openai import OpenAIChatCompletion
 import traceback
 
+from litellm.llms.openai.openai import OpenAIChatCompletion
+
 user_message = "Write a short poem about the sky"
 messages = [{"content": user_message, "role": "user"}]
 
@@ -3777,7 +3785,7 @@ def test_completion_forwards_store_and_prompt_cache_key_to_openai():
 
     client = OpenAI(api_key="fake-api-key")
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
+    with patch.object(client, "post") as mock_client:
         try:
             litellm.completion(
                 model="openai/gpt-4o",
@@ -3790,7 +3798,7 @@ def test_completion_forwards_store_and_prompt_cache_key_to_openai():
             print(e)
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
         assert request_body["store"] is False
         assert request_body["prompt_cache_key"] == "test-cache-key"
 
@@ -3805,7 +3813,7 @@ async def test_acompletion_forwards_store_and_prompt_cache_key_to_openai():
 
     client = AsyncOpenAI(api_key="fake-api-key")
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
+    with patch.object(client, "post") as mock_client:
         try:
             await litellm.acompletion(
                 model="openai/gpt-4o",
@@ -3818,7 +3826,7 @@ async def test_acompletion_forwards_store_and_prompt_cache_key_to_openai():
             print(e)
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
         assert request_body["store"] is False
         assert request_body["prompt_cache_key"] == "test-cache-key"
 
@@ -3832,7 +3840,7 @@ def test_completion_omits_store_and_prompt_cache_key_when_not_passed():
 
     client = OpenAI(api_key="fake-api-key")
 
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
+    with patch.object(client, "post") as mock_client:
         try:
             litellm.completion(
                 model="openai/gpt-4o",
@@ -3843,7 +3851,7 @@ def test_completion_omits_store_and_prompt_cache_key_when_not_passed():
             print(e)
 
         mock_client.assert_called_once()
-        request_body = mock_client.call_args.kwargs
+        request_body = mock_client.call_args.kwargs["body"]
         assert "store" not in request_body
         assert "prompt_cache_key" not in request_body
 
@@ -4038,8 +4046,7 @@ def test_openai_model_without_a_provider_still_routes_to_openai():
     from openai import OpenAI
 
     client = OpenAI(api_key="fake-key")
-    raw_response = client.chat.completions.with_raw_response
-    with patch.object(raw_response, "create") as mock_create, contextlib.suppress(Exception):
+    with patch.object(client, "post") as mock_create, contextlib.suppress(Exception):
         litellm.completion(
             model="gpt-4o",
             messages=[{"role": "user", "content": "hello"}],
@@ -4050,7 +4057,7 @@ def test_openai_model_without_a_provider_still_routes_to_openai():
 
 
 def _openai_chat_create_kwargs(client, **completion_kwargs):
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_client:
+    with patch.object(client, "post") as mock_client:
         with contextlib.suppress(Exception):
             litellm.completion(
                 messages=[{"role": "system", "content": "sys"}, {"role": "user", "content": "hi"}],
@@ -4075,7 +4082,7 @@ def test_completion_custom_api_base_sends_no_prompt_cache_breakpoint_for_gpt_5_6
     from openai import OpenAI
 
     client = OpenAI(api_key="fake-api-key", base_url="http://127.0.0.1:9/v1")
-    request_body = _openai_chat_create_kwargs(client, model="gpt-5.6", api_base="http://127.0.0.1:9/v1")
+    request_body = _openai_chat_create_kwargs(client, model="gpt-5.6", api_base="http://127.0.0.1:9/v1")["body"]
 
     assert request_body["messages"][0] == {"role": "system", "content": "sys", "cache_control": {"type": "ephemeral"}}
     assert "prompt_cache_breakpoint" not in json.dumps(request_body["messages"])
@@ -4087,7 +4094,7 @@ def test_completion_custom_base_url_sends_no_prompt_cache_breakpoint_for_gpt_5_6
     from openai import OpenAI
 
     client = OpenAI(api_key="fake-api-key", base_url="http://127.0.0.1:9/v1")
-    request_body = _openai_chat_create_kwargs(client, model="gpt-5.6", base_url="http://127.0.0.1:9/v1")
+    request_body = _openai_chat_create_kwargs(client, model="gpt-5.6", base_url="http://127.0.0.1:9/v1")["body"]
 
