beaaaan's blog

GPT o3-mini is out
By beaaaan, history, 5 weeks ago, In English

Apparently, it's codeforces performance beats both o1-mini, o1 and deepseek r1 in codeforces rating:

It's price is on par with o1-mini, I guess they felt the heat from deepseek r1:

  • Vote: I like it
  • +5
  • Vote: I do not like it

»
5 weeks ago, # |
  Vote: I like it +20 Vote: I do not like it

o1 is claimed to be 1800 and it fails div2b, i'm suspicious of this

→ Reply
  • »
    »
    5 weeks ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    which round is it from may i ask?

    → Reply
    • »
      »
      »
      5 weeks ago, # ^ |
        Vote: I like it +6 Vote: I do not like it

      I remember someone posting a blog about how o1-pro (the 200$ a month version) couldn't solve 2040B - Paint a Strip, but o3-mini can solve it :(

      → Reply
      • »
        »
        »
        »
        5 weeks ago, # ^ |
          Vote: I like it +26 Vote: I do not like it

        Maybe just because o1 doesn't include it in a training data while o3 does? (round happened right in between their releases)

        → Reply
        • »
          »
          »
          »
          »
          4 weeks ago, # ^ |
            Vote: I like it +4 Vote: I do not like it

          I would honestly doubt that the training data cutoff is that recent, though I could be wrong, which I do hope

          → Reply
    • »
      »
      »
      5 weeks ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      most of the rounds

      → Reply
»
5 weeks ago, # |
  Vote: I like it +8 Vote: I do not like it

I just tested it on all the problems that Deepseek R1 failed that I had tested (from https://codeforces.net/blog/entry/138735 ), it solved all of them (though it took 2 attempts on Maximum AND Queries (Easy version)). I also tested it on Paint a Strip, which o1-pro (which was 200$) wasn't able to solve.

I'm also on the free plan, meaning my o3-mini is on low compute (if it means anything, it also has way shorter wait times)

→ Reply
»
5 weeks ago, # |
← Rev. 2   Vote: I like it +3 Vote: I do not like it

nvm

→ Reply
  • »
    »
    5 weeks ago, # ^ |
      Vote: I like it 0 Vote: I do not like it

    the edit...

    It's nice to know that the lower rated problems aren't entirely screwed, though it's a bit nerve wracking seeing it solve problems on the free plan that it used to not be able to solve on the 200$ plan

    → Reply
»
5 weeks ago, # |
← Rev. 2   Vote: I like it 0 Vote: I do not like it

Source

→ Reply
  • »
    »
    5 weeks ago, # ^ |
      Vote: I like it +3 Vote: I do not like it

    Note: o3-mini has already achieved a rating(allegedly) of 2130 (above cf master) with the setting set to high reasoning.

    → Reply
»
5 weeks ago, # |
← Rev. 2   Vote: I like it 0 Vote: I do not like it

If you have access to this model, please try this question — 2060D - Subtract Min Sort. This is a simple 1100 rated problem that DeepSeek-R1 gets wrong despite me giving it more than enough hints.

(update: I tried it myself. It failed on first but after giving it a very obvious hint, it was able to solve it. https://chatgpt.com/share/679e03a1-e82c-8004-b362-a745b412bf52)

→ Reply
»
5 weeks ago, # |
← Rev. 2   Vote: I like it +32 Vote: I do not like it

cf is cooked

→ Reply
»
5 weeks ago, # |
  Vote: I like it +1 Vote: I do not like it

well its not gpt-o7 yet xdd

→ Reply
»
4 weeks ago, # |
← Rev. 2   Vote: I like it 0 Vote: I do not like it

2036 is medium version. High version is 2130. Only 2400 people higher right now.

→ Reply
»
4 weeks ago, # |
  Vote: I like it +8 Vote: I do not like it

I tried o3-mini-high on last completed round. I tried for A, B, C, D, E1.

Results:

A -> solved in 1m 12s

B -> solved in 2m 17s

C -> WA on test 2 (surprising, I thought it'll be able crack this)

D -> WA on test 1

E1 -> WA on test 2

→ Reply
  • »
    »
    4 weeks ago, # ^ |
      Vote: I like it +4 Vote: I do not like it

    Sometimes o3-mini-low is better than o3-mini-high

    It solved C in one-shot 303999547

    → Reply
    • »
      »
      »
      4 weeks ago, # ^ |
        Vote: I like it 0 Vote: I do not like it

      it's not necessarily better. you probably got lucky with the sampling.

      → Reply
»
4 weeks ago, # |
  Vote: I like it +2 Vote: I do not like it

Gpt o3 mini low is really insane, it has solved almost all the problem i've asked upto 1900 rated.It's going to ruin all the fun. I think codeforces rating system just got cooked,as free users can access it.Contests should be unrated untill they find any solutions.

→ Reply
»
4 weeks ago, # |
  Vote: I like it 0 Vote: I do not like it

I have access to R1 Pro and it runs out of token before answering any problems. LOL

→ Reply
»
4 weeks ago, # |
  Vote: I like it 0 Vote: I do not like it

Bruh, o3 will beat me for sure...

→ Reply
  • »
    »
    4 weeks ago, # ^ |
      Vote: I like it -12 Vote: I do not like it

    I don't think so, these models most probably solving problems whose solutions are already available on internet.

    → Reply