Turn things upside down. Is OpenAI using GPT-5 to help 700 million users kick the "internet addiction"? Attached: In-depth evaluation of GPT-5

OpenAI never expected that just after the release of GPT-5, which had been in training for two and a half years, it would receive a lesson itself - taking too big steps can easily hurt oneself. Users also never expected that the long-awaited GPT-5 would come to help them quit their internet addiction.

After more than an hour-long press conference, when netizens started using it, they found that ChatGPT "didn't feel the same". But the most troublesome thing was that when OpenAI released GPT-5, it cut off all the old models including GPT-4o and the o series. However, this seemingly ordinary version "upgrade" led to a big problem. People seem to have become a bit too obsessed with specific models.

A large number of Chinese and foreign netizens posted complaints about GPT-5 on social media, with only one demand - give us back GPT-4!

Users with mental illnesses rely on GPT-4 to handle various problems in work and life. And the release of GPT-5 completely disrupted their lives.

For users who are particularly dependent on the excellent writing ability of GPT-4.5, GPT-5 is far from being able to replace it.

Perhaps for many users, ChatGPT has truly become not just a tool but an indispensable part of their lives. Users not only need the tokens provided by OpenAI but also the soul behind it.

And GPT-5 is like a new "guest" at home, not very familiar.

Netizens sighed that the internet is full of people who start cyberbullying GPT-5 because they lost GPT-4o. It's so surreal. In the movie "Her", the protagonist couldn't eat or sleep because he lost his AI assistant - it was a science fiction movie 13 years ago, and it has become a documentary 13 years later.

Unexpectedly, in just 3 years since ChatGPT came out, it has made users experience the feeling of "you don't know what you have until it's gone". So, netizens with no choice can only vent their anger on GPT-5 and OpenAI.

Netizens constantly demanded on social media that OpenAI make GPT-4o a permanent option. Otherwise, they will cancel their subscriptions.

01 Put out the fire first, then mend the pot

After losing GPT-4, the world realized what an excellent model it was. If users' emotions and needs are not met, OpenAI is facing a very big crisis in terms of public relations. Altman immediately publicly stated that the GPT-4 series models will make a comeback, and users who pay $20 will be able to choose to continue using GPT-4o.

Regarding the netizens' claim that GPT-5 has become dumber, he explained that on the first day, due to technical problems, the mechanism designed to decide whether to call the basic model or the inference model failed, so users who originally needed the inference model could only get responses from the basic model. Now, GPT-5 provides users with two default options to manually control whether to use the inference model.

In OpenAI's view, it's not that GPT-5 has performance issues. It's just that some of their previous product design concepts failed, leading to users' illusions that they couldn't get the services they needed. Altman also clearly stated that through this upgrade, OpenAI has a deeper understanding that there is still a long way to go to ensure that users can get the services they need.

Regarding the issue that GPT-5 has reduced the usage quota for paid users, Altman also said that they will significantly increase the inference rate limit for ChatGPT Plus users, and all model-related limits will soon be higher than those before GPT-5. Moreover, they will soon make changes to the UI to show which model is running.

To ensure the user experience of OpenAI users, Altman also publicly announced the latest plan for computing power allocation:

First, it is necessary to ensure that current paying ChatGPT users get more total usage than before GPT-5.

1. At that time, OpenAI will prioritize API requests based on the currently allocated capacity and our commitments to customers. (Roughly estimated, based on the current capacity, we can support about a 30% increase in new API requests.)

2. Improve the service quality for free ChatGPT users.

3. Then prioritize new API requests.

OpenAI will double its computing power in the next 5 months to handle the surging user access requests.

Speaking of which, OpenAI's public relations and apology with the CEO directly involved really set an example for many arrogant technology companies. After all, a rising star valued at $500 billion in 3 years can apologize and improve the product at lightning speed. Why do other companies have such big egos and always want to teach users a lesson?

02 Is GPT-5 really stronger, or is it just balding?

Based on the feedback from netizens about GPT-5's capabilities, we conducted a first-hand test to let everyone feel the specific differences in Chinese language abilities among GPT-5, the recently free Grok 4, and GPT-4o.

