The overall cost is the same per image. Perhaps it is just their way of representing the same cost with the lower per token cost? Which just highlights how much cheaper the text tokens are.
3 Likes
Yes, that could indeed be a hint. Great development!
1 Like
Thx mate, I looked through my code and I realises I’m white washing specific models, so the issue was in my own code … 
(Embarrassed blush!)
Do you mean “white listing”? That’s rather different!! 
Btw the modern phrase we are supposed to use is “allow listing” 
1 Like
Will it be possible to fine-tune this model?
So, is GPT-4o also supports 16,384 max_tokens?
Hehe, yes 
I meant white listing 
Yes, @jamie10 - exact same issue. I already thought it might be a fallback or some additional context, that mini needs.
If i do it in Chat comparrison, upload a picture and ask open AI to explain the Picture, GPT4o will use around 800 Tokes, GPT4o-mini will consume 25,000 tokens
Replying to myself…turns out some of my functions needed some updating and care. I can gladly report that function calling on GPT-4o-mini is phenomenal and I’m so glad to be done with the nightmares of 3.5 this past year.
Thank you so much and looking forward to building more with it!
3 Likes
Is fine tuning not available yet or have I coded it wrong?
openai.BadRequestError: Error code: 400 - {'error': {'message': 'Model gpt-4o-mini is not available for fine-tuning or does not exist.', 'type': 'invalid_request_error', 'param': None, 'code': 'model_not_available'}}
Fine-tuning will be available shortly for the 4o-Mini model.
4 Likes
Is there any news on how much fine tuned models will cost at inference time?
Other fine-tuned models have generally been 4–6 times as expensive, I would expect Mini to likely follow that trend.
Although GPT-4o mini theoretically supports a 16k token output, my tests suggest that it appears to be constrained to around 4k tokens, even to the extent of generating hallucinations to stay within this limit. When I attempt to push it to extend beyond 4k tokens, the model often resorts to ellipses or claims to have generated content that doesn’t exist.
I could be mistaken in my assessment. Has anyone else tested this feature?
1 Like
I’ve managed to get the model to generate ~10k tokens by telling it that it can do that and forcing it to write a story that “never ends”. I have no idea how useable it would be for more serious use cases though.
3 Likes
how long will it take for the fine-tune to roll out? Is there a way to request access to it now?
Has anyone done any vision speed response testing between the 4o and 4o-mini api?
Thanks for the updat and this is very interesting
Can you tell us about the details of the pricing
You can find pricing at openai.com/api/pricing