From 5737cab258b405689a55e1e6dcaec385673fa572 Mon Sep 17 00:00:00 2001
From: shivam <shivam@berri.ai>
Date: Thu, 10 Sep 2026 22:27:58 +0000
Subject: [PATCH] fix(cost): cap nested cached modality counts at cached_tokens

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 litellm/cost_calculator.py                    |  2 --
 .../litellm_core_utils/llm_cost_calc/utils.py | 16 +++++++++++----
 .../llm_cost_calc/test_llm_cost_calc_utils.py | 20 +++++++++++++++++++
 3 files changed, 32 insertions(+), 6 deletions(-)

diff --git a/litellm/cost_calculator.py b/litellm/cost_calculator.py
index 8daa2de416b..b865318f3af 100644
--- a/litellm/cost_calculator.py
+++ b/litellm/cost_calculator.py
@@ -2332,8 +2332,6 @@ def _combine_prompt_tokens_details(combined: Usage, usage: Usage) -> None:
     if not hasattr(combined, "prompt_tokens_details") or not combined.prompt_tokens_details:
         combined.prompt_tokens_details = PromptTokensDetailsWrapper()
 
-    # Check what keys exist in the model's prompt_tokens_details
-    # Access model_fields on the class, not the instance, to avoid Pydantic 2.11+ deprecation warnings
     for attr in _summable_prompt_token_fields(usage.prompt_tokens_details):
         if (
             hasattr(usage.prompt_tokens_details, attr)
diff --git a/litellm/litellm_core_utils/llm_cost_calc/utils.py b/litellm/litellm_core_utils/llm_cost_calc/utils.py
index 18ef99597a0..dc689ca9618 100644
--- a/litellm/litellm_core_utils/llm_cost_calc/utils.py
+++ b/litellm/litellm_core_utils/llm_cost_calc/utils.py
@@ -806,9 +806,17 @@ def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
         or None
     )
     cached_tokens_details: Final = getattr(usage.prompt_tokens_details, "cached_tokens_details", None)
-    cached_text_tokens: Final = _get_token_detail_value(cached_tokens_details, "text_tokens") or 0
-    cached_audio_tokens: Final = _get_token_detail_value(cached_tokens_details, "audio_tokens") or 0
-    cached_image_tokens: Final = _get_token_detail_value(cached_tokens_details, "image_tokens") or 0
+    cached_audio_tokens: Final = min(
+        _get_token_detail_value(cached_tokens_details, "audio_tokens") or 0, cache_hit_tokens
+    )
+    cached_text_tokens: Final = min(
+        _get_token_detail_value(cached_tokens_details, "text_tokens") or 0,
+        cache_hit_tokens - cached_audio_tokens,
+    )
+    cached_image_tokens: Final = min(
+        _get_token_detail_value(cached_tokens_details, "image_tokens") or 0,
+        cache_hit_tokens - cached_audio_tokens - cached_text_tokens,
+    )
     text_tokens: Final = max(
         (
             cast(int | None, getattr(usage.prompt_tokens_details, "text_tokens", None))
@@ -852,7 +860,7 @@ def parse_prompt_tokens_details(usage: Usage) -> PromptTokensDetailsResult:
 
     return PromptTokensDetailsResult(
         cache_hit_tokens=cache_hit_tokens,
-        cache_hit_audio_tokens=min(cached_audio_tokens, cache_hit_tokens),
+        cache_hit_audio_tokens=cached_audio_tokens,
         cache_creation_tokens=cache_creation_tokens,
         cache_creation_token_details=cache_creation_token_details,
         text_tokens=text_tokens,
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 3709b526c3b..33825c8dd01 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
@@ -5215,3 +5215,23 @@ def test_cached_audio_tokens_fall_back_to_cache_read_input_token_cost() -> None:
     )
     expected = 52 * 4e-6 + 64 * 5e-7 + 39 * 32e-6 + 128 * 5e-7
     assert prompt_cost == pytest.approx(expected)
+
+
+def test_cached_audio_tokens_capped_at_cached_tokens(_local_model_cost_map: None) -> None:
+    """Nested cached_tokens_details exceeding cached_tokens must not over-subtract the audio bucket."""
+    usage = Usage(
+        prompt_tokens=283,
+        completion_tokens=0,
+        total_tokens=283,
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            text_tokens=116,
+            audio_tokens=167,
+            cached_tokens=100,
+            cached_tokens_details={"audio_tokens": 128},
+        ),
+    )
+
+    prompt_cost, _ = generic_cost_per_token(
+        model="gpt-realtime-2", usage=usage, custom_llm_provider="openai"
+    )
+    assert prompt_cost == pytest.approx(116 * 4e-6 + (167 - 100) * 32e-6 + 100 * 4e-7)