From c7c382402a6ec0e26d32b45b30a2ee1c10ad64d5 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 12:04:44 -0700
Subject: [PATCH 1/6] feat(proxy): add /v1/responses/input_tokens token
 counting endpoint

---
 litellm/proxy/_types.py                       |   3 +
 .../proxy/response_api_endpoints/endpoints.py | 157 ++++++++++++-
 .../response_api_endpoints/test_endpoints.py  | 215 +++++++++++++-----
 ui/litellm-dashboard/src/lib/http/schema.d.ts | 153 +++++++++++++
 4 files changed, 467 insertions(+), 61 deletions(-)

diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py
index 0f2d97b8b1c..9574eb36ee5 100644
--- a/litellm/proxy/_types.py
+++ b/litellm/proxy/_types.py
@@ -422,6 +422,9 @@ class LiteLLMRoutes(enum.Enum):
         "/responses/{response_id}/cancel",
         "/v1/responses/{response_id}/cancel",
         "/openai/v1/responses/{response_id}/cancel",
+        "/responses/input_tokens",
+        "/v1/responses/input_tokens",
+        "/openai/v1/responses/input_tokens",
         # vector stores
         "/vector_stores",
         "/v1/vector_stores",
diff --git a/litellm/proxy/response_api_endpoints/endpoints.py b/litellm/proxy/response_api_endpoints/endpoints.py
index 5e56e822484..100b42a9e2a 100644
--- a/litellm/proxy/response_api_endpoints/endpoints.py
+++ b/litellm/proxy/response_api_endpoints/endpoints.py
@@ -1,14 +1,18 @@
 import asyncio
 import json
 import time
-from collections.abc import AsyncIterator, Mapping
+from collections.abc import AsyncIterator, Awaitable, Mapping
+from enum import Enum
 from types import MappingProxyType
-from typing import TYPE_CHECKING, Any, Final, NamedTuple, cast, get_args
+from typing import TYPE_CHECKING, Any, Final, NamedTuple, Protocol, cast, get_args
 from uuid import uuid4
 
 import fastapi
 from fastapi import APIRouter, Depends, HTTPException, Request, Response
+from fastapi.responses import JSONResponse
+from openai.types.responses.response_create_params import ResponseInputParam
 from starlette.websockets import WebSocket, WebSocketDisconnect
+from typing_extensions import ReadOnly, TypedDict
 
 from litellm._logging import verbose_proxy_logger
 from litellm.integrations.custom_guardrail import ModifyResponseException
@@ -26,8 +30,13 @@ from litellm.proxy.common_utils.http_parsing_utils import (
     _read_request_body,
     _safe_set_request_parsed_body,
 )
-from litellm.types.llms.openai import REASONING_EFFORT, ResponsesAPIResponse
+from litellm.types.llms.openai import (
+    REASONING_EFFORT,
+    ResponsesAPIOptionalRequestParams,
+    ResponsesAPIResponse,
+)
 from litellm.types.responses.main import DeleteResponseResult
+from litellm.types.utils import TokenCountResponse
 
 if TYPE_CHECKING:
     from litellm.router import Router
@@ -35,7 +44,7 @@ if TYPE_CHECKING:
 router: Final = APIRouter()
 
 _user_api_key_auth_dep: Final = Depends(user_api_key_auth)
-_RESPONSES_TAGS: Final = ["responses"]  # mutable-ok: fastapi's route signature requires List[str] tags
+_RESPONSES_TAGS: Final[list[str | Enum]] = ["responses"]  # mutable-ok: fastapi's route signature requires list tags
 
 _TOOL_PAYLOAD_KEYS: Final[Mapping[str, tuple[str, ...]]] = MappingProxyType(
     {
@@ -1017,6 +1026,146 @@ async def compact_response(
         )
 
 
+class _ResponsesApiErrorDetail(TypedDict):
+    message: ReadOnly[str]
+    type: ReadOnly[str]
+    param: ReadOnly[str | None]
+    code: ReadOnly[str | None]
+
+
+class _ResponsesApiErrorBody(TypedDict):
+    error: ReadOnly[_ResponsesApiErrorDetail]
+
+
+class _ResponsesInputTokensResult(TypedDict):
+    object: ReadOnly[str]
+    input_tokens: ReadOnly[int]
+
+
+class _TokenCountPayload(TypedDict):
+    model: ReadOnly[str]
+    messages: ReadOnly[tuple[Mapping[str, object], ...]]
+    tools: ReadOnly[object]
+
+
+class _TokenCounter(Protocol):
+    def __call__(self, request: TokenCountRequest, call_endpoint: bool) -> Awaitable[TokenCountResponse]: ...
+
+
+def _proxy_token_counter() -> _TokenCounter:
+    from litellm.proxy.proxy_server import token_counter
+
+    return token_counter
+
+
+_token_counter_dep: Final = Depends(_proxy_token_counter)
+
+
+def _responses_invalid_request_response(message: str, param: str | None, code: str | None) -> JSONResponse:
+    body: Final[_ResponsesApiErrorBody] = {
+        "error": {
+            "message": message,
+            "type": "invalid_request_error",
+            "param": param,
+            "code": code,
+        }
+    }
+    return JSONResponse(status_code=400, content=body)
+
+
+def _missing_responses_param_response(param: str) -> JSONResponse:
+    return _responses_invalid_request_response(
+        message=f"Missing required parameter: '{param}'.",
+        param=param,
+        code="missing_required_parameter",
+    )
+
+
+def _responses_input_as_token_count_messages(
+    input_value: str | ResponseInputParam,
+    instructions: str | None,
+) -> tuple[Mapping[str, object], ...]:
+    from litellm.responses.litellm_completion_transformation.transformation import (
+        LiteLLMCompletionResponsesConfig,
+    )
+
+    request_params: Final[ResponsesAPIOptionalRequestParams] = {"instructions": instructions}
+    transformed: Final = LiteLLMCompletionResponsesConfig.transform_responses_api_input_to_messages(
+        input=input_value,
+        responses_api_request=request_params,
+    )
+    return tuple(
+        message if isinstance(message, dict) else message.model_dump(exclude_none=True) for message in transformed
+    )
+
+
+@router.post(
+    "/v1/responses/input_tokens",
+    dependencies=(_user_api_key_auth_dep,),
+    tags=_RESPONSES_TAGS,
+)
+@router.post(
+    "/responses/input_tokens",
+    dependencies=(_user_api_key_auth_dep,),
+    tags=_RESPONSES_TAGS,
+)
+@router.post(
+    "/openai/v1/responses/input_tokens",
+    dependencies=(_user_api_key_auth_dep,),
+    tags=_RESPONSES_TAGS,
+)
+async def responses_input_tokens(
+    request: Request,
+    token_counter: _TokenCounter = _token_counter_dep,
+):
+    """
+    Count the input tokens of a Responses API request without calling the model.
+
+    Follows the OpenAI Responses API spec: https://platform.openai.com/docs/api-reference/responses/input-tokens
+
+    ```bash
+    curl -X POST http://localhost:4000/v1/responses/input_tokens \
+    -H "Content-Type: application/json" \
+    -H "Authorization: Bearer sk-1234" \
+    -d '{
+        "model": "gpt-4o",
+        "input": "Hello, how are you?"
+    }'
+    ```
+
+    Returns: `{"object": "response.input_tokens", "input_tokens": <count>}`
+    """
+    data: Final = await _read_request_body(request=request)
+    model_name: Final = data.get("model")
+    input_value: Final = data.get("input")
+    if not isinstance(model_name, str) or not model_name:
+        return _missing_responses_param_response("model")
+    if input_value is None:
+        return _missing_responses_param_response("input")
+
+    try:
+        payload: Final[_TokenCountPayload] = {
+            "model": model_name,
+            "messages": _responses_input_as_token_count_messages(
+                input_value=input_value,
+                instructions=data.get("instructions"),
+            ),
+            "tools": data.get("tools"),
+        }
+        token_request: Final = TokenCountRequest.model_validate(payload)
+    except Exception as e:
+        return _responses_invalid_request_response(
+            message=f"Invalid request for token counting: {e}", param=None, code=None
+        )
+
+    token_response: Final = await token_counter(request=token_request, call_endpoint=True)
+    result: Final[_ResponsesInputTokensResult] = {
+        "object": "response.input_tokens",
+        "input_tokens": token_response.total_tokens,
+    }
+    return result
+
+
 @router.post(
     "/v1/responses/{response_id}/cancel",
     dependencies=[Depends(user_api_key_auth)],
diff --git a/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py b/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
index 791d64c6428..dc43e7f5c06 100644
--- a/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
+++ b/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
@@ -82,11 +82,7 @@ class TestResponsesAPIEndpoints(unittest.TestCase):
                 ResponseOutputMessage(
                     type="message",
                     role="assistant",
-                    content=[
-                        ResponseOutputText(
-                            type="output_text", text="Hello from Cursor!"
-                        )
-                    ],
+                    content=[ResponseOutputText(type="output_text", text="Hello from Cursor!")],
                 )
             ],
         )
@@ -121,9 +117,7 @@ class TestResponsesAPIEndpoints(unittest.TestCase):
     @pytest.mark.asyncio
     @patch("litellm.proxy.proxy_server.llm_router")
     @patch("litellm.proxy.proxy_server.user_api_key_auth")
-    async def test_responses_api_key_spend_header_includes_response_cost(
-        self, mock_auth, mock_router
-    ):
+    async def test_responses_api_key_spend_header_includes_response_cost(self, mock_auth, mock_router):
         """
         Test that x-litellm-key-spend header includes the current request's response_cost
         for /v1/responses endpoint.
@@ -159,9 +153,7 @@ class TestResponsesAPIEndpoints(unittest.TestCase):
                 ResponseOutputMessage(
                     type="message",
                     role="assistant",
-                    content=[
-                        ResponseOutputText(type="output_text", text="Test response")
-                    ],
+                    content=[ResponseOutputText(type="output_text", text="Test response")],
                 )
             ],
         )
@@ -356,6 +348,7 @@ class TestWSModelExtraction:
         from litellm.proxy.response_api_endpoints.endpoints import (
             _extract_model_from_first_ws_event,
         )
+
         event = {"type": "response.create", "model": "gpt-4o", "input": "hello"}
         assert _extract_model_from_first_ws_event(event) == "gpt-4o"
 
