fix(token-counter): count reasoning summaries in message content

litellm/litellm_core_utils/token_counter.py

@@ -46,6 +46,7 @@ from litellm.types.llms.openai import (
     AllMessageValues,
     ChatCompletionDocumentObject,
     ChatCompletionNamedToolChoiceParam,
+    ChatCompletionReasoningItem,
     ChatCompletionToolParam,
     OpenAIMessageContentListBlock,
 )
@@ -846,6 +847,7 @@ def _count_content_list(
         | AnthropicMessagesTextParam
         | AnthropicMessagesImageParam
         | AnthropicMessagesDocumentParam
+        | ChatCompletionReasoningItem
     ],
     use_default_image_token_count: bool,
     default_token_count: int | None,
@@ -893,6 +895,10 @@ def _count_content_list(
                 thinking_text = str(c.get("thinking", ""))
                 if thinking_text:
                     num_tokens += count_function(thinking_text)
+            elif c["type"] == "reasoning":
+                num_tokens += sum(
+                    count_function(text) for summary in c.get("summary", ()) if (text := summary.get("text"))
+                )
             elif c["type"] == "tool_reference":
                 # Anthropic tool-search reference block: a lightweight pointer to
                 # a deferred tool, e.g. {"type": "tool_reference", "tool_name": ...}.
@@ -909,7 +915,7 @@ 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, file, tool_use, tool_result, thinking, tool_reference)."
+                    f"(text, image_url, image, document, file, tool_use, tool_result, thinking, reasoning, tool_reference)."
                 )
         return num_tokens
     except Exception as e:

tests/test_litellm/litellm_core_utils/test_token_counter.py

@@ -7,6 +7,7 @@ import threading
 import time
 import traceback
 from concurrent.futures import Future, wait
+from copy import deepcopy
 from typing import Final
 from unittest.mock import MagicMock
 
@@ -31,6 +32,7 @@ from litellm.litellm_core_utils.token_counter import (
     offload_token_count,
 )
 from litellm.litellm_core_utils.token_counter import token_counter as token_counter_new
+from litellm.types.llms.openai import ChatCompletionReasoningItem
 from tests.large_text import text
 from tests.test_litellm.litellm_core_utils.event_loop_lag import (
     assert_loop_stayed_free,
@@ -1250,6 +1252,77 @@ def test_token_counter_with_thinking_content():
     ), f"Expected minimal token count for empty thinking block, got {tokens_no_thinking}"
 
 
+@pytest.mark.parametrize("model", ["gemini/gemini-3.8-flash", "gpt-6-astra"])
+@pytest.mark.parametrize(
+    ("reasoning", "summary_texts"),
+    [
+        ({"type": "reasoning"}, ()),
+        ({"type": "reasoning", "summary": []}, ()),
+        ({"type": "reasoning", "encrypted_content": "opaque " * 1000}, ()),
+        (
+            {"type": "reasoning", "summary": [{"type": "summary_text", "text": "A short synthetic summary."}]},
+            ("A short synthetic summary.",),
+        ),
+        (
+            {
+                "type": "reasoning",
+                "summary": [
+                    {"type": "summary_text", "text": "First thought."},
+                    {"type": "summary_text", "text": ""},
+                    {"type": "summary_text", "text": "Second thought."},
+                ],
+                "id": "rs_opaque_identifier",
+                "encrypted_content": "opaque " * 1000,
+            },
+            ("First thought.", "", "Second thought."),
+        ),
+    ],
+    ids=["missing-summary", "empty-summary", "encrypted-only", "summary", "multiple-summaries-with-metadata"],
+)
+def test_token_counter_with_reasoning_content(
+    model: str, reasoning: ChatCompletionReasoningItem, summary_texts: tuple[str, ...]
+) -> None:
+    original: Final = deepcopy(reasoning)
+    messages: Final = [
+        {"role": "user", "content": "Context marker."},
+        {"role": "assistant", "content": [reasoning, {"type": "text", "text": "Ready."}]},
+    ]
+    equivalent_messages: Final = [
+        {"role": "user", "content": "Context marker."},
+        {
+            "role": "assistant",
+            "content": [
+                *({"type": "text", "text": value} for value in summary_texts),
+                {"type": "text", "text": "Ready."},
+            ],
+        },
+    ]
+
+    assert token_counter(model=model, messages=messages) == token_counter(model=model, messages=equivalent_messages)
+    assert reasoning == original
+
+
+def test_reasoning_content_preserves_prompt_cache_eligibility() -> None:
+    model: Final = "gemini/gemini-3.8-flash"
+    messages: Final = [
+        {"role": "user", "content": "Context marker. " * 1800},
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "reasoning", "summary": [{"type": "summary_text", "text": "A short synthetic summary."}]},
+                {"type": "text", "text": "Ready."},
+            ],
+        },
+        {"role": "user", "content": "Reply with exactly OK."},
+    ]
+    count: Final = token_counter(model=model, messages=messages)
+
+    assert count >= litellm.utils.get_prompt_cache_min_tokens(model)
+    assert litellm.utils.is_prompt_caching_valid_prompt(model=model, messages=messages)
+    assert litellm.utils.is_prompt_caching_valid_prompt(model=model, messages=messages, min_token_count=count)
+    assert not litellm.utils.is_prompt_caching_valid_prompt(model=model, messages=messages, min_token_count=count + 1)
+
+
 def test_token_counter_with_tool_reference_block():
     """
     Regression test: a message containing an Anthropic tool-search