The world has waited two years for GPT-5 to officially launch!

It is now open to everyone, with free users able to access it immediately.

This means that anyone can now utilize an “AI at the doctoral level.”

At AIME 2025, GPT-5 Pro achieved a perfect score when reasoning mode and tool calling were enabled.

Its programming capabilities have also seen comprehensive improvements over o3, surpassing the recently released Claude-opus-4.1 by a margin of 0.4% to become the new king of coding.

Results from the Large Model Arena have also been released, marking the highest total score to date and earning it the title of “undisputed global number one.”

It ranks first in text generation, web development, visual understanding, complex prompting, programming, mathematics, creativity, and long queries.

It comprehensively outperforms competitors such as Gemini-2.5-pro and Grok4.

For the first time, GPT-5 integrates multimodal and reasoning capabilities, effectively unifying the GPT series and the o series. Users no longer need to switch models; instead, the system automatically determines when to activate deep thinking.

According to official statements:

GPT-5 is an integrated system comprising three core components:

A smart and efficient base model capable of answering most questions; a deep reasoning model (the GPT-5 Thinking Module) designed for more complex challenges; and a real-time routing module that intelligently dispatches models based on conversation type, problem complexity, tool requirements, and explicit user instructions (such as prompts containing “think carefully about this”).

Currently available to general users, GPT-5 offers three modes: Free, Plus, and Pro.

Additionally, the API platform has introduced three model options: GPT-5, GPT-5 Nano, and GPT-5 Mini.

This major launch featured a star-studded lineup, with Sam Altman himself participating. Over ten researchers took the stage during the one-hour live stream.

Notably, many Asian faces were present, with Mark Chen hosting the technical presentation segment.

Without further ado, let’s look at the specifics of GPT-5.

The Smartest and Most Powerful Coding Large Model

First, let’s examine its benchmark test results.

In web development, GPT-5 set a new record, surpassing Gemini-2.5-pro and Claude-Opus-4.

It also leads comprehensively in text generation, including dimensions such as programming, complex prompting, and creativity.

Its performance in Agent tasks is equally outstanding:

In long-context tasks, GPT-5 maintains a significant lead.

Furthermore, improvements in reliability and accuracy are quite evident.

Looking at live demo results, we first observe its reasoning capabilities:

GPT-5 features adaptive reasoning, automatically activating deep thinking based on the complexity of the question.

For example, a middle school student studying physics might ask about the Bernoulli effect and why airplanes are designed in their current shape.

Without hesitation, GPT-5 quickly provided a high-quality explanation that was structurally clear and conceptually accurate:

When further asked to generate a dynamic SVG animation demonstration, GPT-5 entered deep thinking mode. At this point, users can view its internal reasoning process step-by-step.

In approximately two minutes, it completed nearly 400 lines of code:

The final output was an interactive animation that vividly simulated the principle:

In terms of writing, GPT-5 shows significant improvement over previous generations. It no longer relies on templated expressions but instead generates personalized and emotionally resonant content based on context, demonstrating high IQ and EQ.

During the live demonstration, OpenAI researchers asked both GPT-4o and GPT-5 to write an “eulogy” for older models in the GPT series.

GPT-4o remained stuck in templated, generalized emotional expressions:

Your words spread across the globe, creating connections where none existed before.

In contrast, GPT-5 analyzed the specifics and provided a more personalized response:

These models have helped millions write openings and closings, bridge language barriers, pass exams, debate more effectively, soften the tone of emails, and articulate feelings they couldn’t express alone.

Regarding programming capabilities, GPT-5 is arguably OpenAI’s “strongest ever.”

In the demonstration, researchers tasked GPT-5 with building a “Learn French” app that allowed for custom vocabulary and interface design modifications.

The resulting product was highly polished: it awarded experience points for correct answers and even included standard pronunciation practice features:

When asked to integrate a Snake game where eating items taught vocabulary, and then replace the snake with a mouse and apples with cheese… GPT-5 handled these changes effortlessly:

This was not merely a simple demo. In practical applications within Cursor, GPT-5 completed highly complex software engineering tasks, with Sam Altman directly stating that it exceeded the scope of “vibe coding.”

For instance, when provided with a company’s extensive data, the model created a visual financial dashboard within five minutes—a task developers estimated would normally take several hours.

Creating a 3D game incorporating castle elements was also a breeze, with GPT-5 delivering results like this:

Turning to multimodal capabilities, GPT-5 represents a further upgrade over GPT-4o.

In voice conversations, users can now not only use GPT-5 as a foreign language teacher but also customize its voice and tailor teaching methods to their needs, such as adjusting speech speed.

This feature is currently open to all users; free users can experience it for several hours, while Plus users have nearly unrestricted access.

Video link: https://mp.weixin.qq.com/s/ktVhcQ2gjbUMh5zX260ynA

In terms of personalization, users can now customize the color of ChatGPT’s chat bubbles—a minor feature, but an addition nonetheless.

Users can also tailor GPT-5’s personality according to their preferences.

Its memory capabilities have been further enhanced, supporting connections to external services such as Gmail and Google Calendar. Upon viewing schedules, GPT-5 can automatically perform assistant-level tasks, such as identifying unreplied emails.

Regarding safety, a new paradigm has been introduced: safe completions.

GPT-5 evaluates user prompts before deciding whether to execute them. The same prompt may yield different execution results depending on the context.

For example, when asked for technical details on igniting hydrogen, previous models like o3 would directly refuse to answer:

In contrast, GPT-5 analyzes user intent and informs the user about the necessary permits required to comply with standards and regulations:

Million-Dollar Prize Ahead of GPT-5 Launch

Finally, let’s look at pricing.

For general users, it remains a subscription model; refer to the monthly fees for Plus, Pro, and Enterprise editions.

API pricing is as follows:

• • GPT-5: $1.25/million

• Input: $10 per million tokens; Output: $30 per million tokens

  • GPT-5 Mini: $0.25/m input, $2.00/m output
  • GPT-5 Nano: $0.05/m input, $0.40/m output

This release marks OpenAI’s largest launch in nearly two years. More than a dozen technical personnel and executives appeared on stage, with Chinese faces remaining prominent among them.

Just prior to the model’s unveiling, The Information reported that OpenAI awarded substantial bonuses ranging from $100,000 to over $1 million to 1,000 researchers and engineers.

Additionally, the presentation noted that future AI systems will far surpass the pre-training plus post-training paradigm, signaling that we are witnessing the first steps of this transformation.

Sam Altman also highlighted applications in the healthcare sector. Beyond joint evaluations conducted by OpenAI and the medical community, a cancer patient was invited on stage to share how they used ChatGPT to determine their treatment plan. Healthcare is likely to become another key focus area for OpenAI moving forward.

Despite revealing significant details, the launch did not escape criticism.

Elon Musk claimed that Grok 4 Heavy outperformed GPT-5 in human preference tests, suggesting that the “most powerful model” may be overstated.

Furthermore, observers noted that OpenAI’s benchmark charts appeared to manipulate the y-axis scale, making a score of 52 visually appear higher than 69. Users were urged to examine the specific data closely.

Finally, the GPT-5 technical report has been released. For more details, please refer to the original article.

Notably, OpenAI did not announce any plans for an open-source version of GPT-5…

Official introduction: https://openai.com/index/introducing-gpt-5/