How are people finding the results from 4o-mini? I find that a prompt that returns great answers in 4o, will make up answers and become real round-about in 4o-mini. It is like 4o knows I want a final answer and 4o-mini just goes around and around, making things up.
Curious to hear other’s results with 4o-mini. I need to lower my API costs and I was hoping mini would be the way but these results just are not the same quality as I get using 4o.
I had a similer experience, 4o model gave an exact answer with accuracy, but the 4o mini model was struggling. But I gave it clear instructions in the system prompt which massively improved its performance, just think of it like someone who needs clearer directions to do the same task
1 Like
Since it’s a lot cheaper (4o-mini), you can try a one-shot or two-shot example which usually helps a lot and still ends up being cheaper than 4o.
2 Likes