Unwanted injection into new GPT-4o model system messages by OpenAI

I’m seeing this in mini as well as gpt-4o-2024-08-06.

Yes, still gpt-4o-2024-08-06 as reported earlier…

in json_schema, for example…

{your system}

# Response Formats

## response_output_name

{"type":"object","properties":{"reproduction_string":{"type":"string"}}}

You are trained on data up to October 2023.

It seems while justification was provided for mini, we get to live with it in the larger model.

They’re giving away free input tokens, though. If you go by the API usage returned. :smile:

I am having the same issue with gpt-4o-moni (version 2024-07-18) today.