Ah this was likely because of a change we made to make imports faster, for now you should be able to workaround this with import openai.resources, can you let me know if that fixes the issue?

Could you also share what you're using the openai.resources import for? Trying to assess how common this breakage could be.

We're hooking in to openai.resources.chat.Completions.create and intercepting arguments

        original_create = openai.resources.chat.Completions.create

        def intercepted_create(_self: openai.resources.chat.Completions, **kwargs: Any) -> Any:
            ...

        # Replace the original method with our intercepted version
        openai.resources.chat.Completions.create = intercepted_create

as well as type hinting

completions: openai.resources.chat.Completions

The former usage is a bit obscure, but the latter is common and surprising not to work in my opinion.

Thanks for looking in to this.

Can you try this?

python -m pip install git+https://github.com/openai/openai-python@robert/fix-imports

Works! Thanks again.

Just another use case: I have tests that monkeypatch openai.resources.chat.Completions/AsyncCompletions to not send actual requests during most tests.