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.

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
+                    content = chat_content_blocks_to_responses_content(content)
-                    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)
                 input_items.append({"role": "user", "content": content})
             elif role == "assistant":
                 # Map tool_calls to Responses API function_call items

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()

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