Unable to use GPT-5 with Langchain's create_react_agent

[genaiconference]

[Bug Report] GPT-5 "Unsupported parameter: 'stop'" Error — Needs Official LangChain Fix

Summary

When using **LangChain’s react agent using create_react_agent, AgentExecutor with GPT-5 (via AzureChatOpenAI), executions consistently fail with the following error:

BadRequestError: Error code: 400 - {'error': {'message': "Unsupported parameter: 'stop' is not supported with this model." ...}}

Using the below versions

langchain==0.3.27
langchain-openai==0.3.32

Details

• I do not explicitly set "stop" in my agent, model, or executor configuration.
• Despite this, the "stop" parameter is somehow included in the POST request to the GPT-5 endpoint.

Attempted Fixes

• Removed "stop" from _default_params and model_kwargs after LLM initialization.
• Tried to patch methods to filter "stop" (not possible due to Pydantic restrictions).
• Verified that no legacy params are set elsewhere.

➡️ No success — "stop" keeps returning in the payload and causing the error.

Impact

This prevents using GPT-5 models with LangChain agents, even on the latest versions (langchain-core, langchain-openai, and Azure deployments).

Request

Could LangChain developers address the following?

1. Filter/remove unsupported parameters like "stop" automatically for GPT-5 and future models.
2. Provide clearer guidance or an official workaround for users hitting this issue.

⚠️ This is a blocker for integrating GPT-5 with LangChain agents and impacts all downstream async/react setups.

[Herc-Ch]

If you are using the create_react_agent, it forces a stop argument itself (due to the construction of the react prompt so that the llm doesnt hallucinate after the first observation). You can check it in debug in agent -> steps -> RunnableBinding( kwargs={'stop': ['\nObservation']}).
Until this is fixed, either change agent-style (tool-agent) or use an llm that accepts like gtp-4 etc.