From e77b59a59ec5678c8a4029bf183217ac5f41894c Mon Sep 17 00:00:00 2001
From: yangpengtao <34332614+totoyang@users.noreply.github.com>
Date: Wed, 28 Oct 2026 19:28:11 +0800
Subject: [PATCH] fix(token_counter): count assistant refusal content parts

---
 litellm/litellm_core_utils/token_counter.py   | 12 +++++++--
 .../litellm_core_utils/test_token_counter.py  | 26 ++++++++++++++++++-
 2 files changed, 35 insertions(+), 3 deletions(-)

diff --git a/litellm/litellm_core_utils/token_counter.py b/litellm/litellm_core_utils/token_counter.py
index df8038e756c..8b7d2bf76a2 100644
--- a/litellm/litellm_core_utils/token_counter.py
+++ b/litellm/litellm_core_utils/token_counter.py
@@ -17,6 +17,7 @@ import anyio
 import anyio.lowlevel
 import httpx
 import tiktoken
+from openai.types.chat import ChatCompletionContentPartRefusalParam
 from pydantic import TypeAdapter
 from tiktoken.model import MODEL_TO_ENCODING
 from typing_extensions import ParamSpec, TypeVar
@@ -1691,7 +1692,13 @@ def _count_anthropic_content(
 
 
 _TEXT_FIELD_BY_BLOCK_TYPE: Final = MappingProxyType(
-    {"text": "text", "thinking": "thinking", "redacted_thinking": "thinking", "compaction": "content"}
+    {
+        "text": "text",
+        "refusal": "refusal",
+        "thinking": "thinking",
+        "redacted_thinking": "thinking",
+        "compaction": "content",
+    }
 )
 
 
@@ -1700,6 +1707,7 @@ def _count_content_list(
     content_list: str
     | Iterable[
         OpenAIMessageContentListBlock
+        | ChatCompletionContentPartRefusalParam
         | ChatCompletionThinkingBlock
         | ChatCompletionRedactedThinkingBlock
         | AnthropicMessagesTextParam
@@ -1770,7 +1778,7 @@ def _count_content_list(
                 raise UnsupportedContentTypeError(
                     f"Invalid content item type: {content_type}. "
                     f"Expected str or dict with 'type' field "
-                    f"(text, image_url, image, document, file, tool_use, tool_result, thinking, redacted_thinking, "
+                    f"(text, refusal, image_url, image, document, file, tool_use, tool_result, thinking, redacted_thinking, "
                     f"compaction, tool_reference, search_result)."
                 )
         return num_tokens
diff --git a/tests/unit/litellm_core_utils/test_token_counter.py b/tests/unit/litellm_core_utils/test_token_counter.py
index 8ef118154d1..4f1230bcc80 100644
--- a/tests/unit/litellm_core_utils/test_token_counter.py
+++ b/tests/unit/litellm_core_utils/test_token_counter.py
@@ -21,6 +21,7 @@ from unittest.mock import MagicMock, sentinel
 import anyio.to_thread
 import pytest
 import tiktoken
+from openai.types.chat import ChatCompletionContentPartRefusalParam
 from PIL import Image
 from tokenizers import Regex, Tokenizer, models, pre_tokenizers
 
@@ -37,6 +38,7 @@ from litellm.constants import (
 )
 from litellm.litellm_core_utils.asyncify import asyncify
 from litellm.litellm_core_utils.token_counter import (
+    _count_content_list,
     _encoding_count,
     _get_exact_count_function,
     _get_extrapolating_count_function,
@@ -49,7 +51,7 @@ from litellm.litellm_core_utils.token_counter import (
 )
 from litellm.litellm_core_utils.token_counter import token_counter as token_counter_new
 from litellm.llms.anthropic.messages.adapters.transformation import LiteLLMAnthropicMessagesAdapter
-from litellm.types.llms.openai import AllMessageValues, ChatCompletionToolParam
+from litellm.types.llms.openai import AllMessageValues, ChatCompletionTextObject, ChatCompletionToolParam
 from litellm.types.utils import SelectTokenizerResponse
 from tests.large_text import text
 from tests.unit.litellm_core_utils.event_loop_lag import (
@@ -93,6 +95,28 @@ def test_token_counter_basic():
     )
 
 
+@pytest.mark.parametrize("refusal_text", ["I cannot help with that request.", "", "抱歉，我无法帮助处理该请求。"])
+@pytest.mark.parametrize("default_token_count", [None, 0])
+def test_content_list_counts_refusal_text(refusal_text: str, default_token_count: int | None) -> None:
+    encoding: Final = tiktoken.get_encoding("cl100k_base")
+    count_function: Final[Callable[[str], int]] = lambda text: _encoding_count(encoding, text)
+    content: Final[list[ChatCompletionTextObject | ChatCompletionContentPartRefusalParam]] = [
+        {"type": "text", "text": "Safety notice."},
+        {"type": "refusal", "refusal": refusal_text},
+    ]
+    expected: Final = count_function("Safety notice.") + count_function(refusal_text)
+
+    assert (
+        _count_content_list(
+            count_function=count_function,
+            content_list=content,
+            use_default_image_token_count=False,
+            default_token_count=default_token_count,
+        )
+        == expected
+    )
+
+
 def test_token_counter_large_repeated_text_is_encoded_in_bounded_chunks():
     text_length: Final = 1024 * 1024
     messages: Final = [{"role": "user", "content": [{"type": "text", "text": "A" * text_length}]}]