fix(cost): bill per-second transcription models outside chat modes (#44458)

* fix(cost): bill per-second transcription models outside chat modes

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(cost): restore request-time billing for per-second transcription models

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

* fix(cost): bill audio length for per-second transcription models when known

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

---------

Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/cost_calculator.py

@@ -326,7 +326,7 @@ class OCRPricing(TypedDict, total=False):
     annotation_cost_per_page: ReadOnly[float | None]
 
 
-_WALL_CLOCK_PRICED_MODES: Final = frozenset({"chat", "completion", "embedding", "responses"})
+_WALL_CLOCK_PRICED_MODES: Final = frozenset({"audio_transcription", "chat", "completion", "embedding", "responses"})
 
 
 def _has_token_or_tiered_pricing(model_info: ModelInfoBase) -> bool:
@@ -346,6 +346,7 @@ def _per_second_pricing_cost(
     model: str,
     custom_llm_provider: str | None,
     response_time_ms: float | None,
+    audio_seconds: float = 0.0,
 ) -> tuple[float, float] | None:
     try:
         model_info: Final = _cached_get_model_info_helper(model=model, custom_llm_provider=custom_llm_provider)
@@ -366,7 +367,11 @@ def _per_second_pricing_cost(
     if resolved_cost_per_second is None:
         return None
 
-    seconds: Final = (response_time_ms or 0.0) / 1000
+    seconds: Final = (
+        audio_seconds
+        if audio_seconds > 0 and model_info.get("mode") == "audio_transcription"
+        else (response_time_ms or 0.0) / 1000
+    )
     verbose_logger.debug(
         "For model=%s - cost_per_second: %s; response time: %s",
         model,
@@ -667,6 +672,7 @@ def cost_per_token(
             model=model,
             custom_llm_provider=custom_llm_provider,
             response_time_ms=response_time_ms,
+            audio_seconds=audio_transcription_file_duration,
         )
     ) is not None:
         return per_second_cost

tests/integration/pricing/test_per_second_pricing.py

@@ -4,6 +4,7 @@ from collections.abc import Mapping
 from typing import Final
 
 import httpx
+import litellm
 import pytest
 from pydantic import JsonValue
 
@@ -197,3 +198,41 @@ def test_streaming_chat_per_second_pricing_covers_the_full_stream(
             f"spend={spend}, total frame delay={total_frame_delay_seconds}s, body={body}"
         )
         assert not PRICING_FIELDS.intersection(body), body
+
+
+@pytest.mark.parametrize(
+    ("model", "custom_llm_provider", "expected_cost"),
+    (
+        ("deepgram/nova-3", "deepgram", (0.0007167, 0.0)),
+        ("mistral/voxtral-mini-transcribe-realtime-latest", "mistral", (0.001, 0.0)),
+    ),
+    ids=("deepgram_nova_3", "mistral_voxtral_realtime"),
+)
+def test_transcription_per_second_model_bills_request_time_through_cost_per_token(
+    model: str, custom_llm_provider: str, expected_cost: tuple[float, float]
+) -> None:
+    cost: Final = litellm.cost_per_token(model=model, custom_llm_provider=custom_llm_provider, response_time_ms=10_000.0)
+
+    assert cost == pytest.approx(expected_cost)
+
+
+def test_transcription_per_second_model_bills_audio_length_through_cost_per_token() -> None:
+    cost: Final = litellm.cost_per_token(
+        model="deepgram/nova-3",
+        custom_llm_provider="deepgram",
+        response_time_ms=10_000.0,
+        audio_transcription_file_duration=60.0,
+    )
+
+    assert cost == pytest.approx((0.0043002, 0.0))
+
+
+def test_transcription_response_still_bills_audio_length_through_completion_cost() -> None:
+    response: Final = litellm.TranscriptionResponse(text="hello")
+    response.duration = 60.0
+
+    cost: Final = litellm.completion_cost(
+        completion_response=response, model="deepgram/nova-3", custom_llm_provider="deepgram", total_time=10.0
+    )
+
+    assert cost == pytest.approx(0.0043002)

tests/unit/test_cost_calculator.py

@@ -3299,7 +3299,7 @@ def test_completion_cost_per_second_deployment_bills_the_call_duration(
     assert cost == pytest.approx(0.02 * expected_seconds)
 
 
-@pytest.mark.parametrize("mode", ["audio_transcription", "audio_speech", "video_generation", "realtime"])
+@pytest.mark.parametrize("mode", ["audio_speech", "video_generation", "realtime"])
 def test_cost_per_token_leaves_media_second_rates_to_their_dedicated_paths(monkeypatch, mode: str):
     """
     A media-mode entry's per-second rates price audio or video seconds, which the dedicated
@@ -3316,6 +3316,29 @@ def test_cost_per_token_leaves_media_second_rates_to_their_dedicated_paths(monke
     assert cost_per_token(model=model, custom_llm_provider="openai", response_time_ms=2000.0) == (0.0, 0.0)
 
 
+@pytest.mark.parametrize(
+    ("audio_seconds", "expected_cost"),
+    [(0.0, (0.04, 0.0)), (60.0, (1.2, 0.0))],
+    ids=["no_audio_length_bills_request_time", "audio_length_bills_audio_seconds"],
+)
+def test_cost_per_token_bills_transcription_second_rates(
+    monkeypatch: pytest.MonkeyPatch, audio_seconds: float, expected_cost: tuple[float, float]
+) -> None:
+    model: Final = "test-transcription-per-second"
+    monkeypatch.setitem(
+        litellm.model_cost,
+        model,
+        {"input_cost_per_second": 0.02, "litellm_provider": "deepgram", "mode": "audio_transcription"},
+    )
+
+    assert cost_per_token(
+        model=model,
+        custom_llm_provider="deepgram",
+        response_time_ms=2000.0,
+        audio_transcription_file_duration=audio_seconds,
+    ) == pytest.approx(expected_cost)
+
+
 def test_completion_cost_video_status_poll_bills_nothing_on_a_per_second_video_model(monkeypatch):
     """
     Polling a video job returns a ``VideoObject`` with no stamped duration, so the cost path falls