GPT-5 is Coming, and It's Powerful! Leaked Benchmarks + Minecraft Test Directly Dropped Jaws Across the Internet: Codenamed Zenith, GPT-5-pro Performs Seamlessly in Games, Dubbed "Magic-Level AI". Rumors Suggest It Will Be Released on July 31st, Outperforming Grok 4 Heavy! Is OpenAI About to Overturn the Table Again?
Early in the morning, news about GPT-5 has arrived again.
These leaked GPT-5 benchmarks are likely to be real.
[Image links omitted]
There's even a shocking message: GPT-5 will be released on July 31st.
As a result, all GPT-5 models have now officially withdrawn from the WebDev arena.
[Image links omitted]
However, there's another perspective from Menlo Ventures investor Deedy, and media outlets like The Verge and The Information, suggesting GPT-5 will debut in August.
[Image links omitted]
Although GPT-5 hasn't arrived yet, tests about it are already flying all over the internet.
Just recently, someone released a test of GPT-5 replicating the Minecraft game. To be precise, it's the internal codename Zenith GPT-5-pro.
This netizen commented: "Impressive, it's like magic! OpenAI has indeed created something incredible."
In this video, GPT-5 smoothly completed game tasks in one go, with truly stunning performance.
With expectations raised so high, GPT-5's official release must be incredibly explosive, otherwise, they won't know how to wrap it up.
[Image links omitted]
There's another heavyweight leak from famous leaker Jimmy Apple.
According to him, many internal evaluators rate GPT-5 as stronger than Grok 4 Heavy.
[Image links omitted]
GPT-5 is Coming, Everyone is Holding Their Breath
Now, the presence of GPT-5 is getting closer and closer.
Some even discovered that when selecting o3 in the app, they accidentally tested a version of GPT-5.
[Image links omitted]
More and more people are accidentally testing GPT-5.
[Image links omitted]
News about its launch this week has been increasingly confirmed.
[Image links omitted]
The Verge's perspective is slightly different, with their intelligence suggesting GPT-5 will be released in early August, including mini and nano versions.
[Image links omitted]
Previously, developers discovered that GPT-5 was internally named "Reasoning Alpha Version".
Meanwhile, a model codenamed "o3-alpha" was quickly taken down just 12 hours after going online, and many believe this is the early shell of GPT-5.
According to OpenAI's convention, the interval between testing and release is as short as 4 days, so GPT-5 is indeed very close.
Just yesterday, everyone discovered that GPT-5 can be used on LMArena. The Zenith model was also discovered simultaneously.
The following examples have already gone viral across the internet.
Generating starship control panels from a distant future.
[Image links omitted]
Creating a streaming website.
[Image links omitted]
Perfectly presenting SVG animation in robot walking.
[Image links omitted]
The best pineapple defense game.
[Image links omitted]
Merging O-Series and GPT Series
Undoubtedly, GPT-5 is now the most anticipated model globally.
Many believe GPT-5 will be a significant milestone that will attract millions of users to the AI ecosystem.
Next, we will sort out the various clues about GPT-5 mentioned during this period.
GPT-5 was mentioned in a live broadcast about OpenAI's agent.
[Image links omitted]
The key information at the time was: This amazing frontier model will first unify models from two series, concentrating the o-series' breakthroughs in reasoning and the GPT series' breakthroughs in multi-modality.
[Image links omitted]
Because ChatGPT has various model types, each model has its unique functions and outstanding characteristics. If GPT-5 is truly a collection of the best parts of each individual model, user experience will obviously be completely transformed.
[Image links omitted]
For instance, those who have used o3 know how crazy the leap from GPT-4o to o3 is.
This was confirmed by OpenAI CPO Kevin Weil as early as February this year.
[Image links omitted]
A netizen asked: Will you create model routers, or will they be more unified in a systematic way? Weil stated that it will be more unified.
There is another leak from a suspected OpenAI internal employee. He said that researchers did indeed try routing methods, but they generated many hallucinations.
So, they are testing a model that can plan, reason, and use intelligent agents as an extension.
Then there are some leaks from the foreign media The Information.
In summary, GPT-5's coding is extremely strong.
In natural sciences, reasoning is more in-depth;
Automatically completing complex tasks in browsers;
Writing is more fluent, logic is more online;
More importantly: a breakthrough in coding!
According to an experiencer, GPT-5 is not only better at solving academic and programming competition problems, but even performs more impressively when handling actual programming tasks faced by real-world engineers.
For instance, it can modify complex code libraries containing a lot of legacy code without hesitation.
It is precisely this delicate ability to handle complex scenarios that has kept OpenAI's models behind Anthropic. After all, in the developer community, everyone acknowledges that Claude is the true programming king.
An experiencer who personally tested it said that GPT-5 in programming even directly surpasses Anthropic's Claude Sonnet 4!
Another perspective is that GPT-5 is not a unified model, but a router mechanism.
It will send your query to a GPT large model that excels in casual chat, or an o-series model that excels in logic and reasoning, depending on the type of question.
The performance of GPT-5 that we see is the result of these two models' combined punch.
Even OpenAI executives privately predict—
We are confident in achieving GPT-8 without changing the architecture.
In other words, OpenAI does not intend to compete with new architectures, but rather push existing technology to the extreme through smarter scheduling, stronger reasoning, and more post-training data.
What will GPT-5 bring to the world?
Meanwhile, Altman's recent statement about "GPT-5 making him feel useless" has raised even more expectations.
Some say that GPT-5 might be one of the most dangerous things currently happening in the AI field.
For instance, Altman mentioned in this interview that many people chat with AI all day, even treating it as their boyfriend or girlfriend.
Some children are obtaining dopamine entirely through scrolling screens during their growth. These things are very dangerous.
When the host asked: How to prevent AI from having the same negative impacts as social media? Altman honestly admitted: I'm very afraid of this, and I don't have an answer.
Worryingly, just a few days ago, an OpenAI investor admitted that due to using ChatGPT all day, he has experienced some abnormal conditions.
In other words, even wealthy people can be triggered into mental illness by chatting with AI.
Altman even stated that he is very interested in providing free GPT-5 to every person on Earth.
When these AI products and services are provided at 1/100 of the cost, it is obvious that certain economies will quickly transform and collapse.
However, no matter what kind of wave it will cause in the world, the momentum of GPT-5's launch is now unstoppable.
References:
https://x.com/chetaslua/status/1949905375546708242
https://www.youtube.com/watch?v=0jDsWemXi3U
This article is from the WeChat public account "New Intelligence", author: New Intelligence, published by 36Kr with authorization.
