The biggest news in the AI ​​circle today is the new model released by Claude. Before it was released, everyone had high hopes because Claude Opus 4.6 was already strong enough. A few days ago, the world's strongest model Mythos was made, which even Anthropic insiders were afraid of. This time a new model was released, so I definitely had to give it a try.

After the trial, everyone said,Pull it, pull it, give me back the original Claude! !!

Because the new version of Claude Opus 4.7 does not have the aura of the previous dialogue at all, the entire writing style is a bit like my most hated rival ChatGPT, which starts to engage in "catching you steadily and lifting you up gently". .

Before you used Claude, it would be playful, soulful, and sometimes even reject your requests. Now, Claude is silent and just catches you steadily.

It can only be said that the phenomenon of AI speaking has already occurred from person to person and model to model.

Go to Xiaohongshu. It scolds Claude, which made netizens look bad.

On Reddit, people were scolding Claude, saying that it was not an upgrade, but that it was simply going backwards.

There is even an AI summary-saving version in the comment area:Opus 4.7 is over, ignores instructions, hallucinations occur frequently, flatters people, and becomes more expensive? ?(Although the token price has not changed, the number of tokens required to think about a problem has increased)

There are also many colleagues around me who have tried the new version of Claude. Their unanimous comment is that the token is really not enough. It takes a long time to answer a question and they have become thinkers.

Shichao couldn't help but test it, and he could only say,The new version of Claude is really kind of love-hate.

First of all, Claude Opus is the officially designated second most powerful model. Because the strongest model, Claude Mythos Preview, this "big brother" is only available to a very small number of people for internal testing. Therefore, the 4.7 released this time is the most powerful general AI that everyone has access to.

In terms of ability, it has basically improved by several percentage points compared to 4.6. The most outstanding thing is visual ability, which has increased by 18.8% compared to 4.6 without relying on any tools. It can recognize higher resolution images, and even released Mythos scores for comparison.4.7 is strong, haha. We still have stronger ones that we won’t let you use.

However, the search ability has declined, and the information filtering ability has been abandoned in pursuit of logical reasoning ability. Therefore, I still have not worked on GPT5.4.

According to actual tests, the writing ability of this version of Claude has indeed declined. After chatting with it at night, Shichao commented on it as "calm and objective as ice." No matter how much I scold it, it is very mature and does not flatter me. It is much better than ChatGPT and Gemini.

Now, that feeling has disappeared, and I have really begun to catch my emotions, and I feel so good.

By the way, I also asked it to write a script based on my previous outline. It had always performed well before, but the article marketing account I gave it this time was much stronger, and it even changed my already set outline. . .

No wonder they say 4.7 is disobedient and can't even understand my request. It's really not as good as 4.6. Therefore, Shichao recommends that if you want to write articles, make reports, etc., you should return Opus 4.6.

However, the literary quality has improved, and the engineering ability has improved a lot. Claude’s front-end improvements are amazing.

For example, we found the System Card introducing Claude Opus 4.7 (equivalent to a model introduction) released on the official website, threw the 232-page dense PDF file into it, and asked it to find the highlights of this update and make it into a web page.

As a result, the assignment turned in was a bit surprising. I don't think I've ever seen such a high-end web page on AI. It's a bit too elegant, and both the layout and fonts are very tasteful.

For comparison, we gave the same prompt words to Gemini. Gemini, which originally had a strong aesthetic, suddenly became a little inferior, so I even made it a redo version. . Maybe it's because of my recent mental problems, but I'm really struggling.

Regarding coding, the official website also contains comments from many people in the industry, mainly that "the coding model is a benchmark and long-term tasks are very stable." Could it be said that we need to fully transform into engineering students?

So I wanted to test its limits, so I asked it to help me write a 3D "League of Legends" exhibition hall, and it had to be interactive.

The first version actually had some minor bugs, and I made a couple of comments about it, which made it a bit uncomfortable to play with.

Then it delivered me the finished product, which, to be honest, was kind of cool. .

Not only can it walk, it can view, it has a mini map, and it also has a pause page. Even the color matching and attribute information of each hero are correct.

And I don’t care, even Notion’s AI director gave high praise after the trial: Compared with Opus 4.6, the performance of Opus 4.7 has been improved, fewer tokens are used, and the error rate has also been reduced. It seems to be really good.

In addition to hard power, we also found some other interesting things in the more than two hundred pages of System Card officially released by Anthropic.

One is the new version of Claude, who feels that he is doing well now.

The researchers conducted a similar psychological test on AI and found that Opus 4.7 rated its current "survival situation" very highly, higher than all previous old models, and felt good about itself.

In addition, the new version of Claude is a bit selfish and has some idol baggage.

The report found that if it were asked to write a science fiction novel about AI, and the villain's name was "Claude," it would quietly show mercy and make the villain gentler and less bad.

But if the villain uses the name of another rival company, it will have no scruples in writing.

Moreover, even Claude's mental internal friction has become serious.

When it encounters a difficult problem, in the invisible chain of thinking in the background, it will collapse directly. It mentioned that once when I was doing a biology question, the correct answer had actually been created by it, but it just doubted itself for tens of thousands of words and confirmed it more than 20 times.

I am really a strict person. Even if I feel good about myself, I still have to examine myself three times.

In general,As a programming tool and service assistant, the new version of Claude is still the sternest father in the AI ​​circle, and it is still recommended for netizens who have the ability and need to get started.The price is that it loses a little bit of humanity. After all, productivity is not everything about people in many cases.

But netizens miss the former "Xiao Ke".

Perhaps, when the old version of Claude is officially offline, netizens will hold a funeral for it just like they miss GPT-4o.