From 20545eda8076f9a8936ef24f4caa0aa4228c30ae Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <devin-ai-integration[bot]@users.noreply.github.com>
Date: Sat, 8 Aug 2026 19:09:56 +0000
Subject: [PATCH] fix(anthropic): use provider-reported thinking_tokens for
 reasoning_tokens

---
 litellm/llms/anthropic/chat/transformation.py | 22 +++++-
 .../test_anthropic_chat_transformation.py     | 79 +++++++++++++++++++
 2 files changed, 100 insertions(+), 1 deletion(-)

diff --git a/litellm/llms/anthropic/chat/transformation.py b/litellm/llms/anthropic/chat/transformation.py
index 1161c92232a..7d1d646da55 100644
--- a/litellm/llms/anthropic/chat/transformation.py
+++ b/litellm/llms/anthropic/chat/transformation.py
@@ -2117,6 +2117,22 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
             return False
         return any(key in usage_object for key in ("cache_read_input_tokens", "cache_creation_input_tokens"))
 
+    @staticmethod
+    def _get_reported_thinking_tokens(usage_object: dict) -> int | None:
+        """Anthropic reports the billed extended-thinking count in ``usage.output_tokens_details.thinking_tokens``.
+
+        It is authoritative: under the default ``display: "omitted"`` the thinking blocks carry no text at all,
+        so estimating from the visible reasoning content always yields 0, and under ``display: "summarized"``
+        the visible summary understates what was billed.
+        """
+        details: Final = usage_object.get("output_tokens_details")
+        if not isinstance(details, dict):
+            return None
+        thinking_tokens: Final = details.get("thinking_tokens")
+        if isinstance(thinking_tokens, bool) or not isinstance(thinking_tokens, (int, float)):
+            return None
+        return int(thinking_tokens)
+
     def calculate_usage(
         self,
         usage_object: dict,
@@ -2199,7 +2215,11 @@ class AnthropicConfig(AnthropicModelInfo, BaseConfig):
         estimated_reasoning_tokens: Final = (
             token_counter(text=reasoning_content, count_response_tokens=True) if reasoning_content else 0
         )
-        reasoning_tokens: Final = min(estimated_reasoning_tokens, completion_tokens)
+        reported_reasoning_tokens: Final = self._get_reported_thinking_tokens(_usage)
+        reasoning_tokens: Final = min(
+            estimated_reasoning_tokens if reported_reasoning_tokens is None else reported_reasoning_tokens,
+            completion_tokens,
+        )
         completion_token_details: Final = CompletionTokensDetailsWrapper(
             reasoning_tokens=max(0, reasoning_tokens),
             text_tokens=(completion_tokens - reasoning_tokens if reasoning_tokens > 0 else completion_tokens),
diff --git a/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py b/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py
index 231d3b48754..6fc61da326a 100644
--- a/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py
+++ b/tests/test_litellm/llms/anthropic/chat/test_anthropic_chat_transformation.py
@@ -2436,6 +2436,85 @@ def test_calculate_usage_completion_tokens_details_with_reasoning():
     assert usage.completion_tokens == 500
 
 
+def test_calculate_usage_prefers_reported_thinking_tokens():
+    """
+    Anthropic reports the billed thinking count in usage.output_tokens_details.thinking_tokens.
+    Under the default display ("omitted") the thinking blocks are empty, so estimating from the
+    visible reasoning text yields 0 even though the model thought (and billed).
+
+    Fixes: https://github.com/BerriAI/litellm/issues/36290
+    """
+    config = AnthropicConfig()
+
+    usage_object = {
+        "input_tokens": 29,
+        "output_tokens": 391,
+        "output_tokens_details": {"thinking_tokens": 180},
+    }
+
+    usage = config.calculate_usage(usage_object=usage_object, reasoning_content="")
+
+    assert usage.completion_tokens_details is not None
+    assert usage.completion_tokens_details.reasoning_tokens == 180
+    assert usage.completion_tokens_details.text_tokens == 211
+
+
+def test_calculate_usage_reported_thinking_tokens_override_text_estimate():
+    """Summarized thinking text understates the billed count, so the reported value wins."""
+    config = AnthropicConfig()
+
+    usage_object = {
+        "input_tokens": 29,
+        "output_tokens": 391,
+        "output_tokens_details": {"thinking_tokens": 180},
+    }
+
+    usage = config.calculate_usage(usage_object=usage_object, reasoning_content="short summary")
+
+    assert usage.completion_tokens_details is not None
+    assert usage.completion_tokens_details.reasoning_tokens == 180
+
+
+@pytest.mark.parametrize(
+    "output_tokens_details",
+    [None, {}, {"thinking_tokens": None}],
+)
+def test_calculate_usage_falls_back_to_estimate_without_reported_thinking_tokens(output_tokens_details):
+    """Older responses have no thinking_tokens, so the visible-text estimate must still apply."""
+    config = AnthropicConfig()
+
+    usage_object = {
+        "input_tokens": 100,
+        "output_tokens": 500,
+        "output_tokens_details": output_tokens_details,
+    }
+
+    usage = config.calculate_usage(
+        usage_object=usage_object,
+        reasoning_content="Let me think about this step by step. " * 10,
+    )
+
+    assert usage.completion_tokens_details is not None
+    assert usage.completion_tokens_details.reasoning_tokens is not None
+    assert usage.completion_tokens_details.reasoning_tokens > 0
+
+
+def test_calculate_usage_clamps_reported_thinking_tokens_to_output_tokens():
+    config = AnthropicConfig()
+
+    usage_object = {
+        "input_tokens": 10,
+        "output_tokens": 50,
+        "output_tokens_details": {"thinking_tokens": 900},
+    }
+
+    usage = config.calculate_usage(usage_object=usage_object, reasoning_content=None)
+
+    assert usage.completion_tokens_details is not None
+    assert usage.completion_tokens_details.reasoning_tokens == 50
+    assert usage.completion_tokens_details.text_tokens == 0
+
+
 # ============ Reasoning Effort Tests ============