Langchain-Openai Defaults To Retired OpenAI Models Causing API Errors

The default model_name fields in langchain_openai/llms/base.py and langchain_openai/chat_models/base.py are hardcoded to retired models. There is no dynamic resolution or validation that prevents usage of deprecated endpoints.

from langchain_openai import ChatOpenAI, OpenAI OpenAI().invoke("Say hi") # Uses retired gpt-3.5-turbo-instruct ChatOpenAI().invoke("Say hi") # Uses soon-to-be-retired gpt-3.5-turbo

This fix introduces breaking changes: users who relied on the old default models must now set model names explicitly. For ChatOpenAI, gpt-5.6-terra is a reasoning model that does not accept temperature or max_tokens (use max_completion_tokens); existing code passing these parameters may break. The OpenAI completions class now raises an error if no model is provided, which may require users to migrate to ChatOpenAI. In multi-threaded environments, ensure model validation is thread-safe (pydantic validators are safe but global default mutation is not recommended). Rollback: revert to previous package version and pin model names in application code to avoid unintended model changes.