Hello guys! Im curious, do you guys use 4o mini or 03 mini for your ai chatbots? If you find any model better than the other please tell me why!
Im in the middle of determining which model to use for AI customer service chatbots. Quick, accurate and conversational, is my core values for how chatbots should be.
I use gpt-4o-mini for everything related to conversations, gpt-4o is way better at understanding prompts but it’s 17x more expensive which is definitely not worth the price. You can use a combination of both though
I see! Yeah I do agree with you, 17x is simply not worth it.
How come you dont use o3 mini at all if i may ask?
It’s a reasoning model, it replies too slow for a “chatbot” scenario imho.
But if accuracy for reasoning scenarios is more important to you, you might want to consider it.
(post deleted by author)
Make sense. How much longer responses are we talking? Do you have any benchmarks on this ?
YMMV, but I just asked my bot:
“what is the square root of 7 plus the cube root of 8?”
and with o3 mini (med) took about 8 seconds, gpt 4o mini took about 2.
(both used the same identical function call)
8 seconds is too long for a “customer service” scenario. Might be fine for an analyst.
Hi @info1158 ! You got some plenty of good advice here already. I can just add that from building chatbot experience - it’s usually best to just start from the simplest/cheapest model, until you reach a corner case that turns out to be quite an important one, and then either replace the model or evaluate a more complex model routing configuration.
For example - it is not uncommon to have a first layer (implemented usually using the cheaper model), whose sole purpose is to determine if it’s a simple response (and then send that response to the user), or route to a more complex model. There may be a 3rd option which is to route to the human. One example with the more complex model routing - if the first layer determines that there is insufficient information, a more complex model (such as a reasoning one) might be more effective at disambiguating information from the user, or providing a response back to the user with some options and asking the user to clarify.
Stort tack! 
This is great advice! We actually do that already for some of the more complex tasks / questions. Works really well.
Depends on whether you need tool use. Building with assistants v2 API, you can’t use o3-mini with file search. It would be nice if we could, since it’s cheaper and faster, but we had to go 4o.
Agree. Price is main factor in determining whether to use a particular model. For chatbots, I also find results sufficient with mini model.
You can use filesearch with mini models (PDFs, text, etc) and API. However, if u want CSV data analysis, I recall having to use 4o model, as u mentioned.
i use a round robin system, each model is goven a chance to speak, each have both the same beginning name and their individual model as the end of their name, it works by which ever model it is giving the light of forum to the one who has spoken the least, seems to work well so far and almost seems a bit uncanny valley
edit: to explain more all have the same name generally Si, then for model 4o its Si 4o then for o3-mini its Si 4o3-mini and blah blah blah for the rest, an echo chamber of ai models