@@ -363,6 +356,7 @@ class TestWSModelExtraction:
         from litellm.proxy.response_api_endpoints.endpoints import (
             _extract_model_from_first_ws_event,
         )
+
         event = {"type": "response.create", "response": {"model": "gpt-4o", "input": "hello"}}
         assert _extract_model_from_first_ws_event(event) == "gpt-4o"
 
@@ -370,6 +364,7 @@ class TestWSModelExtraction:
         from litellm.proxy.response_api_endpoints.endpoints import (
             _extract_model_from_first_ws_event,
         )
+
         event = {
             "type": "response.create",
             "model": "flat-model",
@@ -381,6 +376,7 @@ class TestWSModelExtraction:
         from litellm.proxy.response_api_endpoints.endpoints import (
             _extract_model_from_first_ws_event,
         )
+
         event = {"type": "response.create", "input": "hello"}
         assert _extract_model_from_first_ws_event(event) is None
 
@@ -400,9 +396,7 @@ class TestResponsesWSFirstFrameValidation:
         )
 
         ws = MagicMock()
-        ws.receive_text = AsyncMock(
-            return_value=json.dumps({"type": "session.update", "model": "gpt-4o"})
-        )
+        ws.receive_text = AsyncMock(return_value=json.dumps({"type": "session.update", "model": "gpt-4o"}))
         ws.send_text = AsyncMock()
         ws.close = AsyncMock()
 
@@ -412,10 +406,7 @@ class TestResponsesWSFirstFrameValidation:
         ws.send_text.assert_awaited_once()
         ws.close.assert_awaited_once_with(code=1008, reason="Invalid first message")
         error_payload = json.loads(ws.send_text.await_args.args[0])
-        assert (
-            error_payload["error"]["message"]
-            == "First message must be a response.create JSON object."
-        )
+        assert error_payload["error"]["message"] == "First message must be a response.create JSON object."
 
     @pytest.mark.asyncio
     async def test_rejects_non_object_json_first_frame(self):
@@ -484,16 +475,12 @@ class TestResponsesWSFirstFrameModelAuth:
         ws.url = "ws://testserver/v1/responses"
         ws.accept = AsyncMock()
         ws.receive_text = AsyncMock(
-            return_value=json.dumps(
-                {"type": "response.create", "model": "gpt-4o-mini", "input": []}
-            )
+            return_value=json.dumps({"type": "response.create", "model": "gpt-4o-mini", "input": []})
         )
         ws.close = AsyncMock()
 
         processor = MagicMock()
-        processor.common_processing_pre_call_logic = AsyncMock(
-            return_value=({"model": "gpt-4o-mini"}, MagicMock())
-        )
+        processor.common_processing_pre_call_logic = AsyncMock(return_value=({"model": "gpt-4o-mini"}, MagicMock()))
 
         async def fake_llm_call():
             return None
@@ -529,9 +516,7 @@ class TestResponsesWSFirstFrameModelAuth:
             _enforce_responses_ws_first_frame_model_auth,
         )
 
-        request = Request(
-            {"type": "http", "method": "POST", "path": "/v1/responses", "headers": []}
-        )
+        request = Request({"type": "http", "method": "POST", "path": "/v1/responses", "headers": []})
         user_api_key_dict = MagicMock()
         llm_router = MagicMock()
 
@@ -593,9 +578,7 @@ class TestReadWSModelFromFirstFrameErrors:
 
         assert result is None
         ws.send_text.assert_not_awaited()
-        ws.close.assert_awaited_once_with(
-            code=1008, reason="Timed out waiting for first message"
-        )
+        ws.close.assert_awaited_once_with(code=1008, reason="Timed out waiting for first message")
 
     @pytest.mark.asyncio
     async def test_invalid_json_sends_error_and_closes(self):
@@ -613,9 +596,7 @@ class TestReadWSModelFromFirstFrameErrors:
         assert result is None
         payload = json.loads(ws.send_text.await_args.args[0])
         assert payload["error"]["message"] == "First message is not valid JSON."
-        ws.close.assert_awaited_once_with(
-            code=1008, reason="Invalid JSON in first message"
-        )
+        ws.close.assert_awaited_once_with(code=1008, reason="Invalid JSON in first message")
 
     @pytest.mark.asyncio
     async def test_missing_model_sends_error_and_closes(self):
@@ -624,9 +605,7 @@ class TestReadWSModelFromFirstFrameErrors:
         )
 
         ws = MagicMock()
-        ws.receive_text = AsyncMock(
-            return_value=json.dumps({"type": "response.create", "input": []})
-        )
+        ws.receive_text = AsyncMock(return_value=json.dumps({"type": "response.create", "input": []}))
         ws.send_text = AsyncMock()
         ws.close = AsyncMock()
 
