From c1f809c1e0bb11c93a64a30cfca67e2ff63ad0d5 Mon Sep 17 00:00:00 2001
From: Anuj7411 <ojhaanuj224@gmail.com>
Date: Wed, 26 Aug 2026 19:29:21 +0530
Subject: [PATCH] fix(cost): bill audio input tokens at base rate when model
 has no audio price

Audio prompt tokens are reported separately from text tokens and subtracted from the text bucket before pricing. When a model ships without input_cost_per_audio_token (37 shipped models including gemini-2.5-pro, gemini-3-pro-preview, and xai grok-4-1-fast), those audio tokens were billed at $0 instead of falling back to input_cost_per_token. Image and video input tokens already fall back to the base rate, and output audio falls back to the base completion rate; input audio was the only modality that silently zeroed.
---
 .../litellm_core_utils/llm_cost_calc/utils.py |  8 ++-
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 53 +++++++++++++++++++
 2 files changed, 60 insertions(+), 1 deletion(-)

diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py
index 0a52e1d283e..6a1301fb0ab 100644
--- a/litellm/litellm_core_utils/llm_cost_calc/utils.py
+++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py
@@ -662,7 +662,13 @@ def _calculate_input_cost(
 
     ### AUDIO COST
     if prompt_tokens_details["audio_tokens"]:
-        audio_cost_key: Final = _get_service_tier_cost_key("input_cost_per_audio_token", service_tier)
+        tier_audio_cost_key: Final = _get_service_tier_cost_key("input_cost_per_audio_token", service_tier)
+        has_audio_price: Final = (
+            model_info.get(tier_audio_cost_key) is not None
+            or model_info.get("input_cost_per_audio_token") is not None
+            or model_info.get("input_cost_per_audio_per_second") is not None
+        )
+        audio_cost_key: Final = tier_audio_cost_key if has_audio_price else "input_cost_per_token"
         prompt_cost += calculate_cost_component(model_info, audio_cost_key, prompt_tokens_details["audio_tokens"])
 
     ### IMAGE TOKEN COST
diff --git a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
index 6f513ce1bd4..1e01602dc5f 100644
--- a/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
+++ b/tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py
@@ -1476,6 +1476,59 @@ def test_string_cost_values_edge_cases():
     assert round(completion_cost, 12) == round(expected_completion_cost, 12)
 
 
+def test_audio_input_tokens_fall_back_to_base_rate_when_no_audio_price():
+    """Audio input tokens must bill at input_cost_per_token when the model has no audio rate.
+
+    Regression: models like gemini-2.5-pro report audio prompt tokens separately from text
+    but ship without input_cost_per_audio_token, which previously billed those tokens at $0.
+    """
+    model_info: ModelInfo = {
+        "input_cost_per_token": 1.25e-6,
+        "output_cost_per_token": 1e-5,
+    }
+
+    usage = Usage(
+        prompt_tokens=9700,
+        completion_tokens=0,
+        total_tokens=9700,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            audio_tokens=9600, cached_tokens=0, text_tokens=100, image_tokens=None
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="gemini-2.5-pro", usage=usage, custom_llm_provider="vertex_ai", model_info=model_info
+    )
+
+    expected_prompt_cost = (100 + 9600) * 1.25e-6
+    assert round(prompt_cost, 12) == round(expected_prompt_cost, 12)
+
+
+def test_audio_input_tokens_use_audio_rate_when_present():
+    """When input_cost_per_audio_token exists it takes precedence over the base rate."""
+    model_info: ModelInfo = {
+        "input_cost_per_token": 1.25e-6,
+        "input_cost_per_audio_token": 5e-6,
+        "output_cost_per_token": 1e-5,
+    }
+
+    usage = Usage(
+        prompt_tokens=9700,
+        completion_tokens=0,
+        total_tokens=9700,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            audio_tokens=9600, cached_tokens=0, text_tokens=100, image_tokens=None
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="gemini-2.5-pro", usage=usage, custom_llm_provider="vertex_ai", model_info=model_info
+    )
+
+    expected_prompt_cost = 100 * 1.25e-6 + 9600 * 5e-6
+    assert round(prompt_cost, 12) == round(expected_prompt_cost, 12)
+
+
 def test_string_cost_values_with_threshold():
     """Test that string cost values work correctly with threshold pricing."""
     from unittest.mock import patch