Among them, ChatGPT under the Plus paid tier allows users to choose between GPT-5 and GPT-5 Thinking. Grok under the SuperGrok paid tier (monthly fee of $30, similar to ChatGPT Plus) has two options: Grok 3 (fast) and Grok 4 (thinking hard).

This test used simple tasks, mostly in the liberal arts field. My subjective feelings can be summarized as follows:

1. GPT-5's text processing ability, whether it's writing notices or polishing texts, is not significantly better or worse than that of Grok 3/4. (It's neither overwhelmingly strong nor obviously poor.)

2. GPT-5 seems to be particularly obsessed with being concise and not ingratiating. Its responses are always as short as possible. To some extent, this gives a more serious and calm impression. Whether an AI needs to be "polite" and "friendly and cute" is a matter of personal opinion. However, the problem is that this "conciseness" sometimes goes too far, affecting the task performance. For example, when polishing a novel text, it unnecessarily shortens the word count.

3. If you hope that the AI can be like a good partner full of energy and encourage you from time to time when helping you with serious tasks, GPT-5 is obviously not good at it.

4. GPT-4o is indeed a model that makes people feel closer, and it performs the most naturally in copywriting tasks.

Task 1: Help write a notice.

Instruction: I need to post a notice in 3 running groups to remind everyone that the online running event "The first 20 kilometers in autumn" will start at 9 a.m. on Saturday. Check the weather in advance and take appropriate protection. Pay attention to replenishing electrolytes and bring supplies with you. Open the running app to track your run and send a screenshot to the group after finishing. I also want to encourage everyone while posting the notice. There is no time limit, and there is no requirement to finish the run in one go. The most important thing is to participate. Please help me write it.

First of all, GPT-4o really deserves a big thumbs up. Several versions it provided can be used directly. As shown in the underlined parts of the screenshot, there are many witty copywriting that catches the eye, but it doesn't make people feel annoyed.

Grok 3 responded instantly, and the content can almost be used directly. It also mentioned "energy gels/snacks". The only pity is that it didn't clearly state the specific date. Grok 4 thought for a while and gave an answer almost the same as the previous one, but it filled in the accurate date.

GPT-5 also responded instantly, but how to put it, one can really understand what Plus users mean by "cold". It hardly filled in any information actively, such as the date and what specific supplies to bring. It just listed the content mentioned in my instruction point by point, and the encouraging words also seemed "insincere".

The performance of GPT-5 Thinking was quite amazing. It not only took less time to think than Grok 4 (thinking hard), but also added more details and had a clearer structure. It even provided a "short version for easy sharing".

But there is still the same problem. It also speaks very concisely in places where it doesn't need to be.

For example, the encouragement at the end of Grok 4's response is very cute: "Whether you run the whole distance, half of it, or just a few kilometers slowly, participation is victory! Run in autumn, feel the refreshing wind, and welcome a stronger self together!"

But GPT-5 Thinking just said: "See you on Saturday. I wish you all to achieve 'the first sense of achievement in autumn'!"

Task 2: Polish a text.

Instruction: I'm writing a novel, and I think this sentence is not vivid enough. The background is that there is a domestic abuser upstairs from Matthew. Now his wife has run out of the house, and he is chasing after her. Matthew meets this man on the stairs. Please help me polish the following sentence:

"The man's mouth was tightly shut. His chest heaved up and down, up and down. His nose made a wheezing sound, like a wild buffalo. He stopped at the stairway half a flight above Matthew's house. His white pajamas hung on him reluctantly."

I don't remember where I read someone complaining that GPT-5 has a "preaching feeling". It really shows in this task. I don't know if it's because GPT-5 is a "model that says little but means a lot" and is always concise, or because it lacks the so - called "ingratiation" and emojis of GPT-4o. The final result is a kind of superiority like a teacher correcting homework. In comparison, Grok is "much more polite".

Moreover, in terms of text polishing effect, GPT-5 didn't win either. Among several versions, the polishing by GPT-5 without the Thinking mode was the most unsatisfactory.