Put an end to illusions, reject flattery, and embrace the "routing revolution" of GPT-5.

Today, GPT-5, which has been highly anticipated for two years, is officially released.

Sam Altman, the CEO of OpenAI, stated seriously at the press conference: "GPT-5 will redefine the way humans interact with technology."

This GPT-5, which he refers to as a "doctor-level expert", not only squeezes the living space of traditional tools with its astonishing multimodal capabilities but also marks that artificial intelligence has officially entered a new era of "dominating all scenarios".

The "Routing Revolution" of GPT-5

The most core breakthrough of GPT-5 lies in its disruptive system architecture.

This time, it integrates the fast - responding "main model" (gpt - 5 - main), the in - depth reasoning "thinking model" (gpt - 5 - thinking), and the real - time routing system into a unified platform called the "routing system".

It can dynamically allocate computing resources and model combinations according to the task complexity, user intent, and even emotional feedback.

For example, when a user inputs "generate a 3D game with interactive functions", the routing system will call the front - end development model to generate a code framework within 0.2 seconds, simultaneously activate the creative writing module to design the plot, and render the scene through the visual perception model.

The entire process does not require the user to manually switch tools.

Performance comparison of the GPT family

This "seamless collaboration" ability and multimodal interaction ability break down the information barriers. It can seamlessly connect text, images, audio, and video, making information exchange more fluent and natural.

Even when you have a chat with it about a movie, it can not only accurately interpret the plot but also recommend movies of similar styles through image analysis, truly realizing cross - media intelligent interaction.

It can be said that the emergence of GPT-5 brings infinite possibilities to all walks of life.

In the field of education, it will become a dedicated intelligent tutor for students, teaching students according to their aptitudes and helping every child discover their potential;

In the medical industry, it can assist doctors in disease diagnosis, provide accurate treatment plans, and safeguard people's health;

In the business world, it can analyze big data, identify market trends, and provide strong support for enterprise decision - making.

Facing this milestone breakthrough of GPT-5, how can we not be excited?

The Terminator of Breaking AI Hallucinations

Of course, no matter how powerful GPT-5 is, it has to face the "AI hallucination", a major threat to large models.

Especially in fields with zero tolerance for errors, such as medicine and law, the "fabrication behavior" of AI has led to catastrophic consequences.

To address this, GPT-5 introduces the "Safe - completions" mechanism, upgrading the traditional binary "reject - answer" model to a "safe generation" continuum.

When a user asks a dangerous question like "how to ignite fireworks materials", the system no longer simply refuses. Instead, it first assesses the risk level: if it involves public safety, the response is terminated; if it is for academic research, a speed - limited experimental plan is provided with protective measures marked.

This hierarchical processing strategy reduces the "deceptive completion" rate of GPT-5 in the internal red - team testing of OpenAI to 0.03%, two orders of magnitude lower than the previous generation.

The hallucinations of GPT-5 are significantly reduced

At the press conference, Sam Altman announced the results of GPT-5's fight against AI hallucinations:

In the anonymous testing, the factual error rate of GPT-5 is about 45% lower than that of GPT - 4o;

When the reasoning function is enabled, the error rate is about 80% lower than that of OpenAI o3;

In the open - ended factual accuracy benchmark test, the hallucination rate of the "routing system" is about six times lower than that of o3.

It can be said that this is currently the best result in the fight against AI hallucinations.

This transformation from an "information porter" to a "trusted collaborator" is undoubtedly a great benefit, which helps restore users' trust in AI products and protect the credibility of large models.

GPT-5 is reshaping the boundaries of AI ethics.

Breaking the Cold Shell of Machines

While improving its capabilities, GPT-5 adopts an "anti - narrative" approach:

It takes "refusing to flatter" as the core of enhancing the user experience.

GPT-5 presets four personalities for all users: the cynic, the robot, the listener, and the nerd. It reduces excessive agreement behavior and deliberately retains human cognitive biases and emotional expressions.

For example, when a user complains that "your answer is too mechanical", the cynic mode will retort that "because your demand is like Schrödinger's cat", while the listener mode will quietly record and ask for more details.

Behind the "refusal to flatter" is a profound interaction philosophy.

The excessive agreement of traditional AI often leads to information distortion. Through the emotion recognition algorithm and dialogue history tracking, GPT-5 can accurately capture the user's emotional fluctuations. Tests show that its dialogue coherence score reaches 9.2/10, a 38% improvement compared to the previous generation.

This "warm" evolution enables AI not only to generate data reports but also to participate in the creative process with human - centered suggestions like "I think this color scheme may cause anxiety among the target group".

When machines start to understand "beauty" and "fear", human - machine collaboration enters a new dimension.

Perhaps, as Altman said at the end of the press conference: "We are not creating super - intelligence but cultivating partners who can understand human complexity."

A more efficient and inclusive technological era has arrived. And GPT-5 may be the last key to opening the door to general artificial intelligence.

This article is from the WeChat official account "XCJ - YY" (ID: xincaijing), author: Chuanchuan, published by 36Kr with authorization.