LangChain OpenAI API Error? Fix AuthenticationError, ChatOpenAI & Model Names

1. AuthenticationError — correct key but still failing

The error looks like this:

openai.AuthenticationError: Error code: 401 - {'error': {'message': 'Incorrect API key provided'}}

The two most common causes: the environment variable is named wrong, or it is set after the import. The fix:

import os

# ✅ Set BEFORE importing ChatOpenAI — must be OPENAI_API_KEY exactly
os.environ["OPENAI_API_KEY"] = "sk-..."

from langchain_openai import ChatOpenAI

# ✅ Or pass directly to the constructor (also reads env var as fallback)
llm = ChatOpenAI(model="gpt-4o", api_key="sk-...")

# ❌ Wrong variable names that look right but don't work:
# os.environ["OPENAI_KEY"] = "sk-..."         # missing _API_
# os.environ["openai_api_key"] = "sk-..."      # must be uppercase

• Verify the key in the OpenAI dashboard — if it starts with sk-proj- it is a project key and requires the project to have API access enabled.
• In .env files, load with python-dotenv before any imports: from dotenv import load_dotenv; load_dotenv().
• Shell: export OPENAI_API_KEY=sk-... (Linux/macOS) or $env:OPENAI_API_KEY="sk-..." (PowerShell).

2. ChatOpenAI vs OpenAI — wrong class for chat models

Using the OpenAI class with a chat model name causes an InvalidRequestError because it calls the completions endpoint instead of chat/completions:

from langchain_openai import ChatOpenAI, OpenAI

# ❌ OpenAI class — only for legacy text-completion models (deprecated)
llm = OpenAI(model="gpt-4o")  # InvalidRequestError

# ✅ ChatOpenAI — for ALL current OpenAI chat models
llm = ChatOpenAI(model="gpt-4o")           # recommended, best price/performance
llm = ChatOpenAI(model="gpt-4-turbo")      # if you need a specific turbo version
llm = ChatOpenAI(model="gpt-3.5-turbo")   # cheaper, faster

# ✅ With extra params
llm = ChatOpenAI(
    model="gpt-4o",
    temperature=0.7,
    max_tokens=512,
    streaming=True,        # enables token streaming
)

• Rule of thumb: if the model name contains "gpt-3.5-turbo", "gpt-4", "gpt-4o", or "o1" — use ChatOpenAI.
• Old import path from langchain.chat_models import ChatOpenAI is removed in v0.2 — use from langchain_openai import ChatOpenAI.

3. Model name errors — deprecated or misspelled names

OpenAI returns 404 model_not_found for deprecated snapshot names:

# ❌ Deprecated / snapshot names — will 404
ChatOpenAI(model="gpt-4-0314")       # retired snapshot
ChatOpenAI(model="gpt-4-32k")        # no longer available
ChatOpenAI(model="gpt-3.5-turbo-0301")

# ✅ Current aliases (as of 2025) — prefer these
ChatOpenAI(model="gpt-4o")           # points to latest gpt-4o
ChatOpenAI(model="gpt-4o-mini")      # cheaper, fast
ChatOpenAI(model="gpt-4-turbo")      # latest gpt-4-turbo
ChatOpenAI(model="gpt-3.5-turbo")    # legacy but still active
ChatOpenAI(model="o1")               # reasoning model
ChatOpenAI(model="o1-mini")          # cheaper reasoning

Check platform.openai.com/docs/models for the current list — model availability depends on your account tier.

