From 15d8d5bbffc8b0c4a964ec9ede0951e13d4096ff Mon Sep 17 00:00:00 2001
From: Atishyy27 <sethatishayjain@gmail.com>
Date: Thu, 5 Nov 2026 00:03:15 +0530
Subject: [PATCH] fix(langchain): use the model name reported in the response

ChatOpenAI puts its configured model into llm_output["model_name"], so a
deployment name or LiteLLM alias was recorded instead of the model the
provider returned, and cost came out as zero. Read the model from the
generation's response metadata first and keep llm_output as the fallback.
---
 langfuse/langchain/CallbackHandler.py | 12 ++++++++++
 tests/unit/test_langchain.py          | 32 +++++++++++++++++++++++++++
 2 files changed, 44 insertions(+)

diff --git a/langfuse/langchain/CallbackHandler.py b/langfuse/langchain/CallbackHandler.py
index dac76f4fc..01612cbad 100644
--- a/langfuse/langchain/CallbackHandler.py
+++ b/langfuse/langchain/CallbackHandler.py
@@ -1820,6 +1820,18 @@ def _parse_usage(response: LLMResult) -> Any:
 
 
 def _parse_model(response: LLMResult) -> Any:
+    # llm_output["model_name"] can be the configured alias, so prefer the
+    # model the provider reported on the generation
+    if response.generations and response.generations[-1]:
+        generation = response.generations[-1][-1]
+        message = getattr(generation, "message", None)
+        for metadata in (
+            getattr(message, "response_metadata", None),
+            generation.generation_info,
+        ):
+            if isinstance(metadata, dict) and metadata.get("model_name"):
+                return metadata["model_name"]
+
     # langchain-anthropic uses the usage field
     llm_model_keys = ["model_name"]
     llm_model = None
diff --git a/tests/unit/test_langchain.py b/tests/unit/test_langchain.py
index c55eaa3d2..f2a9e04a8 100644
--- a/tests/unit/test_langchain.py
+++ b/tests/unit/test_langchain.py
@@ -104,6 +104,38 @@ def test_chat_model_callback_exports_generation_span(
     }
 
 
+def test_chat_model_generation_uses_model_reported_in_response(
+    langfuse_memory_client, get_span
+):
+    response = ChatResult(
+        generations=[
+            ChatGeneration(
+                message=AIMessage(
+                    content="bonjour",
+                    response_metadata={"model_name": "gpt-4o-2024-08-06"},
+                ),
+                text="bonjour",
+            )
+        ],
+        llm_output={"model_name": "azure.gpt-4o"},
+    )
+
+    with patch.object(ChatOpenAI, "_generate", return_value=response):
+        handler = CallbackHandler()
+        ChatOpenAI(api_key="test", model="azure.gpt-4o").invoke(
+            [HumanMessage(content="hello")],
+            config={"callbacks": [handler]},
+        )
+
+    langfuse_memory_client.flush()
+    generation_span = get_span("ChatOpenAI")
+
+    assert (
+        generation_span.attributes[LangfuseOtelSpanAttributes.OBSERVATION_MODEL]
+        == "gpt-4o-2024-08-06"
+    )
+
+
 def test_chat_model_input_keeps_message_names(
     langfuse_memory_client, get_span, json_attr
 ):