Merge pull request #13742 from BerriAI/litellm_dev_08_18_2025_p2

Fix - gemini prompt caching cost calculation

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -305,9 +305,9 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
                 return None
 
         for tool in value:
-            openai_function_object: Optional[
-                ChatCompletionToolParamFunctionChunk
-            ] = None
+            openai_function_object: Optional[ChatCompletionToolParamFunctionChunk] = (
+                None
+            )
             if "function" in tool:  # tools list
                 _openai_function_object = ChatCompletionToolParamFunctionChunk(  # type: ignore
                     **tool["function"]
@@ -597,14 +597,14 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
             elif param == "seed":
                 optional_params["seed"] = value
             elif param == "reasoning_effort" and isinstance(value, str):
-                optional_params[
-                    "thinkingConfig"
-                ] = VertexGeminiConfig._map_reasoning_effort_to_thinking_budget(value)
+                optional_params["thinkingConfig"] = (
+                    VertexGeminiConfig._map_reasoning_effort_to_thinking_budget(value)
+                )
             elif param == "thinking":
-                optional_params[
-                    "thinkingConfig"
-                ] = VertexGeminiConfig._map_thinking_param(
-                    cast(AnthropicThinkingParam, value)
+                optional_params["thinkingConfig"] = (
+                    VertexGeminiConfig._map_thinking_param(
+                        cast(AnthropicThinkingParam, value)
+                    )
                 )
             elif param == "modalities" and isinstance(value, list):
                 response_modalities = self.map_response_modalities(value)
@@ -1000,6 +1000,7 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
             GenerateContentResponseBody, BidiGenerateContentServerMessage
         ],
     ) -> Usage:
+
         if (
             completion_response is not None
             and "usageMetadata" not in completion_response
@@ -1038,6 +1039,16 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
                     text_tokens = detail.get("tokenCount", 0)
         if "thoughtsTokenCount" in usage_metadata:
             reasoning_tokens = usage_metadata["thoughtsTokenCount"]
+
+        ## adjust 'text_tokens' to subtract cached tokens
+        if (
+            (audio_tokens is None or audio_tokens == 0)
+            and text_tokens is not None
+            and text_tokens > 0
+            and cached_tokens is not None
+        ):
+            text_tokens = text_tokens - cached_tokens
+
         prompt_tokens_details = PromptTokensDetailsWrapper(
             cached_tokens=cached_tokens,
             audio_tokens=audio_tokens,
@@ -1344,28 +1355,28 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
             ## ADD METADATA TO RESPONSE ##
 
             setattr(model_response, "vertex_ai_grounding_metadata", grounding_metadata)
-            model_response._hidden_params[
-                "vertex_ai_grounding_metadata"
-            ] = grounding_metadata
+            model_response._hidden_params["vertex_ai_grounding_metadata"] = (
+                grounding_metadata
+            )
 
             setattr(
                 model_response, "vertex_ai_url_context_metadata", url_context_metadata
             )
 
-            model_response._hidden_params[
-                "vertex_ai_url_context_metadata"
-            ] = url_context_metadata
+            model_response._hidden_params["vertex_ai_url_context_metadata"] = (
+                url_context_metadata
+            )
 
             setattr(model_response, "vertex_ai_safety_results", safety_ratings)
-            model_response._hidden_params[
-                "vertex_ai_safety_results"
-            ] = safety_ratings  # older approach - maintaining to prevent regressions
+            model_response._hidden_params["vertex_ai_safety_results"] = (
+                safety_ratings  # older approach - maintaining to prevent regressions
+            )
 
             ## ADD CITATION METADATA ##
             setattr(model_response, "vertex_ai_citation_metadata", citation_metadata)
-            model_response._hidden_params[
-                "vertex_ai_citation_metadata"
-            ] = citation_metadata  # older approach - maintaining to prevent regressions
+            model_response._hidden_params["vertex_ai_citation_metadata"] = (
+                citation_metadata  # older approach - maintaining to prevent regressions
+            )
 
         except Exception as e:
             raise VertexAIError(

tests/test_litellm/test_cost_calculator.py

@@ -480,3 +480,69 @@ def test_gemini_25_implicit_caching_cost():
     ), f"Expected cost {expected_cost}, but got {result}"
 
     print(f"✓ Gemini 2.5 implicit caching cost calculation is correct: ${result:.8f}")
+
+
+def test_gemini_25_explicit_caching_cost_direct_usage():
+    """
+    Test that Gemini 2.5 models correctly calculate costs with explicit caching.
+
+    This test reproduces the issue from #11156 where cached tokens should receive
+    a 75% discount.
+    """
+    from litellm.litellm_core_utils.llm_cost_calc.utils import generic_cost_per_token
+    from litellm.types.utils import (
+        CompletionTokensDetailsWrapper,
+        PromptTokensDetailsWrapper,
+        Usage,
+    )
+    from litellm.utils import get_model_info
+
+    model_info = get_model_info(model="gemini-2.5-pro", custom_llm_provider="gemini")
+
+    usage = Usage(
+        completion_tokens=2522,
+        prompt_tokens=42001,
+        total_tokens=44523,
+        completion_tokens_details=CompletionTokensDetailsWrapper(
+            accepted_prediction_tokens=None,
+            audio_tokens=None,
+            reasoning_tokens=1908,
+            rejected_prediction_tokens=None,
+            text_tokens=614,
+        ),
+        prompt_tokens_details=PromptTokensDetailsWrapper(
+            audio_tokens=None, cached_tokens=40938, text_tokens=1063, image_tokens=None
+        ),
+    )
+
+    input_cost, output_cost = generic_cost_per_token(
+        model="gemini/gemini-2.5-pro",
+        usage=usage,
+        custom_llm_provider="gemini",
+    )
+
+    total_cost = input_cost + output_cost
+
+    expected_higher_than_actual_cost = (
+        model_info["input_cost_per_token"] * usage.prompt_tokens
+        + model_info["output_cost_per_token"] * usage.completion_tokens
+    )
+
+    print(f"expected_higher_than_actual_cost: {expected_higher_than_actual_cost}")
+
+    assert expected_higher_than_actual_cost > total_cost
+
+    expected_actual_cost = (
+        model_info["input_cost_per_token"] * usage.prompt_tokens_details.text_tokens
+        + model_info["cache_read_input_token_cost"]
+        * usage.prompt_tokens_details.cached_tokens
+        + model_info["output_cost_per_token"] * usage.completion_tokens
+    )
+
+    print(
+        f"model_info['input_cost_per_token']: {model_info['input_cost_per_token']}, usage.prompt_tokens_details.text_tokens: {usage.prompt_tokens_details.text_tokens}, model_info['cache_read_input_token_cost']: {model_info['cache_read_input_token_cost']}, model_info['output_cost_per_token']: {model_info['output_cost_per_token']}"
+    )
+
+    print(f"Expected actual cost: {expected_actual_cost}")
+
+    assert expected_actual_cost == total_cost