@@ -679,10 +658,7 @@ class TestManagedResponsesSameProvider:
         assert self._handler("gpt-4o")._same_provider("gpt-4o-mini") is True
 
     def test_different_provider_is_not_same(self):
-        assert (
-            self._handler("gpt-4o")._same_provider("vertex_ai/gemini-2.0-flash")
-            is False
-        )
+        assert self._handler("gpt-4o")._same_provider("vertex_ai/gemini-2.0-flash") is False
 
     def test_inject_credentials_keeps_provider_for_same_provider_model(self):
         handler = self._handler("gpt-4o", custom_llm_provider="openai")
@@ -697,18 +673,14 @@ class TestManagedResponsesSameProvider:
         assert "custom_llm_provider" not in call_kwargs
 
     def test_unresolvable_connection_model_falls_back_to_custom_provider(self):
-        handler = self._handler(
-            "my-custom-deployment", custom_llm_provider="openai"
-        )
+        handler = self._handler("my-custom-deployment", custom_llm_provider="openai")
         assert handler._same_provider("gpt-4o-mini") is True
         call_kwargs: dict = {}
         handler._inject_credentials(call_kwargs, model="gpt-4o-mini")
         assert call_kwargs["custom_llm_provider"] == "openai"
 
     def test_unresolvable_connection_model_still_drops_cross_provider(self):
-        handler = self._handler(
-            "my-custom-deployment", custom_llm_provider="openai"
-        )
+        handler = self._handler("my-custom-deployment", custom_llm_provider="openai")
         call_kwargs: dict = {}
         handler._inject_credentials(call_kwargs, model="vertex_ai/gemini-2.0-flash")
         assert "custom_llm_provider" not in call_kwargs
@@ -840,9 +812,7 @@ def test_cursor_chat_completions_input_body_uses_responses_pipeline_and_strips_s
                     type="message",
                     role="assistant",
                     status="completed",
-                    content=[
-                        ResponseOutputText(type="output_text", text="agent reply", annotations=[])
-                    ],
+                    content=[ResponseOutputText(type="output_text", text="agent reply", annotations=[])],
                 )
             ],
         )
@@ -851,9 +821,12 @@ def test_cursor_chat_completions_input_body_uses_responses_pipeline_and_strips_s
 
     app.dependency_overrides[user_api_key_auth] = _auth_override
     try:
-        with patch.object(ps, "llm_router", mock_router), patch(
-            "litellm.proxy.response_api_endpoints.endpoints._read_request_body",
-            side_effect=capturing_read_request_body,
+        with (
+            patch.object(ps, "llm_router", mock_router),
+            patch(
+                "litellm.proxy.response_api_endpoints.endpoints._read_request_body",
+                side_effect=capturing_read_request_body,
+            ),
         ):
             client = TestClient(app)
             response = client.post(
@@ -1488,8 +1461,8 @@ def _router_serving_only(base_model: str) -> MagicMock:
     mock_router.router_general_settings.pass_through_all_models = False
     mock_router.default_deployment = None
     mock_router.pattern_router.patterns = {base_model: ["anthropic/*"]}
-    mock_router.pattern_router.get_pattern.side_effect = (
-        lambda model: [{"model_name": "anthropic/*"}] if model == base_model else None
+    mock_router.pattern_router.get_pattern.side_effect = lambda model: (
+        [{"model_name": "anthropic/*"}] if model == base_model else None
     )
     return mock_router
 
@@ -1739,9 +1712,7 @@ class TestCursorGateRecognizesRoutingGroups:
         from litellm.proxy.response_api_endpoints.endpoints import _resolve_cursor_model_variant
 
