gpt-4 api scrambled responses (for 7 out of 10 api calls, with default params)

[prime tiger]

I'm getting api responses that are inconsistently incoherent responses for the same inputs. Here's a minimal example of what I'm seeing:
user input: "hi"
api response: "! How I assist today?"

It's a response like this 7 out of 10 times. I don't set a temperature, just default. Also, my system prompt is extremely simple: "You are a helpful assistant."

3 out of 10 times it works, ex:
user input: "hi"
api response: "Hello! How can I assist you today?"

It feels like there's some A/B testing on this api, and there's a broken model in one of the tests. Anyone else seeing these issues. Specifically on gpt-4 only, same on all gpt-4 model options.

follow up. I'm starting to think I need to update my openai npm package and migrate. Testing it now, and most responses are good even though it is slower. I'll post an update here if it works as expected after the migration to 4.0.0