From 6e108c1cb47042b4ae303e04504337dda318f648 Mon Sep 17 00:00:00 2001
From: ccurme <chester.curme@gmail.com>
Date: Thu, 7 Aug 2025 09:49:34 -0300
Subject: [PATCH] feat(core): zero-out token costs for cache hits (#32437)

---
 .../language_models/chat_models.py            | 10 ++++++++++
 .../language_models/chat_models/test_cache.py | 20 +++++++++++++++++++
 2 files changed, 30 insertions(+)

diff --git a/libs/core/langchain_core/language_models/chat_models.py b/libs/core/langchain_core/language_models/chat_models.py
index c5e7b43cebf..7873e7aaf5d 100644
--- a/libs/core/langchain_core/language_models/chat_models.py
+++ b/libs/core/langchain_core/language_models/chat_models.py
@@ -666,6 +666,16 @@ class BaseChatModel(BaseLanguageModel[BaseMessage], ABC):
                 converted_generations.append(chat_gen)
             else:
                 # Already a ChatGeneration or other expected type
+                if hasattr(gen, "message") and isinstance(gen.message, AIMessage):
+                    # We zero out cost on cache hits
+                    gen.message = gen.message.model_copy(
+                        update={
+                            "usage_metadata": {
+                                **(gen.message.usage_metadata or {}),
+                                "total_cost": 0,
+                            }
+                        }
+                    )
                 converted_generations.append(gen)
         return converted_generations
 
diff --git a/libs/core/tests/unit_tests/language_models/chat_models/test_cache.py b/libs/core/tests/unit_tests/language_models/chat_models/test_cache.py
index f4f033d3d7f..39e4babc782 100644
--- a/libs/core/tests/unit_tests/language_models/chat_models/test_cache.py
+++ b/libs/core/tests/unit_tests/language_models/chat_models/test_cache.py
@@ -458,3 +458,23 @@ def test_cleanup_serialized() -> None:
         "name": "CustomChat",
         "type": "constructor",
     }
+
+
+def test_token_costs_are_zeroed_out() -> None:
+    # We zero-out token costs for cache hits
+    local_cache = InMemoryCache()
+    messages = [
+        AIMessage(
+            content="Hello, how are you?",
+            usage_metadata={"input_tokens": 5, "output_tokens": 10, "total_tokens": 15},
+        ),
+    ]
+    model = GenericFakeChatModel(messages=iter(messages), cache=local_cache)
+    first_response = model.invoke("Hello")
+    assert isinstance(first_response, AIMessage)
+    assert first_response.usage_metadata
+
+    second_response = model.invoke("Hello")
+    assert isinstance(second_response, AIMessage)
+    assert second_response.usage_metadata
+    assert second_response.usage_metadata["total_cost"] == 0  # type: ignore[typeddict-item]