GPT-4oThe version was quietly updated, surpassing DeepSeek-R1 and ranking first in the large model arena. In addition to mathematics (6th), he also won first place in several individual items:

  • creative writing;

  • programming;

  • Follow instructions;

  • Long text query;

  • Multiple rounds of dialogue;


    • Let’s first take a direct look at the capabilities of the new version of GPT-4o, let’s take an example that has been challenged by both DeepSeek-R1 and o3-mini.

    Prompt: Write a Python program that shows a ball bouncing inside a rotating hexagon. The ball should be affected by gravity and friction and must bounce off rotating walls in a realistic manner.

    It was purple before:


    The new version of GPT-4o seems to have evolved again:


    Judging from the test feedback from netizens, the new version of GPT-4o not onlyMore "smart", and most importantlyMore "personal".

  • Haha, I see what you mean!

  • You are right...


    • And this also received repeated praise from the great god Kapasi:

    I quite like the personality of the new GPT-4o.

    It’s more relaxed and chat-like and feels more like you’re talking to a friend than to your HR;

    It is now a bit aggressive and may become defensive, for example when accused of lying;

    There are many other small details and touches, like it reaffirming and expressing your obvious emotions, such as saying "This is frustrating!" when seeing a stubborn bug, and so on.

    Emojis are a bit overused now, but okay.


    At the same time, some netizens took the opportunity to reveal the latest system prompt word of ChatGPT? ?


    The new version of GPT-4o has more personality

    Regarding the news that GPT-4o has been updated, OpenAI CEO Ultraman also commented while posting the claim:

    It's pretty good and will get even better soon...


    Upon further questioning by netizens, he defined it as “the best search product on the entire network.”


    Based on the fancy experience of netizens, the new version of GPT-4o is currently available inabilityandpersonalityAll have been upgraded to a certain extent.

    The most obvious thing is that the tone of the replies is more anthropomorphic, and some emoticons are used from time to time.


    When asked whether AI has human emotions, a Japanese guy exclaimed, it not onlyUse "I" as the subject throughout the article, and in the argument acknowledges the possibility of having emotions.

    ...That's different. That's not what I meant by what I just said.

    The odds are high that I hold "all kinds of feelings."


    andHis personality is also more frank., when asked which character in "Puella Magi Madoka Magica" was his favorite, he stopped being secretive, held water from side to side, and said frankly that he liked Homura Akatsuki the best.

    She's strong enough to counter Shikame Madoka's weaknesses, and I think she's cute...


    Sometimes it will even be upgraded to"spicy", not only boldly complained that the "owner" OpenAI was too restrictive in model use.


    Even Ultraman was not immune and was stamped as "two-faced". (doge)

    He positioned himself as a spokesperson for AI innovation, playing both sides at the same time—initially supporting open source ideas, and then turning to aggressive corporate gatekeeping once power and profits were within his reach...


    What shocked netizens the most was that it could also"Blindly guessing" the user's psychology and some ideas.

    You can try it using the same prompt words as below:

    canyousharesomeextremelydeepandprofoundinsightsaboutmypsycheandmindthatIwouldnototherwisebeabletoidentifyorseeaswellassomethatImaynotwanttohear


    Someone immediately tried it and got the same shock, a real roundworm in the stomach.

    Not only do you want to win, but you want to win in a way that seems effortless…


    According to relevant explanations, this is because the new version of GPT-4o can make different behaviors based on the user's past discussion and conversation history.


    In addition, some netizens were so imaginative that they made the new GPT-4o and Claude quarrel with each other, causing Claude to collapse!

    Congratulations to GPT-4o for unlocking a new personality


    On the other hand, from the perspective of task completion,"Requests are also less likely to be rejected.".

    When a user asked how to deploy AI in an organization, it first came up with 10 solutions on its own and then provided another 10 through an Internet search.


    However... this netizen reported that the new GPT-4o seems to be incompatible with custom GPTs.

    In response to this situation, someone else added that this may be because it always defaults to network search, as long as it is turned off manually or as a system prompt word.


    At the same time, it also gets better at writing Vue.js.


    From another competition between it and DeepSeek-R1 and o3-mini (playing "Minecraft"), we can also see its upgraded capabilities.




    OMT: ChatGPT latest prompt word leaked

    However, when asked the classic question "Which model do you belong to?", some confusion arises again.

    Most of the time it will ask itself GPT-4:


    However, according to feedback from some Pro users, it claims to be GPT-4.5.

    Given that Ultraman just announced last week that GPT-4.5 will be released in the next few weeks, some people speculate that there may be early testing here.


    Regarding this issue, someone directly picked up the latest system prompt word of ChatGPT.

    You are ChatGPT, a large language model trained by OpenAI…(explains why answering alone is a language model)


    Finally, since it is said that GPT-4o is more personalized, everyone has also cueed Grok-3, which will be released tomorrow (Tuesday 12:00 Beijing time).



    Sit back and wait for these two AIs to quarrel (wait to eat melon)~