         router = Router(
-            model_list=[
-                {"model_name": "member-fast", "litellm_params": {"model": "openai/gpt-4o", "api_key": "fake"}}
-            ],
+            model_list=[{"model_name": "member-fast", "litellm_params": {"model": "openai/gpt-4o", "api_key": "fake"}}],
             routing_groups=[
                 {"group_name": "grouped-thinking-high", "models": ["member-fast"], "routing_strategy": "simple-shuffle"}
             ],
@@ -1836,3 +1807,133 @@ class TestGuardrailBlockedResponsesUsage:
         assert usage["input_tokens"] == 0
         assert usage["output_tokens"] == 0
         assert usage["total_tokens"] == 0
+
+
+class TestResponsesInputTokens:
+    """Regression tests for POST /v1/responses/input_tokens.
+
+    The docs promise OpenAI-format token counting on the proxy, but the route was
+    never registered, so the POST fell through to the GET/DELETE-only
+    /v1/responses/{response_id} route and returned 405."""
+
+    def _post_input_tokens(self, body, path="/v1/responses/input_tokens", counter=None):
+        from litellm.proxy._types import UserAPIKeyAuth
+        from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
+        from litellm.proxy.response_api_endpoints.endpoints import _proxy_token_counter
+        from litellm.types.utils import TokenCountResponse
+
+        token_counter_mock = (
+            counter
+            if counter is not None
+            else AsyncMock(
+                return_value=TokenCountResponse(
+                    total_tokens=13,
+                    request_model=body.get("model", ""),
+                    model_used=body.get("model", ""),
+                    tokenizer_type="openai_api",
+                )
+            )
+        )
+        app.dependency_overrides[user_api_key_auth] = lambda: UserAPIKeyAuth(api_key="sk-test", request_route=path)
+        app.dependency_overrides[_proxy_token_counter] = lambda: token_counter_mock
+        try:
+            client = TestClient(app)
+            response = client.post(path, json=body, headers={"Authorization": "Bearer sk-1234"})
+            return response, token_counter_mock
+        finally:
+            app.dependency_overrides.pop(user_api_key_auth, None)
+            app.dependency_overrides.pop(_proxy_token_counter, None)
+
+    def test_string_input_returns_openai_input_tokens_shape(self):
+        response, counter = self._post_input_tokens({"model": "gpt-4o", "input": "Hello, how are you?"})
+
+        assert response.status_code == 200, response.text
+        assert response.json() == {"object": "response.input_tokens", "input_tokens": 13}
+        counter.assert_awaited_once()
+        assert counter.call_args.kwargs["call_endpoint"] is True
+        token_request = counter.call_args.kwargs["request"]
+        assert token_request.model == "gpt-4o"
+        assert token_request.messages == [{"role": "user", "content": "Hello, how are you?"}]
+
+    def test_every_route_alias_is_registered(self):
+        for path in ("/v1/responses/input_tokens", "/responses/input_tokens", "/openai/v1/responses/input_tokens"):
+            response, _ = self._post_input_tokens({"model": "gpt-4o", "input": "hi"}, path=path)
+            assert response.status_code == 200, f"{path}: {response.status_code} {response.text}"
+
+    def test_input_items_instructions_and_tools_are_forwarded(self):
+        tools = [
+            {
+                "type": "function",
+                "name": "get_weather",
+                "description": "Get weather for a city",
+                "parameters": {"type": "object", "properties": {"city": {"type": "string"}}},
+            }
+        ]
+        response, counter = self._post_input_tokens(
+            {
+                "model": "gpt-4o",
+                "input": [{"role": "user", "content": "What is the weather in Paris?"}],
+                "instructions": "You are terse.",
+                "tools": tools,
+            }
+        )
+
+        assert response.status_code == 200, response.text
+        token_request = counter.call_args.kwargs["request"]
+        assert token_request.messages == [
+            {"role": "system", "content": "You are terse."},
+            {"role": "user", "content": "What is the weather in Paris?"},
+        ]
+        assert token_request.tools == tools
+
+    def test_missing_model_returns_openai_400(self):
+        response, counter = self._post_input_tokens({"input": "Hello"})
+
+        assert response.status_code == 400, response.text
+        assert response.json() == {
+            "error": {
+                "message": "Missing required parameter: 'model'.",
+                "type": "invalid_request_error",
+                "param": "model",
+                "code": "missing_required_parameter",
+            }
+        }
+        counter.assert_not_awaited()
+
+    def test_missing_input_returns_openai_400(self):
+        response, counter = self._post_input_tokens({"model": "gpt-4o"})
+
+        assert response.status_code == 400, response.text
+        assert response.json() == {
+            "error": {
+                "message": "Missing required parameter: 'input'.",
+                "type": "invalid_request_error",
+                "param": "input",
+                "code": "missing_required_parameter",
+            }
+        }
+        counter.assert_not_awaited()
+
+    def test_invalid_tools_returns_openai_400(self):
+        response, counter = self._post_input_tokens({"model": "gpt-4o", "input": "hi", "tools": "not-a-list"})
+
+        assert response.status_code == 400, response.text
+        error = response.json()["error"]
+        assert error["type"] == "invalid_request_error"
+        counter.assert_not_awaited()
+
+    def test_provider_error_maps_status_code(self):
+        from litellm.proxy._types import ProxyException
+
+        failing_counter = AsyncMock(
+            side_effect=ProxyException(
+                message="rate limited",
+                type="token_counting_error",
+                param="model",
+                code="429",
+            )
+        )
+        response, _ = self._post_input_tokens({"model": "gpt-4o", "input": "hi"}, counter=failing_counter)
+
+        assert response.status_code == 429, response.text
+        assert response.json()["error"]["message"] == "rate limited"
diff --git a/ui/litellm-dashboard/src/lib/http/schema.d.ts b/ui/litellm-dashboard/src/lib/http/schema.d.ts
index 81a2fefe32e..01a1b516d11 100644
--- a/ui/litellm-dashboard/src/lib/http/schema.d.ts
+++ b/ui/litellm-dashboard/src/lib/http/schema.d.ts
@@ -9700,6 +9700,37 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/openai/v1/responses/input_tokens": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /**
+         * Responses Input Tokens
+         * @description Count the input tokens of a Responses API request without calling the model.
+         *
+         *     Follows the OpenAI Responses API spec: https://platform.openai.com/docs/api-reference/responses/input-tokens
+         *
+         *     ```bash
+         *     curl -X POST http://localhost:4000/v1/responses/input_tokens     -H "Content-Type: application/json"     -H "Authorization: Bearer sk-1234"     -d '{
+         *         "model": "gpt-4o",
+         *         "input": "Hello, how are you?"
+         *     }'
+         *     ```
+         *
+         *     Returns: `{"object": "response.input_tokens", "input_tokens": <count>}`
+         */
+        post: operations["responses_input_tokens_openai_v1_responses_input_tokens_post"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/openai/v1/responses/{response_id}": {
         parameters: {
             query?: never;
@@ -12619,6 +12650,37 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/responses/input_tokens": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /**
+         * Responses Input Tokens
+         * @description Count the input tokens of a Responses API request without calling the model.
+         *
+         *     Follows the OpenAI Responses API spec: https://platform.openai.com/docs/api-reference/responses/input-tokens
+         *
+         *     ```bash
+         *     curl -X POST http://localhost:4000/v1/responses/input_tokens     -H "Content-Type: application/json"     -H "Authorization: Bearer sk-1234"     -d '{
+         *         "model": "gpt-4o",
+         *         "input": "Hello, how are you?"
+         *     }'
+         *     ```
+         *
+         *     Returns: `{"object": "response.input_tokens", "input_tokens": <count>}`
+         */
+        post: operations["responses_input_tokens_responses_input_tokens_post"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/responses/{response_id}": {
         parameters: {
             query?: never;
@@ -19184,6 +19246,37 @@ export interface paths {
         patch?: never;
         trace?: never;
     };
+    "/v1/responses/input_tokens": {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        get?: never;
+        put?: never;
+        /**
+         * Responses Input Tokens
+         * @description Count the input tokens of a Responses API request without calling the model.
+         *
+         *     Follows the OpenAI Responses API spec: https://platform.openai.com/docs/api-reference/responses/input-tokens
+         *
+         *     ```bash
+         *     curl -X POST http://localhost:4000/v1/responses/input_tokens     -H "Content-Type: application/json"     -H "Authorization: Bearer sk-1234"     -d '{
+         *         "model": "gpt-4o",
+         *         "input": "Hello, how are you?"
+         *     }'
+         *     ```
+         *
+         *     Returns: `{"object": "response.input_tokens", "input_tokens": <count>}`
+         */
+        post: operations["responses_input_tokens_v1_responses_input_tokens_post"];
+        delete?: never;
+        options?: never;
+        head?: never;
+        patch?: never;
+        trace?: never;
+    };
     "/v1/responses/{response_id}": {
         parameters: {
             query?: never;
@@ -51476,6 +51569,26 @@ export interface operations {
             };
         };
     };
+    responses_input_tokens_openai_v1_responses_input_tokens_post: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+        };
+    };
     get_response_openai_v1_responses__response_id__get: {
         parameters: {
             query?: never;
@@ -54440,6 +54553,26 @@ export interface operations {
             };
         };
     };
+    responses_input_tokens_responses_input_tokens_post: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+        };
+    };
     get_response_responses__response_id__get: {
         parameters: {
             query?: never;
@@ -62862,6 +62995,26 @@ export interface operations {
             };
         };
     };
+    responses_input_tokens_v1_responses_input_tokens_post: {
+        parameters: {
+            query?: never;
+            header?: never;
+            path?: never;
+            cookie?: never;
+        };
+        requestBody?: never;
+        responses: {
+            /** @description Successful Response */
+            200: {
+                headers: {
+                    [name: string]: unknown;
+                };
+                content: {
+                    "application/json": unknown;
+                };
+            };
+        };
+    };
     get_response_v1_responses__response_id__get: {
         parameters: {
             query?: never;

From 6b7159323bebc5bb5f2ad6a7b8680942cbfb8ab8 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 12:25:45 -0700
Subject: [PATCH 2/6] fix(proxy): match OpenAI on empty input and skip budget
 reservation for token counting

/v1/responses/input_tokens returned 200 with a count for an empty
"input" ("" or []), while OpenAI returns a 400 missing_required_parameter.
The route also went through optimistic budget reservation, which is only
released by LLM success/failure callbacks that a token count never
reaches, so every call leaked a reservation until TTL expiry and could
429 real traffic. Both routes plus the /openai alias now join
/utils/token_counter in the reservation exemption set.
---
 .../proxy/response_api_endpoints/endpoints.py |  6 +++
 .../spend_tracking/budget_reservation.py      |  9 +++-
 .../response_api_endpoints/test_endpoints.py  | 15 ++++++
 .../spend_tracking/test_budget_reservation.py | 48 +++++++++++++++++++
 4 files changed, 77 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py

diff --git a/litellm/proxy/response_api_endpoints/endpoints.py b/litellm/proxy/response_api_endpoints/endpoints.py
index 100b42a9e2a..aa7595ed13d 100644
--- a/litellm/proxy/response_api_endpoints/endpoints.py
+++ b/litellm/proxy/response_api_endpoints/endpoints.py
@@ -1142,6 +1142,12 @@ async def responses_input_tokens(
         return _missing_responses_param_response("model")
     if input_value is None:
         return _missing_responses_param_response("input")
+    if isinstance(input_value, (str, list)) and not input_value:
+        return _responses_invalid_request_response(
+            message="""One of "input" or "previous_response_id" or 'prompt' or 'conversation' must be provided.""",
+            param=None,
+            code="missing_required_parameter",
+        )
 
