From ca5bcfc8a8cf5cf5d406557a506fe62458835060 Mon Sep 17 00:00:00 2001
From: shivam <shivam@berri.ai>
Date: Thu, 16 Jul 2026 00:48:16 +0000
Subject: [PATCH] fix(cost): price Gemini audio/video prompt tokens at
 tier-aware text rate

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../litellm_core_utils/llm_cost_calc/utils.py | 12 ++-
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 81 +++++++++++++++++++
 2 files changed, 90 insertions(+), 3 deletions(-)

diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py
index 33bf546c239..2a5322416de 100644
--- a/litellm/litellm_core_utils/llm_cost_calc/utils.py
+++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py
@@ -581,7 +581,10 @@ def _calculate_input_cost(
     ### AUDIO COST
     if prompt_tokens_details["audio_tokens"]:
         audio_cost_key = _get_service_tier_cost_key("input_cost_per_audio_token", service_tier)
-        prompt_cost += calculate_cost_component(model_info, audio_cost_key, prompt_tokens_details["audio_tokens"])
+        if model_info.get(audio_cost_key) is None:
+            prompt_cost += float(prompt_tokens_details["audio_tokens"]) * prompt_base_cost
+        else:
+            prompt_cost += calculate_cost_component(model_info, audio_cost_key, prompt_tokens_details["audio_tokens"])
 
     ### IMAGE TOKEN COST
     if prompt_tokens_details["image_tokens"]:
@@ -596,8 +599,11 @@ def _calculate_input_cost(
     if prompt_tokens_details["video_tokens"]:
         video_token_cost_key = "input_cost_per_video_token"
         if model_info.get(video_token_cost_key) is None:
-            video_token_cost_key = "input_cost_per_token"
-        prompt_cost += calculate_cost_component(model_info, video_token_cost_key, prompt_tokens_details["video_tokens"])
+            prompt_cost += float(prompt_tokens_details["video_tokens"]) * prompt_base_cost
+        else:
+            prompt_cost += calculate_cost_component(
+                model_info, video_token_cost_key, prompt_tokens_details["video_tokens"]
+            )
 
     ### CACHE WRITING COST - Now uses tiered pricing
     if (
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 b156faf3ea6..b84cda71ea2 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
@@ -337,6 +337,87 @@ def test_video_input_tokens_gemini_omni_flash_preview():
     )
 
 
+def test_audio_input_tokens_gemini_priced_at_text_rate():
+    """Regression for LIT-4474: audio prompt tokens must fall back to the text input rate.
+
+    gemini-3-pro-preview declares supports_audio_input but has no input_cost_per_audio_token,
+    so audio tokens were previously billed at $0.
+    """
+    model = "gemini-3-pro-preview"
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    model_cost_map = litellm.model_cost[f"gemini/{model}"]
+    assert model_cost_map.get("input_cost_per_audio_token") is None
+
+    text_tokens = 1000
+    audio_tokens = 5000
+    prompt_tokens = text_tokens + audio_tokens
+    usage = Usage(
+        completion_tokens=0,
+        prompt_tokens=prompt_tokens,
+        total_tokens=prompt_tokens,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=text_tokens,
+            audio_tokens=audio_tokens,
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model=model,
+        usage=usage,
+        custom_llm_provider="gemini",
+    )
+
+    assert round(prompt_cost, 10) == round(
+        model_cost_map["input_cost_per_token"] * prompt_tokens,
+        10,
+    )
+    assert round(prompt_cost, 10) != round(
+        model_cost_map["input_cost_per_token"] * text_tokens,
+        10,
+    )
+
+
+def test_audio_video_input_tokens_gemini_use_above_200k_tier():
+    """Regression for LIT-4474: audio/video fallbacks must honor the >200k long-context tier.
+
+    Falling back to the raw un-tiered input_cost_per_token undercounts when the request crosses
+    the 200k boundary; audio and video should be priced at input_cost_per_token_above_200k_tokens
+    like text.
+    """
+    model = "gemini-3-pro-preview"
+    os.environ["LITELLM_LOCAL_MODEL_COST_MAP"] = "True"
+    litellm.model_cost = litellm.get_model_cost_map(url="")
+
+    model_cost_map = litellm.model_cost[f"gemini/{model}"]
+    hi_rate = model_cost_map["input_cost_per_token_above_200k_tokens"]
+    assert hi_rate != model_cost_map["input_cost_per_token"]
+
+    text_tokens = 100000
+    audio_tokens = 60000
+    video_tokens = 80000
+    prompt_tokens = text_tokens + audio_tokens + video_tokens
+    usage = Usage(
+        completion_tokens=0,
+        prompt_tokens=prompt_tokens,
+        total_tokens=prompt_tokens,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=text_tokens,
+            audio_tokens=audio_tokens,
+            video_tokens=video_tokens,
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model=model,
+        usage=usage,
+        custom_llm_provider="gemini",
+    )
+
+    assert round(prompt_cost, 10) == round(hi_rate * prompt_tokens, 10)
+
+
 def test_video_tokens_fallback_to_base_cost():
     """Video output tokens fall back to the base output rate when output_cost_per_video_token is not set."""
     from unittest.mock import patch