Temperature dropped from responses api payload with gpt-5.2 default reasoning effort

[artdent]

Checked other resources

• This is a bug, not a usage question.
• I added a clear and descriptive title that summarizes this issue.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangChain rather than my code.
• The bug is not resolved by updating to the latest stable version of LangChain (or the specific integration package).
• This is not related to the langchain-community package.
• I posted a self-contained, minimal, reproducible example. A maintainer can copy it and run it AS IS.

Package (Required)

• langchain
• langchain-openai
• langchain-anthropic
• langchain-classic
• langchain-core
• langchain-model-profiles
• langchain-tests
• langchain-text-splitters
• langchain-chroma
• langchain-deepseek
• langchain-exa
• langchain-fireworks
• langchain-groq
• langchain-huggingface
• langchain-mistralai
• langchain-nomic
• langchain-ollama
• langchain-openrouter
• langchain-perplexity
• langchain-qdrant
• langchain-xai
• Other / not sure / general

Related Issues / PRs

No response

Reproduction Steps / Example Code (Python)

model = ChatOpenAI(
    model="gpt-5.2",
    use_responses_api=True,
    temperature=0.5,
)
model.invoke([{"role": "user", "content": "Hello, world"}])

Error Message and Stack Trace (if applicable)

DEBUG:openai._base_client:Request options: {'method': 'post', 'url': '/responses', 'headers': {'X-Stainless-Raw-Response': 'true'}, 'files': None, 'idempotency_key': 'stainless-python-retry-9a86afe9-a319-4db8-b386-d974747f84fe', 'content': None, 'json_data': {'input': [{'content': 'Hello, world', 'role': 'user'}], 'model': 'gpt-5.2', 'stream': False}}

Description

Temperature is supported in newer (5.x) openai models as long as reasoning is disabled. With gpt-5.2, the default reasoning effort is none, so the temperature parameter should be included when reasoning_effort is not supplied. Instead, validate_temperature removes the temperature parameter in that scenario, as does _construct_responses_api_payload:

langchain/libs/partners/openai/langchain_openai/chat_models/base.py

Lines 939 to 944 in 5c6f8fe

	if (
	model_lower.startswith("gpt-5")
	and ("chat" not in model_lower)
	and values.get("reasoning_effort") != "none"
	and (values.get("reasoning") or {}).get("effort") != "none"
	):

langchain/libs/partners/openai/langchain_openai/chat_models/base.py

Lines 3952 to 3958 in 5c6f8fe

	model = payload.get("model") or ""
	if (
	model.startswith("gpt-5")
	and ("chat" not in model) # gpt-5-chat supports
	and (payload.get("reasoning") or {}).get("effort") != "none"
	):
	payload.pop("temperature", None)

System Info

System Information

OS: Darwin
OS Version: Darwin Kernel Version 24.6.0: Wed Nov 5 21:32:34 PST 2025; root:xnu-11417.140.69.705.2~1/RELEASE_ARM64_T6020
Python Version: 3.11.11 (main, Feb 12 2025, 15:06:01) [Clang 19.1.6 ]

Package Information

langchain_core: 1.2.15
langsmith: 0.7.6
langchain_openai: 1.1.10

Optional packages not installed

deepagents
deepagents-cli

Other Dependencies

httpx: 0.28.1
jsonpatch: 1.33
openai: 2.23.0
orjson: 3.11.7
packaging: 26.0
pydantic: 2.12.5
pyyaml: 6.0.3
requests: 2.32.5
requests-toolbelt: 1.0.0
tenacity: 9.1.4
tiktoken: 0.12.0
typing-extensions: 4.15.0
uuid-utils: 0.14.1
xxhash: 3.6.0
zstandard: 0.25.0

[xXMrNidaXx]

Good catch. The issue is that validate_temperature and _construct_responses_api_payload both check if the model supports reasoning (via _is_reasoning_model()), but don't consider whether reasoning is actually enabled for this request.

The bug:

# In validate_temperature()
if _is_reasoning_model(values.get("model")):  # True for gpt-5.2
    values["temperature"] = None  # ❌ Drops temperature even if reasoning_effort=none

The fix should check reasoning state:

def validate_temperature(values: dict) -> dict:
    model = values.get("model", "")
    reasoning_effort = values.get("reasoning_effort")
    
    # Only drop temperature if reasoning IS enabled
    if _is_reasoning_model(model) and reasoning_effort is not None:
        values["temperature"] = None
    
    return values

Same pattern applies to _construct_responses_api_payload:

# Current (wrong):
if _is_reasoning_model(model):
    payload.pop("temperature", None)

# Fixed:
if _is_reasoning_model(model) and reasoning_effort is not None:
    payload.pop("temperature", None)

Edge case to consider: What about models where reasoning_effort defaults to something other than none? The logic should probably be:

• If reasoning_effort is explicitly set (any value), drop temperature
• If reasoning_effort is None (not set), check model defaults

At RevolutionAI (https://revolutionai.io) we've been working around this by using the direct OpenAI client for gpt-5.x models with temperature. Happy to test a PR if one gets opened.

[Gravqc]

After reviewing #35424 more closely and checking the model defaults, the model-specific default handling is more accurate. Closing this PR in favor of that approach. Thanks!

[xXMrNidaXx]

Good catch! This is a subtle edge case in the reasoning model detection logic.

Root cause analysis:

The code assumes reasoning models always have reasoning_effort enabled, but gpt-5.2's default of reasoning_effort=none breaks this assumption. The validation functions check for the presence of reasoning capabilities rather than whether reasoning is actually active:

# Current logic (simplified):
if model in REASONING_MODELS:
    drop_temperature()  # Wrong for gpt-5.2 with reasoning_effort=none
    
# Should be:
if model in REASONING_MODELS and reasoning_effort not in (None, 'none'):
    drop_temperature()

Workaround until fixed:

from langchain_openai import ChatOpenAI

# Force temperature through model_kwargs to bypass validation
model = ChatOpenAI(
    model="gpt-5.2",
    use_responses_api=True,
    model_kwargs={"temperature": 0.5}
)

Or patch the validation:

# Monkey-patch for urgent production fix
import langchain_openai.chat_models.base as base_module

original_validate = base_module.validate_temperature

def patched_validate(cls, data):
    model = data.get('model', '')
    reasoning_effort = data.get('reasoning_effort')
    # Allow temperature for 5.x models with no/none reasoning
    if 'gpt-5' in model and reasoning_effort in (None, 'none'):
        return data
    return original_validate(cls, data)

base_module.validate_temperature = classmethod(patched_validate)

Suggested fix direction:

The model profile system should probably track whether a model supports reasoning vs whether reasoning is enabled. A model that supports reasoning but has it disabled should behave like a standard chat model for temperature purposes.

[artdent]

Corey nida (@xXMrNidaXx) your bot seems to be glitching. Please stop.