     assert request_body["messages"][0] == {"role": "system", "content": "sys", "cache_control": {"type": "ephemeral"}}
     assert "prompt_cache_breakpoint" not in json.dumps(request_body["messages"])
@@ -4100,7 +4107,7 @@ async def test_acompletion_custom_base_url_sends_no_prompt_cache_breakpoint_for_
     from openai import AsyncOpenAI
 
     client = AsyncOpenAI(api_key="fake-api-key", base_url="http://127.0.0.1:9/v1")
-    with patch.object(client.chat.completions.with_raw_response, "create") as mock_create:
+    with patch.object(client, "post") as mock_create:
         with contextlib.suppress(Exception):
             await litellm.acompletion(
                 model="gpt-5.6",
@@ -4111,7 +4118,7 @@ async def test_acompletion_custom_base_url_sends_no_prompt_cache_breakpoint_for_
             )
 
         mock_create.assert_called_once()
-        request_body = mock_create.call_args.kwargs
+        request_body = mock_create.call_args.kwargs["body"]
 
     assert request_body["messages"][0] == {"role": "system", "content": "sys", "cache_control": {"type": "ephemeral"}}
     assert "prompt_cache_breakpoint" not in json.dumps(request_body["messages"])
@@ -4123,12 +4130,12 @@ def test_completion_default_api_base_sends_prompt_cache_breakpoint_for_gpt_5_6()
     from openai import OpenAI
 
     client = OpenAI(api_key="fake-api-key")
-    request_body = _openai_chat_create_kwargs(client, model="gpt-5.6")
+    request = _openai_chat_create_kwargs(client, model="gpt-5.6")
 
-    assert request_body["messages"][0]["content"] == [
+    assert request["body"]["messages"][0]["content"] == [
         {"type": "text", "text": "sys", "prompt_cache_breakpoint": {"mode": "explicit"}}
     ]
-    assert request_body["extra_body"]["prompt_cache_options"] == {"mode": "implicit"}
+    assert request["options"]["extra_json"]["prompt_cache_options"] == {"mode": "implicit"}
 
 
 _SUBSCRIPTION_OAUTH_CREDENTIAL = "Bearer sk-ant-oat01-fake-subscription-token-for-testing-0123456789"
@@ -5141,6 +5148,7 @@ def test_drop_params_false_still_rejects_an_invalid_stream_chunk_size() -> None:
 
 def test_acompletion_params():
     import inspect
+
     from litellm.types.completion import CompletionRequest
 
     acompletion_params_odict = inspect.signature(acompletion).parameters
@@ -5171,7 +5179,7 @@ def test_acompletion_params():
 
 
 def _openai_mock_response(*args: object, **kwargs: object) -> MagicMock:
-    new_response: Final = MagicMock()
+    new_response: Final = MagicMock(spec=LegacyAPIResponse)
     new_response.headers = {"hello": "world"}
     response_object: Final = {
         "id": "chatcmpl-123",
@@ -5206,7 +5214,7 @@ def test_null_role_response():
 
     openai_client = openai.OpenAI()
     with patch.object(
-        openai_client.chat.completions, "create", side_effect=_openai_mock_response
+        openai_client, "post", side_effect=_openai_mock_response
     ) as mock_response:
         response = litellm.completion(
             model="gpt-3.5-turbo",
@@ -5296,7 +5304,7 @@ def test_completion_perplexity_api():
         pydantic_obj = ChatCompletion(**response_object)
 
         def _return_pydantic_obj(*args, **kwargs):
-            new_response = MagicMock()
+            new_response = MagicMock(spec=LegacyAPIResponse)
             new_response.headers = {"hello": "world"}
 
             new_response.parse.return_value = pydantic_obj
@@ -5305,8 +5313,7 @@ def test_completion_perplexity_api():
         openai_client = OpenAI()
 
         with patch.object(
-            openai_client.chat.completions.with_raw_response,
-            "create",
+            openai_client, "post",
             side_effect=_return_pydantic_obj,
         ) as mock_client:
             # litellm.set_verbose= True
@@ -5350,7 +5357,7 @@ async def test_openai_compatible_custom_api_base(provider):
     openai_client = OpenAI(api_key="fake-key")
 
     with patch.object(
-        openai_client.chat.completions, "create", new=MagicMock()
+        openai_client, "post", new=MagicMock()
     ) as mock_call:
         try:
             completion(
@@ -5368,7 +5375,7 @@ async def test_openai_compatible_custom_api_base(provider):
 
         print("Call KWARGS - {}".format(mock_call.call_args.kwargs))
 
