Looks like different models differently processing json_object type.
I am configuring Assistant which supposed to return json object back… and I found that in case set response_format={“type”: “json_object”}, then only gpt-4 returns text.value with json decoration. While gtp-4-turbo and gtp-40 both returns plain text withg no such decoration.
And oppossitely if I set response_format={“type”: “text”}, then gtp-4 process it properly… while two others return json decorated text 
…
I suppose this is bug.