I can’t speak for what OpenAI is doing, but I’ve noticed those types of hallucinations occurring when I quantize a model beyond a certain point.

Maybe they are trying to cut down on memory usage ?