[Fix] UI - Usage: Cached tokens always showing zero for OpenAI models

OpenAI stores cached tokens under prompt_tokens_details.cached_tokens,
but the daily spend writer only checked the top-level
cache_read_input_tokens field (used by Anthropic/DeepSeek). Added
fallback to read from prompt_tokens_details when the top-level field
is absent.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

litellm/proxy/db/db_spend_update_writer.py

@@ -1852,6 +1852,38 @@ class DBSpendUpdateWriter:
             unique_constraint_name="tag_date_api_key_model_custom_llm_provider_mcp_namespaced_tool_name_endpoint",
         )
 
+    @staticmethod
+    def _get_cache_read_input_tokens(usage_obj: dict) -> int:
+        """
+        Extract cache read input tokens from a usage object dict.
+
+        Anthropic/DeepSeek store this as a top-level `cache_read_input_tokens` field.
+        OpenAI stores it nested under `prompt_tokens_details.cached_tokens`.
+        """
+        value = usage_obj.get("cache_read_input_tokens", 0) or 0
+        if value:
+            return value
+        prompt_details = usage_obj.get("prompt_tokens_details") or {}
+        if isinstance(prompt_details, dict):
+            return prompt_details.get("cached_tokens", 0) or 0
+        return 0
+
+    @staticmethod
+    def _get_cache_creation_input_tokens(usage_obj: dict) -> int:
+        """
+        Extract cache creation input tokens from a usage object dict.
+
+        Anthropic stores this as a top-level `cache_creation_input_tokens` field.
+        It may also appear under `prompt_tokens_details.cache_creation_tokens`.
+        """
+        value = usage_obj.get("cache_creation_input_tokens", 0) or 0
+        if value:
+            return value
+        prompt_details = usage_obj.get("prompt_tokens_details") or {}
+        if isinstance(prompt_details, dict):
+            return prompt_details.get("cache_creation_tokens", 0) or 0
+        return 0
+
     async def _common_add_spend_log_transaction_to_daily_transaction(
         self,
         payload: Union[dict, SpendLogsPayload],
@@ -1932,12 +1964,12 @@ class DBSpendUpdateWriter:
                 api_requests=1,
                 successful_requests=1 if request_status == "success" else 0,
                 failed_requests=1 if request_status != "success" else 0,
-                cache_read_input_tokens=usage_obj.get("cache_read_input_tokens", 0)
-                or 0,
-                cache_creation_input_tokens=usage_obj.get(
-                    "cache_creation_input_tokens", 0
-                )
-                or 0,
+                cache_read_input_tokens=self._get_cache_read_input_tokens(
+                    usage_obj
+                ),
+                cache_creation_input_tokens=self._get_cache_creation_input_tokens(
+                    usage_obj
+                ),
             )
             return daily_transaction
         except Exception as e:

tests/test_litellm/proxy/db/test_db_spend_update_writer.py

@@ -1428,3 +1428,52 @@ async def test_commit_spend_updates_uses_pipeline():
     mock_redis_update_buffer.get_all_daily_end_user_spend_update_transactions_from_redis_buffer.assert_not_called()
     mock_redis_update_buffer.get_all_daily_agent_spend_update_transactions_from_redis_buffer.assert_not_called()
     mock_redis_update_buffer.get_all_daily_tag_spend_update_transactions_from_redis_buffer.assert_not_called()
+
+
+class TestGetCacheTokenHelpers:
+    """Tests for _get_cache_read_input_tokens and _get_cache_creation_input_tokens.
+
+    These helpers extract cache token counts from usage objects. Anthropic/DeepSeek
+    store them as top-level fields, while OpenAI nests them under prompt_tokens_details.
+    """
+
+    def test_cache_read_from_top_level_field(self):
+        """Anthropic-style: cache_read_input_tokens at top level."""
+        usage = {"cache_read_input_tokens": 500}
+        assert DBSpendUpdateWriter._get_cache_read_input_tokens(usage) == 500
+
+    def test_cache_read_from_prompt_tokens_details(self):
+        """OpenAI-style: cached_tokens nested in prompt_tokens_details."""
+        usage = {"prompt_tokens_details": {"cached_tokens": 300}}
+        assert DBSpendUpdateWriter._get_cache_read_input_tokens(usage) == 300
+
+    def test_cache_read_top_level_takes_precedence(self):
+        """When both are present, top-level field wins."""
+        usage = {
+            "cache_read_input_tokens": 500,
+            "prompt_tokens_details": {"cached_tokens": 300},
+        }
+        assert DBSpendUpdateWriter._get_cache_read_input_tokens(usage) == 500
+
+    def test_cache_read_empty_usage(self):
+        assert DBSpendUpdateWriter._get_cache_read_input_tokens({}) == 0
+
+    def test_cache_read_none_values(self):
+        usage = {
+            "cache_read_input_tokens": None,
+            "prompt_tokens_details": None,
+        }
+        assert DBSpendUpdateWriter._get_cache_read_input_tokens(usage) == 0
+
+    def test_cache_creation_from_top_level_field(self):
+        """Anthropic-style: cache_creation_input_tokens at top level."""
+        usage = {"cache_creation_input_tokens": 200}
+        assert DBSpendUpdateWriter._get_cache_creation_input_tokens(usage) == 200
+
+    def test_cache_creation_from_prompt_tokens_details(self):
+        """cache_creation_tokens nested in prompt_tokens_details."""
+        usage = {"prompt_tokens_details": {"cache_creation_tokens": 150}}
+        assert DBSpendUpdateWriter._get_cache_creation_input_tokens(usage) == 150
+
+    def test_cache_creation_empty_usage(self):
+        assert DBSpendUpdateWriter._get_cache_creation_input_tokens({}) == 0