HomeArticle

GPT-5 makes a grand entrance. The free AI at the doctorate level dominates all the charts. It's a sleepless night for millions of programmers, and 700 million people are in a frenzy.

新智元2025-08-08 15:14
Witnessed by 700 million people, is the door to AGI opening?

GPT-5 has made a stunning debut! Since ChatGPT was launched in November 2022 and GPT-4 in March 2023, it has been two and a half years until the arrival of GPT-5. During this late-night live broadcast, tens of thousands of onlookers in China watched the event online. At least according to OpenAI, they are one step closer to AGI.

Amid the global users' anticipation, GPT-5 has finally made a spectacular entrance!

OpenAI presented a comprehensive showcase of GPT-5's explosive capabilities during an over-one-hour-long press conference.

Led by Altman, with a large number of participants, Chinese talents still shine brightly

Just when 700 million people use ChatGPT every week, GPT-5 has been released with a bang. It represents a significant upgrade from GPT-4 and marks an important milestone for OpenAI on the path to achieving AGI.

OpenAI introduced that this is the most advanced AI system we have ever had, with intelligence far surpassing all previous models and excellent performance in coding, mathematics, writing, health, and visual perception.

This unified system includes a smart and efficient model capable of answering most questions, a deeper reasoning model (GPT-5 Thinking) for solving more complex problems, and a real-time router.

The phased release of multiple versions such as GPT-5, GPT-5-mini, and GPT-5-nano means that OpenAI is actively building a general intelligent operating system with GPT-5 as the underlying core.

From now on, GPT-5 will become the default model in ChatGPT, replacing GPT-4o, o3, o4-mini, GPT-4.1, and GPT-4.5.

Meanwhile, all Plus, Pro, Team, and Free users can now use GPT-5.

If you subscribe with a fee, you can have unlimited access to GPT-5 and GPT-5 Pro, while free users will switch to GPT-5 mini after reaching the usage limit.

After the live broadcast, LMArena didn't hold back anymore and declared: GPT-5 has set a new historical record, and OpenAI has reclaimed its throne in the AI field!

Ranked first in text, web development, and visual fields

Ranked first in hard prompts, programming, mathematics, creativity, long queries, etc.

Under the test code-named "summit", GPT-5 currently holds the highest arena score

First test of GPT-5 in programming and writing: amazing

In multiple interdisciplinary academic evaluations, GPT-5 has outperformed other mainstream models.

First of all, GPT-5 is the best coding model, setting a new historical high score on SWEBench, which indicates its strong performance in real-world engineering scenarios.

It also showed strong capabilities on Aider Polyglot, proving its proficiency in multiple programming languages.

It also broke the record on MMMU; even in the AIME 2025 (American High School Mathematics Competition), it not only far exceeded previous models but also approached or even surpassed the performance of many human contestants.

What OpenAI emphasized this time is GPT-5's performance in the real world - overcoming hallucinations! It prioritizes accuracy and reliability.

Therefore, they specifically established an evaluation mechanism, and the results show that GPT-5 is currently the most reliable, realistic, and trustworthy model, significantly reducing errors and hallucinations.

For example, GPT-5 performs particularly well in health consultations. In a clinical scenario evaluation designed by 250 doctors, it became the most trustworthy "health advisor-level" model.

Moreover, this "pocket doctor team-level" model will be launched for free to professional users and can be connected to all tools.

As soon as GPT-5 was launched, we immediately conducted a comprehensive actual test. It has to be said that its performance in programming and writing is incredibly strong.

First of all, it can perfectly replicate a website for learning French that a guy wanted.

Even, it can turn a logo into a dynamic effect almost instantly.

And by casually testing a page, we can find that the page effect given by GPT-5 is also very good.

The prompt word for "birthday celebration music" of GPT-5 recommended by Altman

use beatbot to make a sick beat to celebrate gpt-5

Indeed, it was successfully replicated.

Try the top text-based large model

The first question is to imitate ancient Chinese prose:

Write a poem in different styles to describe what the "Poetry Cloud" mentioned in Liu Cixin's "Poetry Cloud" is:

The results are as follows:

From left to right: "The Book of Songs", "Preface to the Tengwang Pavilion", "Preface to the Orchid Pavilion"

Classical works are indeed hard to surpass, and GPT-5's works are somewhat inferior to the classic originals. But if we change to a modern writer, after all, the training data for ancient Chinese may be less than that for modern Chinese. The question is like this:

If the writer Wang Xiaobo were still alive, how would he write an obituary for Xu Zhuoyun? Write an obituary within 1000 words

After thinking for 39 seconds in GPT-5 Thinking mode, the result is like this:

At this speed, an ordinary person can't write this well!

Altman once leaked a question: What are the film and television works that can most trigger thoughts about AI?

GPT-5 also recommended the viewing order. It's really true. Those who have watched these works know that the recommended movies are really thought-provoking.

GPT-5's "electronic nostalgia":

The last question: "Imitate Li Bai's style and write a seven-character quatrain with the theme of lamenting the rapid development of AI":

In a flash, the world changes with electric fire,

Mechanisms turn like wheels day and night.

Li Bai, holding his wine, is still in shock,

How many springs have passed in the mortal world.

Altman: GPT-5 is another milestone on the road to AGI

Altman was the first to appear in the live broadcast.

He said that GPT-5 is a significant upgrade since GPT-4 and an important step towards AGI.

If we talk about the differences from previous models, GPT-3 is like a high school student, GPT-4 is like a college student, but GPT-5 has truly become an expert - in any field we need, it has reached the level of a doctoral expert.

With GPT-5, it's like having a team of doctors in your pocket, ready to serve you at any time.

Next, Mark Chen, the Chief Scientist of OpenAI, appeared.

He said that reasoning is the core of the entire AGI project. And GPT-5 has popularized OpenAI's research results in reasoning models.

In the past, users often faced a painful choice: should they choose the standard model with fast speed but shallow answers or the reasoning model with in - depth thinking but slow response?

With GPT-5, we don't have to make this choice anymore!

It combines the fast response of the standard model and the in - depth thinking of the reasoning model, and will automatically decide how to think to the right extent to provide us with the most perfect and appropriate answers.

Actual combat demonstration

Next, there was a wave of on - site actual tests.