     try:
         payload: Final[_TokenCountPayload] = {
diff --git a/litellm/proxy/spend_tracking/budget_reservation.py b/litellm/proxy/spend_tracking/budget_reservation.py
index 8b2a5dd9312..2d113cfe355 100644
--- a/litellm/proxy/spend_tracking/budget_reservation.py
+++ b/litellm/proxy/spend_tracking/budget_reservation.py
@@ -172,7 +172,14 @@ async def reserve_budget_for_request(
 ) -> dict | None:
     if valid_token is None or not RouteChecks.is_llm_api_route(route=route):
         return None
-    if route in {"/models", "/v1/models", "/utils/token_counter"}:
+    if route in {
+        "/models",
+        "/v1/models",
+        "/utils/token_counter",
+        "/responses/input_tokens",
+        "/v1/responses/input_tokens",
+        "/openai/v1/responses/input_tokens",
+    }:
         return None
     if get_model_from_request(request_body, route, llm_router=llm_router) is None:
         return None
diff --git a/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py b/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
index dc43e7f5c06..b31c53c14a8 100644
--- a/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
+++ b/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
@@ -1914,6 +1914,21 @@ class TestResponsesInputTokens:
         }
         counter.assert_not_awaited()
 
+    @pytest.mark.parametrize("empty_input", ["", []])
+    def test_empty_input_returns_openai_400(self, empty_input):
+        response, counter = self._post_input_tokens({"model": "gpt-4o", "input": empty_input})
+
+        assert response.status_code == 400, response.text
+        assert response.json() == {
+            "error": {
+                "message": """One of "input" or "previous_response_id" or 'prompt' or 'conversation' must be provided.""",
+                "type": "invalid_request_error",
+                "param": None,
+                "code": "missing_required_parameter",
+            }
+        }
+        counter.assert_not_awaited()
+
     def test_invalid_tools_returns_openai_400(self):
         response, counter = self._post_input_tokens({"model": "gpt-4o", "input": "hi", "tools": "not-a-list"})
 
diff --git a/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py b/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py
new file mode 100644
index 00000000000..f65f68812a2
--- /dev/null
+++ b/tests/test_litellm/proxy/spend_tracking/test_budget_reservation.py
@@ -0,0 +1,48 @@
+from typing import Final
+
+import pytest
+
+from litellm.caching import DualCache
+from litellm.proxy._types import UserAPIKeyAuth
+from litellm.proxy.common_utils.user_api_key_cache import UserApiKeyCache
+from litellm.proxy.spend_tracking.budget_reservation import reserve_budget_for_request
+from litellm.proxy.utils import ProxyLogging
+
+TOKEN_COUNTING_ROUTES: Final = (
+    "/responses/input_tokens",
+    "/v1/responses/input_tokens",
+    "/openai/v1/responses/input_tokens",
+    "/utils/token_counter",
+)
+
+
+def _budgeted_token() -> UserAPIKeyAuth:
+    return UserAPIKeyAuth(api_key="sk-test", token="hashed-token", max_budget=100.0, spend=0.0)
+
+
+async def _reserve(route: str) -> dict | None:
+    return await reserve_budget_for_request(
+        request_body={"model": "gpt-4o", "input": "hello"},
+        route=route,
+        llm_router=None,
+        valid_token=_budgeted_token(),
+        team_object=None,
+        user_object=None,
+        prisma_client=None,
+        user_api_key_cache=UserApiKeyCache(),
+        proxy_logging_obj=ProxyLogging(user_api_key_cache=DualCache()),
+    )
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("route", TOKEN_COUNTING_ROUTES)
+async def test_token_counting_routes_are_exempt_from_budget_reservation(route):
+    assert await _reserve(route) is None
+
+
+@pytest.mark.asyncio
+async def test_non_exempt_llm_route_still_reserves_budget():
+    reservation: Final = await _reserve("/v1/responses")
+
+    assert reservation is not None
+    assert reservation["reserved_cost"] > 0

From 73ab647b1c26c8b1fcff137c87733e46e8d90326 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 12:43:17 -0700
Subject: [PATCH 3/6] fix(count_tokens): preserve image inputs when counting
 Responses API tokens

The chat-to-Responses reverse transform kept only text blocks, so an image
input was dropped before the count went to OpenAI. A 256x256 image request
counted 13 tokens instead of 268.
---
 .../responses/count_tokens/transformation.py  | 73 ++++++++++++--
 ...test_openai_count_tokens_transformation.py | 97 +++++++++++++++++++
 .../response_api_endpoints/test_endpoints.py  |  9 +-
 3 files changed, 169 insertions(+), 10 deletions(-)

diff --git a/litellm/llms/openai/responses/count_tokens/transformation.py b/litellm/llms/openai/responses/count_tokens/transformation.py
index 6b2f4535df1..72038ae6f6c 100644
--- a/litellm/llms/openai/responses/count_tokens/transformation.py
+++ b/litellm/llms/openai/responses/count_tokens/transformation.py
@@ -4,7 +4,69 @@ OpenAI Responses API token counting transformation logic.
 This module handles the transformation of requests to OpenAI's /v1/responses/input_tokens endpoint.
 """
 
-from typing import Any, Final
+from collections.abc import Mapping, Sequence
+from typing import Any, Final, Literal
+
+from typing_extensions import ReadOnly, TypedDict
+
+
+class ResponsesInputTextPart(TypedDict):
+    type: ReadOnly[Literal["input_text"]]
+    text: ReadOnly[str]
+
+
+class ResponsesInputImagePart(TypedDict):
+    type: ReadOnly[Literal["input_image"]]
+    image_url: ReadOnly[str]
+    detail: ReadOnly[str]
+
+
+ResponsesInputPart = ResponsesInputTextPart | ResponsesInputImagePart
+
+
+def _chat_image_block_to_responses_part(image_url: object) -> ResponsesInputImagePart | None:
+    url: Final = image_url.get("url") if isinstance(image_url, Mapping) else image_url
+    if not isinstance(url, str) or not url:
+        return None
+    detail: Final = image_url.get("detail") if isinstance(image_url, Mapping) else None
+    part: Final[ResponsesInputImagePart] = {
+        "type": "input_image",
+        "image_url": url,
+        "detail": detail if isinstance(detail, str) and detail else "auto",
+    }
+    return part
+
+
+def _chat_block_to_responses_part(block: object) -> ResponsesInputPart | None:
+    if isinstance(block, str):
+        bare: Final[ResponsesInputTextPart] = {"type": "input_text", "text": block}
+        return bare
+    if not isinstance(block, Mapping):
+        return None
+    match block.get("type"):
+        case "text":
+            text_value: Final = block.get("text")
+            text: Final[ResponsesInputTextPart] = {
+                "type": "input_text",
+                "text": text_value if isinstance(text_value, str) else "",
+            }
+            return text
+        case "image_url":
+            return _chat_image_block_to_responses_part(block.get("image_url"))
+        case _:
+            return None
+
+
+def chat_content_blocks_to_responses_content(
+    content: Sequence[object],
+) -> str | tuple[ResponsesInputPart, ...]:
+    """Text-only content collapses to a joined string, so text-only counts stay unchanged."""
+    parts: Final = tuple(
+        part for part in (_chat_block_to_responses_part(block) for block in content) if part is not None
+    )
+    if any(part["type"] != "input_text" for part in parts):
+        return parts
+    return "\n".join(part["text"] for part in parts if part["type"] == "input_text")
 
 
 class OpenAICountTokensConfig:
@@ -120,14 +182,7 @@ class OpenAICountTokensConfig:
                     instructions_parts.append("\n".join(text_parts))
             elif role == "user":
                 if isinstance(content, list):
-                    # Extract text from content blocks for Responses API
-                    text_parts = []
-                    for block in content:
-                        if isinstance(block, dict) and block.get("type") == "text":
-                            text_parts.append(block.get("text", ""))
-                        elif isinstance(block, str):
-                            text_parts.append(block)
-                    content = "\n".join(text_parts)
+                    content = chat_content_blocks_to_responses_content(content)
                 input_items.append({"role": "user", "content": content})
             elif role == "assistant":
                 # Map tool_calls to Responses API function_call items
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
index e1cc6a92927..ca9e7ab52c3 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
@@ -163,6 +163,103 @@ def test_messages_to_responses_input_with_tool():
     }
 
