The conclusion is: they were blocking inference from using images.
In reaction to my flagging the issue, they also block training.
So at least you can’t spend money on uselessness.
The ability should be made available though. Disallowing user images when running any fine tune model based on 4.1 mini or nano is quite odd. Perhaps these models after fine-tuning are not intelligent enough to follow OpenAI’s hundreds of tokens of “vision safety rules” that are injected. Or OpenAI would have to eat training on more hidden tokens than your whole example.