OpenAI just released 3o mini
- Thread starter Vikjohn
- Start date
- Oct 9, 2013
- 3,482
- 14,527
It's on a par with o1-pro in my tests.
MUCH faster, and of course cheaper, with the biggest thing being it's available in the API! So it's basically cheap fast API o1-pro level.
Don't use it for text gen. It's not for that.
For general purpose stuff 4o is better.
It's for math, science, coding and problem solving.
Vikjohn
Regular Member
- May 10, 2016
- 433
- 213
I tested it, na, this isn't good, even the the o3 mini high,
I might be wrong but but it sucks, feels like 4o
- Oct 9, 2013
- 3,482
- 14,527
Share your tests. If it feels like 4o, then you're probably using it like 4o, and it's a totally different type of model. It has completely different uses than 4o.
but if you share your tests I'll try to shed light into why you're finding it bad.
- Oct 9, 2013
- 3,482
- 14,527
Have to quote my own post here as I can't edit and this is an update..
Something's happened to o3. I've noticed other people on X seeing that o3-mini is suddenly shit.
Same for me. Last night when it first came out it was on par with o1-pro. Now, it's slower and downright dumb sometimes. It's now failing coding tasks that I pass to o1-pro which aces them.
- Oct 11, 2013
- 526
- 424
When I tested it yesterday, o1-mini was great at coding. o1-mini-low was giving errors. o1-mini-high was unnecessarily slower. Will have to check today.
What client do you use for API access?
Edit: - I did some quick tests. It's failing at coding tasks badly now. You are right.
Last edited:
- May 8, 2012
- 4,220
- 3,728
all the models even the new ones are kinda shit at coding still. yes for simple tasks it works fine, but more complex things you constantly have to refine and when you can code yourself you see the flaws and ask the model about it and it just goes on like "oh you are right, this would perform better" then it rewrites the code, but often starts to miss other things and you have to keep questioning it to make it work as expected and every iteration it seems worse. you kinda get stuck in a loop reasoning with it to make it better or whats missing again now. the resulting code is often average and doesnt take CPU usage, RAM and general speed into account even if you instruct it to do so. in the end you would save time just writing it yourself if you are a good coder... for hobby coders who dont develope commercial software its fine though, it gets the task done even if its not the best way to do it... so now im just using it sometimes for simple functions that im too lazy to write myself to save time, but more complex stuff i code myself. AI can't beat good human made code yet 
RedditWarehouse
Junior Member
- Nov 18, 2024
- 189
- 39
i tried today and that was awesome
- Oct 9, 2013
- 3,482
- 14,527
You mean o3, not o1
There is no o1-min-high etc. I'm sure that was a typo.What client? Python ;-). All my AI is custom. I don't use AI tools. I build AI tools. Other than the core models, but even then I do extensive fine-tuning, usually with synthetic data from the current best LLMs.
all the models even the new ones are kinda shit at coding still. yes for simple tasks it works fine, but more complex things you constantly have to refine and when you can code yourself you see the flaws and ask the model about it and it just goes on like "oh you are right, this would perform better" then it rewrites the code, but often starts to miss other things and you have to keep questioning it to make it work as expected and every iteration it seems worse. you kinda get stuck in a loop reasoning with it to make it better or whats missing again now. the resulting code is often average and doesnt take CPU usage, RAM and general speed into account even if you instruct it to do so. in the end you would save time just writing it yourself if you are a good coder... for hobby coders who dont develope commercial software its fine though, it gets the task done even if its not the best way to do it... so now im just using it sometimes for simple functions that im too lazy to write myself to save time, but more complex stuff i code myself. AI can't beat good human made code yet
o1-pro is definitely not. It's the first model where it's making my life easier. sonnet 3.5 etc, no way. I use them like I'd use old school man pages.
But the trick with o1-pro is giving it full context.
I've written a custom dev environment so I pass EVERYTHING to it.
It gets all current code, the hierarchy, the design notes(how to design, technologies to use, practices to follow), the instructions(how to work with all the info) as well as a complete historical log of ALL changes with summaries of every file change, and summary of the main thing it did.
Each change has the task class. I have 3 task classes.
1) Add feature
2) Bug fix
3) (This is a special one) - Analyse code and make suggestions
It's a really deep involved system and I'm constantly adding more.
When you give o1-pro THIS kind of structured info, it'll blow you away.
You can't just paste some code and ask it to write
The same, that you can't just paste some code to me and expect me to magically come up with a solution. The difference is, as a more advanced intelligence I can take the code and formulate designs, instructions, best practices and other things for myself, then follow them..
But if you give o1-pro those things, then it'll perform incredibly.
There's a LOT you can do if you get creative. Not just with code creation, but all aspects of AI.
- Oct 11, 2013
- 526
- 424
Oh Yeah! I meant o3 of course. It was not a typo, but rather misplacing the name in my mind.You mean o3, not o1
What client? Python ;-). All my AI is custom. I don't use AI tools. I build AI tools. Other than the core models, but even then I do extensive fine-tuning, usually with synthetic data from the current best LLMs.
You make AI tools? I don't use AI tools, which are just wrappers around funny prompts.
You don't use an IDE? You don't know what you are missing. I use a throughput of 200k tokens per minute through APIs. Can't do with less than that for larger programs. I don't know if someone who can develop an AI Tool with the capabilities of an IDE. ( Don't call the plugin for an IDE an AI Tool
)
- Oct 9, 2013
- 3,482
- 14,527
Oh Yeah! I meant o3 of course. It was not a typo, but rather misplacing the name in my mind.
You make AI tools? I don't use AI tools, which are just wrappers around funny prompts.
You don't use an IDE? You don't know what you are missing. I use a throughput of 200k tokens per minute through APIs. Can't do with less than that for larger programs. I don't know if someone who can develop an AI Tool with the capabilities of an IDE. ( Don't call the plugin for an IDE an AI Tool
An IDE isn't a client. You asked me what client I used, so I didn't know you meant an IDE
A client generally means something that accesses a server.
I use Cursor just now. I only use Cursor(vscode) because I wrote my convenience AI dev extension in it, but before that I preferred editors and used sublime and emacs. I do all my development on linux servers, so I debug in the console. I'm pretty old school
I've been coding since the 90's. Read my first AI book around 1998.
- Oct 11, 2013
- 526
- 424
An IDE isn't a client. You asked me what client I used, so I didn't know you meant an IDE
A client generally means something that accesses a server.
I use Cursor just now. I only use Cursor(vscode) because I wrote my convenience AI dev extension in it, but before that I preferred editors and used sublime and emacs. I do all my development on linux servers, so I debug in the console. I'm pretty old school
An IDE is also a Client, because you are accessing server data. In this case API. And my question was in regard to API. Hence, Client. The way I see is, we are both right. We just misunderstood each other.
I have been coding since 1997 as well.
I have been using linux for I don't remember when.You read your first AI book in 1998? Din't know you were such a big shot.
P.S. Life is not MMORPG, that you think you are more skilled with age. A 15 year old kid will beat you at many things AI.