Skip to content

The agent engineering platform.

fix(xai): recompute total_tokens after adding reasoning tokens to output (#39667)

Aryan Singh K.1e0dcf77ec15
diff --git a/libs/partners/xai/langchain_xai/chat_models.py b/libs/partners/xai/langchain_xai/chat_models.py
index 38a40680ce..c207cd935c 100644
--- a/libs/partners/xai/langchain_xai/chat_models.py
+++ b/libs/partners/xai/langchain_xai/chat_models.py
@@ -652,6 +652,11 @@ class ChatXAI(BaseChatOpenAI): # type: ignore[override]
rtn.generations[0].message.usage_metadata["output_tokens"] += ( # type: ignore[attr-defined]
reasoning_tokens
)
+ # Keep the documented UsageMetadata invariant
+ # total_tokens == input_tokens + output_tokens (gh #39634).
+ usage_metadata["total_tokens"] = (
+ usage_metadata["input_tokens"] + usage_metadata["output_tokens"]
+ )
return rtn
@@ -700,6 +705,11 @@ class ChatXAI(BaseChatOpenAI): # type: ignore[override]
)
):
generation_chunk.message.usage_metadata["output_tokens"] += reasoning_tokens # type: ignore[attr-defined]
+ # Keep the documented UsageMetadata invariant
+ # total_tokens == input_tokens + output_tokens (gh #39634).
+ usage_metadata["total_tokens"] = (
+ usage_metadata["input_tokens"] + usage_metadata["output_tokens"]
+ )
return generation_chunk
def with_structured_output(
diff --git a/libs/partners/xai/tests/unit_tests/test_chat_models.py b/libs/partners/xai/tests/unit_tests/test_chat_models.py
index 817b01ee5f..a9943484ae 100644
--- a/libs/partners/xai/tests/unit_tests/test_chat_models.py
+++ b/libs/partners/xai/tests/unit_tests/test_chat_models.py
@@ -3,6 +3,7 @@ import json
import pytest # type: ignore[import-not-found]
from langchain_core.messages import (
AIMessage,
+ AIMessageChunk,
FunctionMessage,
HumanMessage,
SystemMessage,
@@ -12,6 +13,13 @@ from langchain_openai.chat_models.base import (
_convert_dict_to_message,
_convert_message_to_dict,
)
+from openai.types.chat import ChatCompletion
+from openai.types.chat.chat_completion import Choice
+from openai.types.chat.chat_completion_message import ChatCompletionMessage
+from openai.types.completion_usage import (
+ CompletionTokensDetails,
+ CompletionUsage,
+)
from pydantic import SecretStr
from langchain_xai import ChatXAI
@@ -278,3 +286,85 @@ def test_metadata_versions() -> None:
assert "langchain-core" in versions
assert "langchain-xai" in versions
assert "langchain-openai" in versions
+
+
+def test_create_chat_result_recomputes_total_tokens_for_reasoning() -> None:
+ """Adding reasoning tokens to output_tokens must keep total_tokens consistent.
+
+ xAI reports reasoning tokens separately from completion tokens, so ChatXAI
+ adds them into output_tokens. total_tokens must be recomputed afterwards to
+ preserve the UsageMetadata invariant total_tokens == input + output
+ (gh #39634).
+ """
+ llm = ChatXAI(model=MODEL_NAME)
+ response = ChatCompletion(
+ id="chatcmpl-1",
+ object="chat.completion",
+ created=0,
+ model=MODEL_NAME,
+ choices=[
+ Choice(
+ index=0,
+ finish_reason="stop",
+ message=ChatCompletionMessage(
+ role="assistant",
+ content="Test response",
+ ),
+ )
+ ],
+ usage=CompletionUsage(
+ prompt_tokens=32,
+ completion_tokens=9,
+ total_tokens=41,
+ completion_tokens_details=CompletionTokensDetails(reasoning_tokens=5),
+ ),
+ )
+
+ result = llm._create_chat_result(response)
+
+ message = result.generations[0].message
+ assert isinstance(message, AIMessage)
+ usage_metadata = message.usage_metadata
+ assert usage_metadata is not None
+ assert usage_metadata["input_tokens"] == 32
+ assert usage_metadata["output_tokens"] == 14 # 9 completion + 5 reasoning
+ assert usage_metadata["total_tokens"] == 46 # 32 + 14, invariant holds
+ assert usage_metadata["output_token_details"]["reasoning"] == 5
+
+
+def test_convert_chunk_recomputes_total_tokens_for_reasoning() -> None:
+ """Streaming chunks must keep the total_tokens invariant as well (gh #39634)."""
+ llm = ChatXAI(model=MODEL_NAME)
+ chunk = {
+ "id": "chatcmpl-1",
+ "object": "chat.completion.chunk",
+ "created": 0,
+ "model": MODEL_NAME,
+ "choices": [
+ {
+ "index": 0,
+ "delta": {"role": "assistant", "content": "Test"},
+ "finish_reason": None,
+ }
+ ],
+ "usage": {
+ "prompt_tokens": 32,
+ "completion_tokens": 9,
+ "total_tokens": 41,
+ "completion_tokens_details": {"reasoning_tokens": 5},
+ },
+ }
+
+ generation_chunk = llm._convert_chunk_to_generation_chunk(
+ chunk, AIMessageChunk, None
+ )
+ assert generation_chunk is not None
+
+ message = generation_chunk.message
+ assert isinstance(message, AIMessageChunk)
+ usage_metadata = message.usage_metadata
+ assert usage_metadata is not None
+ assert usage_metadata["input_tokens"] == 32
+ assert usage_metadata["output_tokens"] == 14 # 9 completion + 5 reasoning
+ assert usage_metadata["total_tokens"] == 46 # 32 + 14, invariant holds
+ assert usage_metadata["output_token_details"]["reasoning"] == 5