 
+def test_messages_to_responses_input_preserves_images():
+    """An image block must survive the round trip, or OpenAI counts only the text.
+
+    A 256x256 image is worth 255 tokens to OpenAI's counting API; dropping it
+    turned a 268-token request into a 13-token one.
+    """
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What is in this image?"},
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "data:image/png;base64,iVBORw0KGgo=", "detail": "high"},
+                },
+            ],
+        }
+    ]
+
+    input_items, instructions = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert instructions is None
+    assert input_items == [
+        {
+            "role": "user",
+            "content": (
+                {"type": "input_text", "text": "What is in this image?"},
+                {
+                    "type": "input_image",
+                    "image_url": "data:image/png;base64,iVBORw0KGgo=",
+                    "detail": "high",
+                },
+            ),
+        }
+    ]
+
+
+def test_messages_to_responses_input_image_without_detail_defaults_to_auto():
+    messages = [
+        {
+            "role": "user",
+            "content": [{"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}}],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items[0]["content"] == (
+        {"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
+    )
+
+
+def test_messages_to_responses_input_bare_string_image_url_is_preserved():
+    messages = [{"role": "user", "content": [{"type": "image_url", "image_url": "https://example.com/cat.png"}]}]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items[0]["content"] == (
+        {"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
+    )
+
+
+def test_messages_to_responses_input_text_only_blocks_stay_a_joined_string():
+    """Text-only content must keep collapsing to a string so existing counts do not shift."""
+    messages = [
+        {
+            "role": "user",
+            "content": [{"type": "text", "text": "first"}, {"type": "text", "text": "second"}],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [{"role": "user", "content": "first\nsecond"}]
+
+
+def test_messages_to_responses_input_drops_unmappable_blocks():
+    """A block with no Responses API equivalent is skipped, never forwarded verbatim."""
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "hi"},
+                {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
+                {"type": "input_audio", "input_audio": {"data": "AAAA", "format": "wav"}},
+            ],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items[0]["content"] == (
+        {"type": "input_text", "text": "hi"},
+        {"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
+    )
+
+
 def test_validate_request_valid():
     """Test that valid requests pass validation."""
     config = OpenAICountTokensConfig()
diff --git a/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py b/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
index b31c53c14a8..d7010de6405 100644
--- a/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
+++ b/tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py
@@ -3,10 +3,12 @@ Test for response_api_endpoints/endpoints.py
 """
 
 import unittest
+from typing import Any
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 from fastapi.testclient import TestClient
+from httpx import Response
 
 import litellm
 from litellm.proxy.proxy_server import app
@@ -1816,7 +1818,12 @@ class TestResponsesInputTokens:
     never registered, so the POST fell through to the GET/DELETE-only
     /v1/responses/{response_id} route and returned 405."""
 
-    def _post_input_tokens(self, body, path="/v1/responses/input_tokens", counter=None):
+    def _post_input_tokens(
+        self,
+        body: dict[str, Any],
+        path: str = "/v1/responses/input_tokens",
+        counter: AsyncMock | None = None,
+    ) -> tuple[Response, AsyncMock]:
         from litellm.proxy._types import UserAPIKeyAuth
         from litellm.proxy.auth.user_api_key_auth import user_api_key_auth
         from litellm.proxy.response_api_endpoints.endpoints import _proxy_token_counter

From fe90c6f6fca6440f302e21bb31f5816225b96770 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 12:59:40 -0700
Subject: [PATCH 4/6] fix(count_tokens): keep assistant turns on the provider
 counting API

Assistant list content was forwarded to /v1/responses/input_tokens as chat
`text` blocks, which the Responses API rejects (it accepts only output_text
and refusal inside an assistant turn). The 400 sent the whole request to the
local tokenizer, so any conversation with an assistant turn silently lost
provider-exact counting, including the image counting added in 73ab647b1c.

Assistant content now collapses to the plain string the Responses API counts
identically, and image parts are kept to user turns where they are legal.
---
 .../responses/count_tokens/transformation.py  | 19 ++++--
 ...test_openai_count_tokens_transformation.py | 63 +++++++++++++++++++
 2 files changed, 77 insertions(+), 5 deletions(-)

diff --git a/litellm/llms/openai/responses/count_tokens/transformation.py b/litellm/llms/openai/responses/count_tokens/transformation.py
index 72038ae6f6c..62dbebb4fe6 100644
--- a/litellm/llms/openai/responses/count_tokens/transformation.py
+++ b/litellm/llms/openai/responses/count_tokens/transformation.py
@@ -23,6 +23,8 @@ class ResponsesInputImagePart(TypedDict):
 
 ResponsesInputPart = ResponsesInputTextPart | ResponsesInputImagePart
 
+ResponsesContentRole = Literal["user", "assistant"]
+
 
 def _chat_image_block_to_responses_part(image_url: object) -> ResponsesInputImagePart | None:
     url: Final = image_url.get("url") if isinstance(image_url, Mapping) else image_url
@@ -37,7 +39,7 @@ def _chat_image_block_to_responses_part(image_url: object) -> ResponsesInputImag
     return part
 
 
-def _chat_block_to_responses_part(block: object) -> ResponsesInputPart | None:
+def _chat_block_to_responses_part(block: object, role: ResponsesContentRole) -> ResponsesInputPart | None:
     if isinstance(block, str):
         bare: Final[ResponsesInputTextPart] = {"type": "input_text", "text": block}
         return bare
@@ -51,7 +53,7 @@ def _chat_block_to_responses_part(block: object) -> ResponsesInputPart | None:
                 "text": text_value if isinstance(text_value, str) else "",
             }
             return text
-        case "image_url":
+        case "image_url" if role == "user":
             return _chat_image_block_to_responses_part(block.get("image_url"))
         case _:
             return None
@@ -59,10 +61,15 @@ def _chat_block_to_responses_part(block: object) -> ResponsesInputPart | None:
 
 def chat_content_blocks_to_responses_content(
     content: Sequence[object],
+    role: ResponsesContentRole,
 ) -> str | tuple[ResponsesInputPart, ...]:
-    """Text-only content collapses to a joined string, so text-only counts stay unchanged."""
+    """Text-only content collapses to a joined string, which every role accepts and counts identically.
+
+    Only a user turn may carry an image part: the Responses API rejects any part but
+    output_text and refusal inside an assistant turn.
+    """
     parts: Final = tuple(
-        part for part in (_chat_block_to_responses_part(block) for block in content) if part is not None
+        part for part in (_chat_block_to_responses_part(block, role) for block in content) if part is not None
     )
     if any(part["type"] != "input_text" for part in parts):
         return parts
@@ -182,11 +189,13 @@ class OpenAICountTokensConfig:
                     instructions_parts.append("\n".join(text_parts))
             elif role == "user":
                 if isinstance(content, list):
-                    content = chat_content_blocks_to_responses_content(content)
+                    content = chat_content_blocks_to_responses_content(content, "user")
                 input_items.append({"role": "user", "content": content})
             elif role == "assistant":
                 # Map tool_calls to Responses API function_call items
                 tool_calls = msg.get("tool_calls")
+                if isinstance(content, list):
+                    content = chat_content_blocks_to_responses_content(content, "assistant")
                 if content:
                     input_items.append({"role": "assistant", "content": content})
                 if tool_calls:
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
index ca9e7ab52c3..22761272321 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
@@ -260,6 +260,69 @@ def test_messages_to_responses_input_drops_unmappable_blocks():
     )
 
