fix(logging): fix turn_off_message_logging redaction for standard logging payload

Two bugs in the redaction path for standard_logging_object["response"]:

1. get_final_response_obj was passing a plain dict (from model_dump()) into
   redact_message_input_output_from_logging. perform_redaction only handles
   ModelResponse/EmbeddingResponse/ResponsesAPIResponse -- a plain dict hits
   the else branch and returns {"text": "redacted-by-litellm"}, destroying
   the choices structure. Fix: pass init_response_obj (the original BaseModel)
   to redaction so isinstance checks work correctly, then call model_dump()
   on the result.

2. _redact_choice_content was not nulling choice.message.audio, so audio
   model responses (gpt-4o-audio-preview) kept the full audio blob in the
   logged payload even with turn_off_message_logging=True. Fix: set
   choice.message.audio = None alongside content redaction.

Fixes:
  test_standard_logging_payload[True-ft:gpt-3.5-turbo:my-org:custom_suffix:id]
  test_standard_logging_payload_audio[True-False]
  test_standard_logging_payload_audio[True-True]

litellm/litellm_core_utils/litellm_logging.py

@@ -4827,9 +4827,16 @@ class StandardLoggingPayloadSetup:
         else:
             final_response_obj = {}
 
+        # Pass the original BaseModel object to redaction so isinstance checks inside
+        # perform_redaction work correctly (a plain dict never matches ModelResponse etc.)
+        redaction_input = (
+            init_response_obj
+            if isinstance(init_response_obj, BaseModel)
+            else final_response_obj
+        )
         modified_final_response_obj = redact_message_input_output_from_logging(
             model_call_details=kwargs,
-            result=final_response_obj,
+            result=redaction_input,
         )
 
         if modified_final_response_obj is not None and isinstance(

litellm/litellm_core_utils/redact_messages.py

@@ -49,6 +49,8 @@ def _redact_choice_content(choice):
             choice.message.reasoning_content = "redacted-by-litellm"
         if hasattr(choice.message, "thinking_blocks"):
             choice.message.thinking_blocks = None
+        if hasattr(choice.message, "audio"):
+            choice.message.audio = None
     elif isinstance(choice, litellm.utils.StreamingChoices):
         choice.delta.content = "redacted-by-litellm"
         if hasattr(choice.delta, "reasoning_content"):