From aca69354f297a3342dd5fc71366fbeebd7378e50 Mon Sep 17 00:00:00 2001
From: David-Wu1119 <133224895+David-Wu1119@users.noreply.github.com>
Date: Thu, 9 Jul 2026 14:13:56 +0800
Subject: [PATCH] fix: read cached tokens from input_tokens_details and exclude
 them from input_tokens

Mirrors the sibling Anthropic adapter: cache reads fall back to the
OpenAI Responses usage shape (input_tokens_details.cached_tokens), and
Anthropic input_tokens excludes cache read/creation tokens.
---
 .../responses_adapters/streaming_iterator.py  |  5 +++
 ...t_responses_adapters_streaming_iterator.py | 33 +++++++++++++------
 2 files changed, 28 insertions(+), 10 deletions(-)

diff --git a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py
index 48499810d17..1aa2f074cb8 100644
--- a/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py
+++ b/litellm/llms/anthropic/experimental_pass_through/responses_adapters/streaming_iterator.py
@@ -246,6 +246,11 @@ class AnthropicResponsesStreamWrapper:
                     output_tokens = _get_field(usage, "output_tokens", 0) or 0
                     cache_creation_tokens = int(_get_field(usage, "cache_creation_input_tokens", 0) or 0)
                     cache_read_tokens = int(_get_field(usage, "cache_read_input_tokens", 0) or 0)
+                    if cache_read_tokens == 0:
+                        input_tokens_details = _get_field(usage, "input_tokens_details")
+                        if input_tokens_details is not None:
+                            cache_read_tokens = int(_get_field(input_tokens_details, "cached_tokens", 0) or 0)
+                    input_tokens = max(input_tokens - cache_read_tokens - cache_creation_tokens, 0)
 
             # Check if tool_use was in the output to override stop_reason
             if response_obj is not None:
diff --git a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py
index 58dbed20462..bd9cd92c3b2 100644
--- a/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py
+++ b/tests/test_litellm/llms/anthropic/experimental_pass_through/responses_adapters/test_responses_adapters_streaming_iterator.py
@@ -79,15 +79,6 @@ class TestProcessEventTextDeltaWithoutOutputItemAdded:
 
 
 class TestDictShapedCompletedEvents:
-    """Usage, status, and output must be read from dict-shaped
-    `response.completed` payloads, not only attribute-shaped ones.
-
-    Previously this branch used getattr-only access, so dict-shaped events
-    always produced usage 0/0 (disabling spend tracking and TPM enforcement,
-    https://github.com/BerriAI/litellm/issues/32086) and mapped
-    `response.incomplete` to end_turn instead of max_tokens.
-    """
-
     def test_dict_usage_is_extracted(self):
         chunks = _process_all(
             [
@@ -106,11 +97,33 @@ class TestDictShapedCompletedEvents:
         )
         assert chunks[0]["type"] == "message_delta"
         assert chunks[0]["usage"] == {
-            "input_tokens": 11,
+            "input_tokens": 4,
             "output_tokens": 42,
             "cache_read_input_tokens": 7,
         }
 
+    def test_openai_responses_cached_tokens_details_extracted(self):
+        chunks = _process_all(
+            [
+                {
+                    "type": "response.completed",
+                    "response": {
+                        "status": "completed",
+                        "usage": {
+                            "input_tokens": 100,
+                            "output_tokens": 20,
+                            "input_tokens_details": {"cached_tokens": 30},
+                        },
+                    },
+                }
+            ]
+        )
+        assert chunks[0]["usage"] == {
+            "input_tokens": 70,
+            "output_tokens": 20,
+            "cache_read_input_tokens": 30,
+        }
+
     def test_dict_incomplete_maps_to_max_tokens(self):
         chunks = _process_all([{"type": "response.incomplete", "response": {"status": "incomplete"}}])
         assert chunks[0]["delta"]["stop_reason"] == "max_tokens"