Gpt-4o-mini spewing nonsense for translated strings

API
1

Anyone else having issues with gpt-4o-mini today?

I have a prompt which takes text in multiple langauges and returns a json schema of translated strings. Up until today this worked perfectly for a long time however this morning all requests return json schema with partially translated, mostly token based nonsense. Json schema still works, but the strings inside are not words.

Switching model to gpt-4o works but is significantly more expensive. Is there some type of ongoing crash or issue for gpt-4o-mini?

Thanks!

2 Likes
2

As of ~7:00 UTC we’re seeing JSON issues when using GPT-4o-mini with JSON schema. The responses are no longer following the schema structure, resulting in parsing errors. This looks like a regression.

3

After ~8h I tested gpt-4o-mini again and now the responses are back to normal. For now everything seems to be working.

It is kind of expected that these might happen although a bit scary in terms of production use. How do you (anyone reading this) handle validation or fallback mechanism? Any common practices for catching madness errors?

Thanks!