Since DeepSeek became popular all over the world, everyone in the AI ​​​​circle has accelerated as if pressing the fast forward button. Not to mention anything else, this week alone includes Musk's Grok3, Anthropic's Claude3.7Sonnet, Alibaba's QwQ-Max-Preview, Tencent's TurboS, Dark Side of the Moon's Kimi-1.6-IoI-High, Google's GeminiCodeAssist, and a lot of other things. There is also DeepSeek Open Source Week.They are simply immortals.

And just last night, Hui Hui was taken out and whipped. OpenAI, which claimed to be holding back every time, finally came out and took out the latest version of the GPT series, GPT-4.5.

According to Ultraman, this time GPT-4.5 is a different type of intelligence, with wonderful features that he has never experienced before. This will be the first one that makes you feel like you aretalk to a thoughtful personmodel.


However, since two months ago, my brother has been chasing them for boring press conferences at 2 a.m. for more than ten days in a row. To be honest, Ultraman has hurt my brother's heart a bit.

So we actually didn’t have much expectations for this GPT-4.5. Even the Ultraman people didn’t come to the press conference. They said they were going home to take care of the baby. Yes, it was the baby he and his husband had.

Anyway, looking at the overall picture, I can only say that this wave of GPT-4.5 can be regarded as the "Tang Tang" debut.

This is not a shame. Most other netizens have the same view on this thing. There is even a discussion on the Internet about whether GPT-4.5 is garbage, because even red necks are not optimistic about GPT and voted for Musk's xAI.


So, what is the drawing method of this thing? Without getting into the details, let’s just come to the conclusion, that is, the performance of GPT-4.5 is not good and the price is high.

The same was released last night, but he and su7u are completely opposite.

Let’s talk about performance first. In a benchmark test officially given by OpenAI, GPT-4.5 is not as good as the o3-mini released last year in science, mathematics, and coding capabilities. The benchmark test is only 5% better than 4o.


In other words, GPT-4.5 is not enough compared to its own o3-mini in difficult academic benchmark tests such as AIME and GPQA, let alone a monster room with DeepSeek-R1 and Claude3.7Sonnet.


Putting aside the data from the official website and looking at actual measurements from netizens, GPT-4.5 is quite different from Claude 3.7, which was also released this week.

For example, in terms of thinking understanding and map generation, Claude can almost move it to ppt to make illustrations, but the pictures drawn by GPT-4.5 are just like the homework in my elementary school computer class...


What's even more outrageous is that this thing runs very slowly...


But that's not the most outrageous thing.What really makes it stand out is its price.

According to their official data, the price of GPT-4.5 per million Tokens is US$75, which is a full 30 times increase compared to 4o, not to mention compared with DeepSeek, which can directly reach 280 times...

If you include DeepSeek's discount, the difference can even be more than 1,000 times!

But what’s funny is that OpenAI’s official website article also says GPT-4.5 “cannot be completely replacedGPT-4o".


However, OpenAI officials don’t care much about this. They think the real power of GPT-4.5 lies in its language capabilities.

They said in the blog on their official website that GPT-4.5 can analyze human emotional needs in conversations and is the best in providing emotional value.

“It combines a deep understanding of the world with better collaboration, resulting in a model that naturally integrates ideas in passionate and intuitive conversations better suited to human collaboration.GPT-4.5 is able to better understand human meaning and interpret subtle cues or implicit expectations with more nuanced “emotional intelligence.”"

For example, if you didn't do well in the exam, he would first comfort you when you tell him, but 4o is very straight and gives you a bunch of plans.


But what can I say? This does seem to be more humane, but training an AI with emotional intelligence does not seem to prove that it is really better than others.

Take Byte's beanbag as an example. If you send it this sentence, it will respond very humanely and even make calls.

Not only us, but also external netizens have tweeted about OpenAI, comparing it with DeepSeekr1 and Grok, openly expressing yin and yang.

To be honest, this is understandable.Labor and capital have spent the most expensive money on the market, but in the end, it requires algorithms to have emotional intelligence, reasoning to have emotional intelligence, and applications to have emotional intelligence...


Former OpenAI employee, well-known foreign AI analystAndrejKarpathyPosted an article saying that the training cost of GPT-4.5 is ten times higher than that of the previous generation, but the IQ is not as good as the inference model, but the focus is on the emotional intelligence of AI.

Although Andrej is quite satisfied with the emotional intelligence of GPT-4.5, saying that this is an improvement from GPT-3.5 to 4.0; he also pointed out thatGPT-4.5 is not an inference model, and may be the last generation of OpenAI's non-inference model.

In this way, it is expected that OpenAI will have better performance when it develops the next generation inference model based on 4.5.


However, from this point of view, the direction of most AI in the future may completely shift to reasoning.

On the one hand, the appearance of GPT-4.5 can actually be said in a sense,Traditionally, miracles can only be achieved through hard work, and ScalingLaw (the law of scale), which has a crazy amount of computing power, has begun to slow down.

On the other hand, the open source model camp is already on the road in this direction.


If nothing else, these days of DeepSeek Open Source Week,Every day, the core equipment used in the training and reasoning process of our own V3 and R1 are released for everyone to use for free.

For example, the FlashMLA architecture on the first day is equivalent to teaching you how to tune NVIDIA GPU, and teaching you step by step to squeeze out the computing power resources of H800;

In the following days, a bunch of databases and algorithms such as DeepEP, DeepGEMM, DualPipe, and EPLB were gradually opened; finally, a 3FS and Smallpond data processing framework was given to squeeze the performance of solid-state drives.


In communities like GitHub, developers in the AI ​​field have been very happy these days. Open source data such as DeepSeek tops the GitHub hot list almost every day. This wave can be said to be the new "source god".

While the performance of GPT-4.5 is mediocre, DeepSeek has made everyone have a gun. As a result, it is estimated that it will be difficult to see traditional computing power competitions in future AI training, and more cost-effective training is expected to become king.