fix: include realtime transcription cost in breakdown

litellm/cost_calculator.py

@@ -2564,7 +2564,16 @@ def handle_realtime_stream_cost_calculation(
         input_cost_per_token += _input_cost_per_token
         output_cost_per_token += _output_cost_per_token
         break  # exit if we find a valid model
-    total_cost = input_cost_per_token + output_cost_per_token
+    transcription_cost = (
+        handle_realtime_transcription_cost_calculation(
+            results=results,
+            custom_llm_provider=custom_llm_provider,
+            litellm_model_name=litellm_model_name,
+        )
+        if any(r.get("type") == _TRANSCRIPTION_COMPLETED_EVENT_TYPE for r in results)
+        else 0.0
+    )
+    total_cost = input_cost_per_token + output_cost_per_token + transcription_cost
 
     _store_cost_breakdown_in_logging_obj(
         litellm_logging_obj=litellm_logging_obj,
@@ -2574,13 +2583,6 @@ def handle_realtime_stream_cost_calculation(
         total_cost_usd_dollar=total_cost,
     )
 
-    if any(r.get("type") == _TRANSCRIPTION_COMPLETED_EVENT_TYPE for r in results):
-        total_cost += handle_realtime_transcription_cost_calculation(
-            results=results,
-            custom_llm_provider=custom_llm_provider,
-            litellm_model_name=litellm_model_name,
-        )
-
     return total_cost
 
 

tests/test_litellm/test_cost_calculator.py

@@ -637,6 +637,10 @@ def test_realtime_transcription_duration_cost(monkeypatch):
     ($0.017/min). The .completed events carry usage {type: duration, seconds: N};
     cost must equal total_seconds * input_cost_per_second.
     """
+    from datetime import datetime
+
+    from litellm.litellm_core_utils.litellm_logging import Logging
+
     monkeypatch.setenv("LITELLM_LOCAL_MODEL_COST_MAP", "True")
     monkeypatch.setattr(litellm, "model_cost", litellm.get_model_cost_map(url=""))
 
@@ -667,17 +671,29 @@ def test_realtime_transcription_duration_cost(monkeypatch):
     combined = RealtimeAPITokenUsageProcessor.collect_and_combine_usage_from_realtime_stream_results(
         results=results
     )
+    logging_obj = Logging(
+        model="gpt-realtime-whisper",
+        messages=[],
+        stream=False,
+        call_type="_arealtime",
+        start_time=datetime.now(),
+        litellm_call_id="realtime-transcription-cost-breakdown-test",
+        function_id="realtime-transcription-cost-breakdown-test",
+    )
     cost = handle_realtime_stream_cost_calculation(
         results=results,
         combined_usage_object=combined,
         custom_llm_provider="openai",
         litellm_model_name="gpt-realtime-whisper",
+        litellm_logging_obj=logging_obj,
     )
 
     # 90 seconds at $0.017/minute.
     expected = 90.0 * (0.017 / 60)
     assert abs(cost - expected) < 1e-9
     assert cost > 0  # guards against the duration branch being dropped
+    assert logging_obj.cost_breakdown is not None
+    assert abs(logging_obj.cost_breakdown["total_cost"] - cost) < 1e-9
 
 
 def test_realtime_transcription_duration_cost_resolves_model_from_litellm_name(