-        assert "hello" in mock_call.call_args.kwargs["extra_body"]
+        assert "hello" in mock_call.call_args.kwargs["options"]["extra_json"]
 
 
 @pytest.mark.parametrize(
@@ -5401,7 +5408,7 @@ async def test_openai_compatible_custom_api_video(provider):
     openai_client = OpenAI(api_key="fake-key")
 
     with patch.object(
-        openai_client.chat.completions, "create", new=MagicMock()
+        openai_client, "post", new=MagicMock()
     ) as mock_call:
         try:
             completion(
@@ -7264,7 +7271,7 @@ def _openai_hallucinated_tool_call_mock_response(
     *args: object,
     **kwargs: object,
 ) -> MagicMock:
-    new_response: Final = MagicMock()
+    new_response: Final = MagicMock(spec=LegacyAPIResponse)
     new_response.headers = {"hello": "world"}
     response_object: Final = {
         "id": "chatcmpl-123",
@@ -7322,8 +7329,8 @@ def test_openai_hallucinated_tool_call():
 
     openai_client = openai.OpenAI()
     with patch.object(
-        openai_client.chat.completions,
-        "create",
+        openai_client,
+        "post",
         side_effect=_openai_hallucinated_tool_call_mock_response,
     ) as mock_response:
         response = litellm.completion(
@@ -7445,7 +7452,7 @@ def test_completion_novita_ai():
     openai_client = OpenAI(api_key="fake-key")
 
     with patch.object(
-        openai_client.chat.completions.with_raw_response, "create"
+        openai_client, "post", return_value=MagicMock(spec=LegacyAPIResponse)
     ) as mock_call:
         mock_call.return_value.headers = {}
         mock_call.return_value.parse.return_value = litellm.ModelResponse(
@@ -7464,11 +7471,11 @@ def test_completion_novita_ai():
 
             # Verify model is passed correctly
             assert (
-                mock_call.call_args.kwargs["model"]
+                mock_call.call_args.kwargs["body"]["model"]
                 == "meta-llama/llama-3.3-70b-instruct"
             )
             # Verify messages are passed correctly
-            assert mock_call.call_args.kwargs["messages"] == messages
+            assert mock_call.call_args.kwargs["body"]["messages"] == messages
 
         except Exception as e:
             pytest.fail(f"Error occurred: {e}")
@@ -7491,8 +7498,8 @@ def test_completion_novita_ai_dynamic_params(api_key):
         openai_client = OpenAI(api_key="fake-key")
 
         with patch.object(
-            openai_client.chat.completions,
-            "create",
+            openai_client,
+            "post",
             side_effect=Exception("Invalid API key"),
         ) as mock_call:
             with pytest.raises(Exception, match="Invalid API key") as exc_info:
@@ -7524,7 +7531,7 @@ def test_completion_openai_metadata(monkeypatch, enable_preview_features):
 
     monkeypatch.setattr(litellm, "enable_preview_features", enable_preview_features)
     with patch.object(
-        client.chat.completions.with_raw_response, "create", return_value=MagicMock()
+        client, "post", return_value=MagicMock()
     ) as mock_completion:
         try:
             resp = litellm.completion(
@@ -7538,11 +7545,11 @@ def test_completion_openai_metadata(monkeypatch, enable_preview_features):
 
         mock_completion.assert_called_once()
         if enable_preview_features:
-            assert mock_completion.call_args.kwargs["metadata"] == {
+            assert mock_completion.call_args.kwargs["body"]["metadata"] == {
                 "my-test-key": "my-test-value"
             }
         else:
-            assert "metadata" not in mock_completion.call_args.kwargs
+            assert "metadata" not in mock_completion.call_args.kwargs["body"]
 
 
 AZURE_TTS_BASE: Final = "https://tts.example.azure.com"