Es ist ein Massenleak von Informationen über GPT-5 aufgetreten. Alle vier Versionen sind öffentlich geworden. Auch kostenlose Nutzer können es nutzen, und es gibt noch eine geheime Funktion.

GPT-5 is finally coming.

Early this morning, OpenAI announced that it will hold a press conference at 1 am Beijing time tomorrow. Different from the previous short live broadcasts of about twenty minutes, Altman just previewed that the press conference will last for an hour and be full of valuable information.

Note that S has been replaced with 5.

On the eve of its official debut, a blogger revealed that Altman just announced that all employees will be given a two-year bonus totaling $1.5 million, even new employees who have just joined the company are treated equally.

In the past half day, leaked information about GPT-5 has gradually surfaced. For example, GitHub suddenly published a blog post with a straightforward title:

"GPT-5 is now generally available in GitHub Models".

This blog also revealed in advance that GPT-5 will have four versions:

gpt-5: The standard version, designed for logical reasoning and multi-step tasks;

gpt-5-mini: The lightweight version, suitable for cost-sensitive product scenarios;

gpt-5-nano: Emphasizes running speed and response time, targeting extremely low-latency requirements;

gpt-5-chat: Designed for complex, natural, multi-modal conversations with context understanding ability, suitable for enterprise-level scenarios;

Judging from the ability description, GPT-5 has been upgraded in terms of reasoning ability, code quality, and interaction understanding. It can complete complex programming tasks with fewer prompt words, and also adds more powerful Agent capabilities.

According to a revelation by X user @testingcatalog, access to GPT-5 will be opened in layers: Free users can use the basic version of GPT-5, Plus users will be connected to the GPT-5 model with stronger reasoning ability, while Pro users can access the GPT-5 Pro version that uses more resources and is designed for complex tasks.

The exposed code also adds multiple model tags, including GPT-5, GPT-5-Thinking, and GPT-5 Pro, which is highly consistent with the rumors of model ability layering and access permission differences.

In addition, the mysterious GPT-5 Pro is clearly marked as only available to Pro users and is defined as "research-level intelligence". It is the high-end version with the strongest performance and most comprehensive capabilities in this series. Such expressions have never appeared in OpenAI's previous product promotions.

The Team account has the most comprehensive usage permissions, covering all basic functions and advanced models.

In addition to the official website blog, community users have also started to share their actual experiences.

X user @hunoematic shared that he accessed GPT-5 through Copilot and exceeded humans in the SimpleBench reasoning test, achieving an accuracy rate of 90% (9 correct answers out of 10 questions).

However, to achieve this effect, you need to prompt the model to "reason" in advance to activate its more complex thinking. The following is the only question that GPT-5 answered incorrectly:

Facing complex problems, GPT-5 will automatically activate its reasoning ability.

Netizens also shared cases of using GPT-5 for programming. The good news is that the model performance has indeed improved to a certain extent, but the bad news is that if these cases were generated by the full-powered GPT-5, then such an improvement can only be described as mediocre.

Cloning GTA is the entry-level dream of game development. Although the graphics of this work by GPT-5 are rough, its structure is clear and the prototype of the gameplay is already in place.

A netizen asked it to design a web button with a "futuristic technology feel". After clicking, it will display a random AI-related fact. Although the interface is simple, the design style is very appropriate.

However, in another task of "creating a 3D flight simulator using Three.js", the UI performance is quite poor.

In more complex visualization tasks, GPT-5 has also produced some impressive works, such as a 3D interactive solar system model implemented with an HTML file (with functions such as animation, orbital operation, information click, and time control).

On the eve of the release, the X platform has been flooded with various news about GPT-5. As soon as a user posted a comparison chart of the parameters of GPT-4 and GPT-5, Altman started to be cryptic.

OpenAI's internal employees have also forwarded the official promotion one after another, adding fuel to the fire. In response, Miles Brundage, the former head of OpenAI's AGI Readiness team, couldn't help but post a complaint:

Dear OpenAI friends:

I'm sorry to be the one to tell you this, but as a former insider/now an outsider, I can confidently say that vague posts do make (most) outsiders uncomfortable. OpenAI can definitely choose a more clear and transparent way of communication - after all, you've done it before.

In fact, APPSO has previously compiled some core details of GPT-5. Welcome to review:

Unified model architecture: Integrating GPT and reasoning models: GPT-5 is no longer a single model, but integrates the GPT series (good at language) and the o series (good at reasoning), with the ability to automatically schedule different sub-models. It implements an "intelligent routing system" that dynamically calls the optimal model according to the task difficulty, ending users' choice phobia.

Complete multi-modal capabilities: Achieve true multi-modal integration, supporting the understanding and generation of information in forms such as images, text, audio, and video. The generated content is highly consistent visually and logically and rich in details.

Significantly improved software engineering capabilities: GPT-5 can deeply process old, large, and complex enterprise-level code libraries, not limited to solving algorithmic problems. Microsoft's internal tests show that while improving code quality, it does not significantly increase the consumption of computing resources.

Agent-driven upgrade: Better at handling multi-step complex tasks, with better task understanding and execution capabilities.

However, referring to a report by The Information, we may not have overly high expectations for GPT-5. A core issue is that as the Scaling Laws seem to have "hit a wall", the marginal benefits brought by pre-training are actually decreasing as the model gets larger.

In addition, the depletion of high-quality data has also brought the training resources of OpenAI to a bottleneck. The report mentioned that as of June this year, OpenAI's internal staff still had not confirmed which model version was worthy of being called "GPT-5".

In this wave of revelations related to GPT-5, another piece of news has attracted wide attention. Blogger @Yuchenj_UW revealed:

"My friends at OpenAI are all super excited now - not because GPT-5 is about to be released, but because Sam announced that each employee will be given a $1.5 million bonus over the next two years. 78% of NVIDIA employees are millionaires, while at OpenAI, this ratio is 100%."

He then added, "Considering that I didn't make it clear before: Every OpenAI employee, even new employees, will receive a $1.5 million bonus over the next two years."

This revelation is not unfounded. According to a Reuters report, OpenAI is preparing for a new round of financing, and its valuation is expected to reach $500 billion.

At the same time, OpenAI