 
+def test_messages_to_responses_input_assistant_blocks_collapse_to_a_string():
+    """An assistant turn must never forward chat `text` blocks.
+
+    The Responses API only accepts output_text and refusal inside an assistant turn, so
+    forwarding them 400s the whole request and silently drops the count back to the local
+    tokenizer, which is exactly what defeats the image fix above.
+    """
+    messages = [
+        {"role": "user", "content": [{"type": "text", "text": "What is the capital of France?"}]},
+        {"role": "assistant", "content": [{"type": "text", "text": "Paris."}]},
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [
+        {"role": "user", "content": "What is the capital of France?"},
+        {"role": "assistant", "content": "Paris."},
+    ]
+
+
+def test_messages_to_responses_input_assistant_image_block_is_dropped():
+    """An image part is illegal inside an assistant turn, so it must not reach the provider."""
+    messages = [
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "text", "text": "Here it is"},
+                {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
+            ],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [{"role": "assistant", "content": "Here it is"}]
+
+
+def test_messages_to_responses_input_keeps_user_image_alongside_an_assistant_turn():
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What is in this image?"},
+                {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
+            ],
+        },
+        {"role": "assistant", "content": [{"type": "text", "text": "A cat."}]},
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [
+        {
+            "role": "user",
+            "content": (
+                {"type": "input_text", "text": "What is in this image?"},
+                {"type": "input_image", "image_url": "https://example.com/cat.png", "detail": "auto"},
+            ),
+        },
+        {"role": "assistant", "content": "A cat."},
+    ]
+
+
 def test_validate_request_valid():
     """Test that valid requests pass validation."""
     config = OpenAICountTokensConfig()

From c9908ffabb732f902f3409a7d763c5ea6a1541a8 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 13:17:43 -0700
Subject: [PATCH 5/6] fix(responses): count input_file tokens instead of
 silently dropping the file

The Responses-to-chat transform dropped the filename OpenAI requires next to
file_data, so a request carrying an inline PDF counted 13 tokens instead of 36
and a real completion through the chat bridge got a 400.
---
 .../responses/count_tokens/transformation.py  | 28 ++++++-
 .../transformation.py                         |  2 +
 ...test_openai_count_tokens_transformation.py | 74 +++++++++++++++++++
 .../test_litellm_completion_responses.py      | 19 +++++
 4 files changed, 121 insertions(+), 2 deletions(-)

diff --git a/litellm/llms/openai/responses/count_tokens/transformation.py b/litellm/llms/openai/responses/count_tokens/transformation.py
index 62dbebb4fe6..88f04c59e01 100644
--- a/litellm/llms/openai/responses/count_tokens/transformation.py
+++ b/litellm/llms/openai/responses/count_tokens/transformation.py
@@ -21,7 +21,13 @@ class ResponsesInputImagePart(TypedDict):
     detail: ReadOnly[str]
 
 
-ResponsesInputPart = ResponsesInputTextPart | ResponsesInputImagePart
+class ResponsesInputFilePart(TypedDict):
+    type: ReadOnly[Literal["input_file"]]
+    filename: ReadOnly[str]
+    file_data: ReadOnly[str]
+
+
+ResponsesInputPart = ResponsesInputTextPart | ResponsesInputImagePart | ResponsesInputFilePart
 
 ResponsesContentRole = Literal["user", "assistant"]
 
@@ -39,6 +45,22 @@ def _chat_image_block_to_responses_part(image_url: object) -> ResponsesInputImag
     return part
 
 
+def _chat_file_block_to_responses_part(file_value: object) -> ResponsesInputFilePart | None:
+    """Only an inline file round trips: OpenAI rejects `file_data` without the `filename` beside it."""
+    if not isinstance(file_value, Mapping):
+        return None
+    filename: Final = file_value.get("filename")
+    file_data: Final = file_value.get("file_data")
+    if not isinstance(filename, str) or not filename or not isinstance(file_data, str) or not file_data:
+        return None
+    part: Final[ResponsesInputFilePart] = {
+        "type": "input_file",
+        "filename": filename,
+        "file_data": file_data,
+    }
+    return part
+
+
 def _chat_block_to_responses_part(block: object, role: ResponsesContentRole) -> ResponsesInputPart | None:
     if isinstance(block, str):
         bare: Final[ResponsesInputTextPart] = {"type": "input_text", "text": block}
@@ -55,6 +77,8 @@ def _chat_block_to_responses_part(block: object, role: ResponsesContentRole) ->
             return text
         case "image_url" if role == "user":
             return _chat_image_block_to_responses_part(block.get("image_url"))
+        case "file" if role == "user":
+            return _chat_file_block_to_responses_part(block.get("file"))
         case _:
             return None
 
@@ -65,7 +89,7 @@ def chat_content_blocks_to_responses_content(
 ) -> str | tuple[ResponsesInputPart, ...]:
     """Text-only content collapses to a joined string, which every role accepts and counts identically.
 
-    Only a user turn may carry an image part: the Responses API rejects any part but
+    Only a user turn may carry an image or file part: the Responses API rejects any part but
     output_text and refusal inside an assistant turn.
     """
     parts: Final = tuple(
diff --git a/litellm/responses/litellm_completion_transformation/transformation.py b/litellm/responses/litellm_completion_transformation/transformation.py
index f39df38d069..4a416f61e1b 100644
--- a/litellm/responses/litellm_completion_transformation/transformation.py
+++ b/litellm/responses/litellm_completion_transformation/transformation.py
@@ -1629,6 +1629,8 @@ class LiteLLMCompletionResponsesConfig:
             file_dict["file_id"] = file_id
         if item.get("file_data"):
             file_dict["file_data"] = item["file_data"]
+        if item.get("filename"):
+            file_dict["filename"] = item["filename"]
 
         new_item: Final[dict[str, object]] = {"type": "file", "file": file_dict}
         if "cache_control" in item:
diff --git a/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py b/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
index 22761272321..c2efc1acdb9 100644
--- a/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
+++ b/tests/test_litellm/llms/openai/responses/test_openai_count_tokens_transformation.py
@@ -323,6 +323,80 @@ def test_messages_to_responses_input_keeps_user_image_alongside_an_assistant_tur
     ]
 
