module 'openai' has no attribute 'ChatCompletion', tested

openai.ChatCompletion belongs to the old SDK, before 1.0. Code written for it fails on any version since, but not always with the same error. I ran the same old call on five versions:

So the AttributeError is the rarer one. It came from two places in my tests:

• A version older than 0.27. ChatCompletion didn’t exist yet.
• A file of your own named openai.py. With one next to my script, import openai loaded that file instead of the package, and every version gave the same AttributeError. Python 3.14 printed no hint about the name clash.

If you have an openai.py, rename it. Otherwise, check your version with pip show openai.

Rewrite the call for openai 1.0 and later

From 1.0 on, the calls hang off a client object.

from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY
response = client.chat.completions.create(
    model="gpt-4.1-mini",
    messages=[{"role": "user", "content": "Hello"}],
)
print(response.choices[0].message.content)

The old names map like this. openai.Image, openai.Moderation and openai.Model raised the same APIRemovedInV1 as the rest:

Two more errors right after the rewrite

The reply is no longer a dict. Old code read response["choices"][0]["message"]["content"]. On the new SDK, that line raised:

TypeError: 'ChatCompletion' object is not subscriptable

openai.error is gone. except openai.error.RateLimitError: raised AttributeError: module 'openai' has no attribute 'error' on 1.55.3 and 3.37.0. The exception classes now sit on the module itself.

except openai.RateLimitError:

openai.api_base is ignored without an error

This one doesn’t raise anything. On 1.55.3 and 3.37.0, I set openai.api_base to a local server and called openai.chat.completions.create(). The request skipped my server and went to api.openai.com, where my test key got a 401. If you used api_base to reach Azure, a proxy or a local model, the new SDK quietly calls OpenAI instead.

openai.base_url = "http://localhost:8000/v1/"
# or, on a client:
client = OpenAI(base_url="http://localhost:8000/v1/")

Or keep the old code on 0.28

The error message itself suggests pip install openai==0.28. That still worked: on Python 3.14.5, 0.28.1 installed, and the old openai.ChatCompletion.create() call reached the real API. The API answered with my key’s credit error, so the request itself was accepted. It’s a stopgap, though: 0.28 gets no fixes, and no new API features such as the Responses API.

How this was tested

Windows 11 Pro, Python 3.14.5, October 23, 2026. Each openai version ran in its own venv against a local HTTP server that recorded every request and answered with a fixed reply. Only the 0.28.1 check went to the real API, with a key that has no credits. The shadowing case put a one-line openai.py next to the script. I didn’t run the openai migrate command that the error message mentions. The test scripts are public.