4. chain.run() removed in v0.2 — use chain.invoke()

LangChain v0.2 removed .run() and .__call__() from LCEL chains. The error is AttributeError: 'RunnableSequence' object has no attribute 'run':

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

llm = ChatOpenAI(model="gpt-4o")
prompt = ChatPromptTemplate.from_template("Summarise: {text}")
chain = prompt | llm | StrOutputParser()

# ❌ v0.1 style — removed in v0.2
result = chain.run("My text here")           # AttributeError
result = chain({"text": "My text here"})    # AttributeError

# ✅ v0.2+ style
result = chain.invoke({"text": "My text here"})          # sync
result = await chain.ainvoke({"text": "My text here"})    # async
for chunk in chain.stream({"text": "My text here"}):     # streaming
    print(chunk, end="", flush=True)

• Legacy LLMChain: if you used LLMChain(llm=..., prompt=...), its .run() still works in v0.2 but is deprecated — it emits a warning and will be removed. Migrate to LCEL.
• invoke() input: pass a dict matching your prompt's input variables. A single-variable prompt can also accept a plain string: chain.invoke("hello").

5. "Cannot read properties of undefined" — accessing .content on wrong object

When you call llm.invoke() directly (without a parser), it returns an AIMessage object, not a string:

from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage

llm = ChatOpenAI(model="gpt-4o")

# llm.invoke() returns AIMessage, not str
response = llm.invoke("Tell me a joke")
print(response)            # AIMessage(content='Why did the...', ...)
print(response.content)    # ✅ the text string

# With a prompt list
response = llm.invoke([HumanMessage(content="Tell me a joke")])
print(response.content)    # ✅

# Add StrOutputParser to get a plain string from the chain
from langchain_core.output_parsers import StrOutputParser
chain = llm | StrOutputParser()
text = chain.invoke("Tell me a joke")
print(text)  # ✅ plain str

• Structured output: use llm.with_structured_output(MyPydanticModel) — this returns a Pydantic object, not AIMessage. Do not call .content on its result.
• JavaScript/TypeScript: @langchain/openai returns the same AIMessage structure — access response.content as a string.

6. Version pinning — langchain vs @langchain/openai (JS/TS)

In Python, langchain and langchain-openai must be from the same release generation. In JS/TS, langchain and @langchain/openai are separate packages with independent versioning:

# Python — check all versions together
pip show langchain langchain-core langchain-openai

# Upgrade everything at once to avoid mismatches
pip install --upgrade langchain langchain-core langchain-openai

# TS/JS — install separately
npm install langchain @langchain/openai @langchain/core

# TS usage
import { ChatOpenAI } from "@langchain/openai";
import { ChatPromptTemplate } from "@langchain/core/prompts";
import { StringOutputParser } from "@langchain/core/output_parsers";

const llm = new ChatOpenAI({ model: "gpt-4o", temperature: 0 });
const prompt = ChatPromptTemplate.fromTemplate("Answer: {question}");
const chain = prompt.pipe(llm).pipe(new StringOutputParser());
const result = await chain.invoke({ question: "What is 2+2?" });

• Python: langchain[openai] extras syntax is deprecated — install langchain-openai directly.
• JS/TS peer dep errors: pin @langchain/core to the version that both langchain and @langchain/openai agree on; check with npm ls @langchain/core.

7. LangSmith tracing causing slowdowns

If each chain call takes 2–5 seconds longer than expected, LangSmith tracing is likely adding blocking network calls. Disable or configure it:

# Disable tracing entirely
import os
os.environ["LANGCHAIN_TRACING_V2"] = "false"

# Or unset if you don't need it in production
# del os.environ["LANGCHAIN_TRACING_V2"]

# Enable tracing correctly (all 3 vars required)
os.environ["LANGCHAIN_TRACING_V2"] = "true"
os.environ["LANGCHAIN_API_KEY"] = "ls__..."    # from smith.langchain.com
os.environ["LANGCHAIN_PROJECT"] = "my-project" # optional project name

# Suppress tracing for a single call without changing env vars
result = chain.invoke({"input": "hello"}, config={"callbacks": []})

• Check LangSmith status on Prismix — if LangSmith is degraded, tracing blocks every call until timeout.
• Background tracing: set LANGCHAIN_TRACING_BACKGROUND=true to make tracing non-blocking (experimental in some versions).