From 29fe342eadd76a9e7c40a274c15d22606d95096e Mon Sep 17 00:00:00 2001
From: Ahmed N <34286755+hMED22@users.noreply.github.com>
Date: Fri, 14 Aug 2026 23:12:37 +0100
Subject: [PATCH] fix(transcription): stop a zero output rate from zeroing
 transcription cost (#36914)

cost_per_second treated a declared-but-zero output_cost_per_second as a real
rate, so the output branch claimed the call and the elif locked out
input_cost_per_second. Every transcription model shipping
output_cost_per_second 0.0 next to a real input rate billed $0, which covers
43 of the 55 per-second entries in the cost map: all 36 deepgram models, both
assemblyai, both elevenlabs scribe, both groq whisper and azure-stt. Custom
deployments pairing the two fields the same way billed $0 as well

Take the output branch only when that rate is actually billable, so a zero
falls through to the input rate. Entries that duplicate one rate into both
fields, whisper-1 among them, keep billing exactly what they bill today
---
 litellm/llms/openai/cost_calculation.py       |  7 +-
 .../llms/openai/test_cost_calculation.py      | 83 +++++++++++++++++++
 2 files changed, 87 insertions(+), 3 deletions(-)
 create mode 100644 tests/test_litellm/llms/openai/test_cost_calculation.py

diff --git a/litellm/llms/openai/cost_calculation.py b/litellm/llms/openai/cost_calculation.py
index eafabdb880d..0352d246c09 100644
--- a/litellm/llms/openai/cost_calculation.py
+++ b/litellm/llms/openai/cost_calculation.py
@@ -109,15 +109,16 @@ def cost_per_second(model: str, custom_llm_provider: str | None, duration: float
     prompt_cost = 0.0
     completion_cost = 0.0
     ## Speech / Audio cost calculation
-    if "output_cost_per_second" in model_info and model_info["output_cost_per_second"] is not None:
+    output_cost_per_second: Final = model_info.get("output_cost_per_second")
+    if output_cost_per_second is not None and output_cost_per_second > 0:
         verbose_logger.debug(
             "For model=%s - output_cost_per_second: %s; duration: %s",
             model,
-            model_info.get("output_cost_per_second"),
+            output_cost_per_second,
             duration,
         )
         ## COST PER SECOND ##
-        completion_cost = model_info["output_cost_per_second"] * duration
+        completion_cost = output_cost_per_second * duration
     elif "input_cost_per_second" in model_info and model_info["input_cost_per_second"] is not None:
         verbose_logger.debug(
             "For model=%s - input_cost_per_second: %s; duration: %s",
diff --git a/tests/test_litellm/llms/openai/test_cost_calculation.py b/tests/test_litellm/llms/openai/test_cost_calculation.py
new file mode 100644
index 00000000000..9b6aec1966c
--- /dev/null
+++ b/tests/test_litellm/llms/openai/test_cost_calculation.py
@@ -0,0 +1,83 @@
+"""Tests for per-second transcription cost calculation."""
+
+import pytest
+
+import litellm
+from litellm.llms.openai.cost_calculation import cost_per_second
+
+
+def _register_stt(name: str, **pricing: float) -> None:
+    litellm.register_model(
+        {
+            name: {
+                "mode": "audio_transcription",
+                "litellm_provider": "openai",
+                **pricing,
+            }
+        },
+        persist_across_reloads=False,
+    )
+
+
+def test_input_rate_bills_when_output_rate_is_zero():
+    """A declared-but-zero output rate must not suppress the real input rate."""
+    _register_stt(
+        "test-stt-zero-output",
+        input_cost_per_second=5e-05,
+        output_cost_per_second=0.0,
+    )
+
+    prompt_cost, completion_cost = cost_per_second(
+        model="test-stt-zero-output", custom_llm_provider="openai", duration=300.0
+    )
+
+    assert prompt_cost == pytest.approx(0.015)
+    assert completion_cost == 0.0
+
+
+def test_output_rate_takes_precedence_when_both_are_billable():
+    """Entries duplicating one rate into both fields must not be billed twice."""
+    _register_stt(
+        "test-stt-both-rates",
+        input_cost_per_second=1e-04,
+        output_cost_per_second=1e-04,
+    )
+
+    prompt_cost, completion_cost = cost_per_second(
+        model="test-stt-both-rates", custom_llm_provider="openai", duration=10.0
+    )
+
+    assert prompt_cost + completion_cost == pytest.approx(1e-03)
+
+
+def test_output_rate_alone_still_bills():
+    _register_stt("test-stt-output-only", output_cost_per_second=3e-05)
+
+    prompt_cost, completion_cost = cost_per_second(
+        model="test-stt-output-only", custom_llm_provider="openai", duration=60.0
+    )
+
+    assert prompt_cost == 0.0
+    assert completion_cost == pytest.approx(1.8e-03)
+
+
+@pytest.mark.parametrize(
+    "model, provider",
+    [
+        ("deepgram/nova-3", "deepgram"),
+        ("groq/whisper-large-v3", "groq"),
+        ("elevenlabs/scribe_v1", "elevenlabs"),
+        ("assemblyai/best", "assemblyai"),
+        ("whisper-1", "openai"),
+    ],
+)
+def test_shipped_per_second_models_bill_a_non_zero_cost(model, provider):
+    prompt_cost, completion_cost = cost_per_second(model=model, custom_llm_provider=provider, duration=60.0)
+
+    assert prompt_cost + completion_cost > 0.0
+
+
+def test_whisper_bills_its_documented_rate_once():
+    prompt_cost, completion_cost = cost_per_second(model="whisper-1", custom_llm_provider="openai", duration=30.0)
+
+    assert prompt_cost + completion_cost == pytest.approx(0.003)