GPT-4 ChatCompletion gives error 400 Bad Request

[upper geyser]

The Auto-GPT CI pipeline has been failing regularly with HTTP 400 Bad Request errors in the past days:
https://github.com/Significant-Gravitas/Auto-GPT/actions/runs/5526234696/jobs/10091185356

I haven't been able to figure out why, also because the error seems to disappear after re-running a job a number of times. Sometimes it takes 3 attempts, sometimes 9. However, the job linked above is the first instance where it consistently fails all 3 tests that need to make API calls. Adding a log statement to dump the headers and the content of the request shows a valid structure AFAICS.

[upper geyser]

Update: making the exact same request directly works fine. Could it be that OpenAI has tightened the screws on their anti-proxy tooling?

We use that proxy to avoid passing API keys to unaudited code

[tired flower]

this is very likely the cause

I strongly reocmend using a Cloudflare Worker for doing that proxy as that for sure would not be blocked by the very strict VPN filters