 
+def test_messages_to_responses_input_preserves_inline_files():
+    """An inline file must survive the round trip, or the count silently drops the file.
+
+    A small PDF is worth 36 tokens to OpenAI's counting API; dropping it left the same
+    request counting 13, the text-only total.
+    """
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Summarize this file."},
+                {
+                    "type": "file",
+                    "file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0="},
+                },
+            ],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [
+        {
+            "role": "user",
+            "content": (
+                {"type": "input_text", "text": "Summarize this file."},
+                {
+                    "type": "input_file",
+                    "filename": "report.pdf",
+                    "file_data": "data:application/pdf;base64,JVBERi0=",
+                },
+            ),
+        }
+    ]
+
+
+def test_messages_to_responses_input_drops_a_file_with_no_inline_data():
+    """OpenAI rejects `file_data` without a `filename`, and a rejected request loses the whole count."""
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Summarize this file."},
+                {"type": "file", "file": {"file_data": "data:application/pdf;base64,JVBERi0="}},
+                {"type": "file", "file": {"file_id": "file-abc123"}},
+            ],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [{"role": "user", "content": "Summarize this file."}]
+
+
+def test_messages_to_responses_input_assistant_file_block_is_dropped():
+    """A file part is illegal inside an assistant turn, so it must not reach the provider."""
+    messages = [
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "text", "text": "Here it is"},
+                {
+                    "type": "file",
+                    "file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0="},
+                },
+            ],
+        }
+    ]
+
+    input_items, _ = OpenAICountTokensConfig.messages_to_responses_input(messages)
+
+    assert input_items == [{"role": "assistant", "content": "Here it is"}]
+
+
 def test_validate_request_valid():
     """Test that valid requests pass validation."""
     config = OpenAICountTokensConfig()
diff --git a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
index b96d2eb5322..bb59e576568 100644
--- a/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
+++ b/tests/test_litellm/responses/litellm_completion_transformation/test_litellm_completion_responses.py
@@ -124,6 +124,25 @@ class TestLiteLLMCompletionResponsesConfig:
         assert "extra_field" not in result["file"]
         assert "another_field" not in result["file"]
 
+    def test_transform_input_file_item_to_file_item_keeps_filename(self):
+        """OpenAI rejects file_data with no filename beside it, so dropping it 400s the request"""
+        result = (
+            LiteLLMCompletionResponsesConfig._transform_input_file_item_to_file_item(
+                {
+                    "type": "input_file",
+                    "filename": "report.pdf",
+                    "file_data": "data:application/pdf;base64,JVBERi0=",
+                }
+            )
+        )
+        assert result == {
+            "type": "file",
+            "file": {
+                "file_data": "data:application/pdf;base64,JVBERi0=",
+                "filename": "report.pdf",
+            },
+        }
+
     def test_transform_input_file_item_to_file_item_with_file_url(self):
         """file_url should be mapped to file_id for downstream URL handling"""
         result = (

From b7da4717843e61cf6bd1aeb10b1cbaf70e946e1d Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Mon, 31 Aug 2026 13:53:22 -0700
Subject: [PATCH 6/6] fix(count_tokens): price an inline file block instead of
 raising on it

`ChatCompletionFileObject` is in the union `_count_content_list` accepts, but
`file` was missing from its match, so every local count of a Responses
`input_file` raised `Invalid content item type: file`. On
/v1/responses/input_tokens that surfaced as an opaque 500 whenever the model's
provider counting API refused the block and the local tokenizer took over.

Count it the way the module already counts the same thing in Anthropic's
dialect: the filename like a document title, the inline bytes through the
image pricer.
---
 litellm/litellm_core_utils/token_counter.py   | 28 ++++++++++++++-
 .../litellm_core_utils/test_token_counter.py  | 35 +++++++++++++++++++
 2 files changed, 62 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/token_counter.py b/litellm/litellm_core_utils/token_counter.py
index 256bee7b348..c350bc5569e 100644
--- a/litellm/litellm_core_utils/token_counter.py
+++ b/litellm/litellm_core_utils/token_counter.py
@@ -693,6 +693,26 @@ def _count_document_tokens(
     )
 
 
+def _count_file_tokens(
+    file_value: object,
+    count_function: TokenCounterFunction,
+    use_default_image_token_count: bool,
+) -> int:
+    """An OpenAI `file` block is the chat-completions spelling of a document, so it prices like one."""
+    if not isinstance(file_value, Mapping):
+        return 0
+    filename: Final = file_value.get("filename")
+    file_data: Final = file_value.get("file_data")
+    name_tokens: Final = count_function(filename) if isinstance(filename, str) and filename else 0
+    if not isinstance(file_data, str) or not file_data:
+        return name_tokens
+    return name_tokens + calculate_img_tokens(
+        data=file_data,
+        mode="auto",
+        use_default_image_token_count=use_default_image_token_count,
+    )
+
+
 def _count_anthropic_content(
     content: Mapping[str, Any],
     count_function: TokenCounterFunction,
@@ -778,6 +798,12 @@ def _count_content_list(
                     use_default_image_token_count,
                     default_token_count,
                 )
+            elif c["type"] == "file":
+                num_tokens += _count_file_tokens(
+                    c.get("file"),
+                    count_function,
+                    use_default_image_token_count,
+                )
             elif c["type"] in ("tool_use", "tool_result"):
                 num_tokens += _count_anthropic_content(
                     c,
@@ -807,7 +833,7 @@ def _count_content_list(
                 raise ValueError(
                     f"Invalid content item type: {content_type}. "
                     f"Expected str or dict with 'type' field "
-                    f"(text, image_url, image, document, tool_use, tool_result, thinking, tool_reference)."
+                    f"(text, image_url, image, document, file, tool_use, tool_result, thinking, tool_reference)."
                 )
         return num_tokens
     except Exception as e:
diff --git a/tests/test_litellm/litellm_core_utils/test_token_counter.py b/tests/test_litellm/litellm_core_utils/test_token_counter.py
index 572b505e94c..4694fa8fbed 100644
--- a/tests/test_litellm/litellm_core_utils/test_token_counter.py
+++ b/tests/test_litellm/litellm_core_utils/test_token_counter.py
@@ -1377,3 +1377,38 @@ def test_anthropic_document_title_and_context_add_their_tokens():
             {"type": "document", "source": source},
         ]
     )
+
+
+def test_openai_file_block_prices_like_the_equivalent_anthropic_document():
+    """An inline `file` is a `document` in the chat-completions dialect, so it must price identically, not raise.
+
+    Before the fix `file` was missing from the content-block match even though `ChatCompletionFileObject`
+    is in the union this counter accepts, so every local count of a Responses `input_file` raised
+    `Invalid content item type: file` and surfaced as a 500 on /v1/responses/input_tokens.
+    """
+    prompt = {"type": "text", "text": "Summarize this file."}
+    inline_file = {
+        "type": "file",
+        "file": {"filename": "report.pdf", "file_data": "data:application/pdf;base64,JVBERi0xLjQK"},
+    }
+    document = {
+        "type": "document",
+        "title": "report.pdf",
+        "source": {"type": "base64", "media_type": "application/pdf", "data": "JVBERi0xLjQK"},
+    }
+
+    assert _count_user_content([prompt, inline_file]) == _count_user_content([prompt, document])
+    assert _count_user_content([prompt, inline_file]) > _count_user_content([prompt])
+
+
+def test_openai_file_block_without_inline_bytes_counts_what_it_carries():
+    """A `file` block naming an uploaded file has no bytes to price, so it adds only the filename's tokens."""
+    prompt = {"type": "text", "text": "Summarize this file."}
+
+    by_id = {"type": "file", "file": {"file_id": "file-abc123"}}
+    assert _count_user_content([prompt, by_id]) == _count_user_content([prompt])
+
+    named = {"type": "file", "file": {"file_id": "file-abc123", "filename": "report.pdf"}}
+    assert _count_user_content([prompt, named]) == _count_user_content(
+        [prompt, {"type": "text", "text": "report.pdf"}]
+    )