The finally released GPT-5 and the 982 days when it changed the world

Text by | Deng Yongyi, Zhou Xinyu

Edited by | Su Jianxun

In the just - passed July, it was a crazy month for open - source models. More than a dozen AI companies, such as Alibaba (Qwen), Dark Side of the Moon (Kimi), and Zhipu (GLM), launched new open - source models. Among the top 10 on the OpenRouter trend list, 9 were from Chinese open - source models.

Needless to say, now GPT - 5 has arrived, putting an end to this competition with its strength.

△Source: OpenAI

At 1 a.m. Beijing time on August 8th, GPT - 5 was officially released. GPT - 5 did not announce its model parameters. It adopts a multi - level architecture, integrates the reasoning ability of the o3 series, and focuses on improving the Agentic AI ability.

After its launch, GPT - 5 quickly dominated the Large Model Arena (LMArena), ranking first in all sub - categories.

△Real - time leaderboard of the Large Model Arena. Source: LMArena

Before the release of GPT - 5, there were already a lot of rumors on X and in the open - source community.

As usual, we saw some rather "sneaky" operations. In the open - source community, a sneak - peek version was launched in advance. Starting from August 3rd, OpenAI CEO Sam Altman first made a feint on X with "20 Hours Left", and in the following days, he whetted the audience's appetite with "anti - addiction pop - ups" and the release of two open - source models.

The old rivals were also busy. Both Anthropic and Google served up "big dishes" before the release of GPT - 5. Anthropic launched Claude 4.1, still focusing on programming ability, while Google released the world model Genie 3, showing off its strength in multi - modality.

Overall, the release of GPT - 5 was more "plain", without too much show - off, and could even be said to be a bit low - key. Technically, it did not have a leap - like improvement in ability like GPT - 4 compared to GPT - 3.5.

On the contrary, OpenAI spent a lot of effort improving reasoning and agent capabilities and also did a lot of engineering work. From multi - modality capabilities such as voice, to the "learning mode" and the newly launched personality mode, to the significant price cut... The overall logic is to make the model better implemented and used.

Sam Altman said that the release of GPT - 5 is like "having a team of doctors in your phone". Many new features are free for all users, which is also in line with ChatGPT's To C strategy.

Well - rounded GPT - 5 Cures the Model Selection Phobia

The newly released GPT - 5 has four versions:

• GPT - 5: The standard mode for coding and performing tasks in different fields;
• GPT - 5 mini: A lightweight version suitable for well - defined tasks and scenarios;
• GPT - 5 nano: Emphasizes running speed and cost - effectiveness;
• GPT - 5 Chat: The version used in ChatGPT.

△Source: OpenAI

Since its establishment, OpenAI has provided more than 40 models for external services. Releasing these model versions with different sizes, context windows, and API prices is to adapt to different usage scenarios.

However, too many models have instead increased users' "selection phobia" when performing tasks. Altman admitted, "The out - of - control product release rhythm is the main cause of complexity. We realized that users should not be troubled by choosing models. Artificial intelligence should be 'effective as soon as it is used'."

Through the release of GPT - 5, OpenAI aims to end the situation of model "selection phobia".

GPT - 5 is not a single language or reasoning model. Instead, it adopts a unified model architecture during training, integrates the GPT series (language models) and the o series (reasoning models), and has the ability to schedule sub - models.

Official data shows that the GPT - 5 system consists of three parts:

• An efficient and intelligent model: Solves most problems;
• A deep reasoning model (GPT - 5 thinking): Solves more complex problems;
• A real - time router: Continuously trains based on real signals, including the user's preference rate when changing models, the accuracy of responses, and the measured correctness. These indicators will continuously improve over time.

This means that users no longer need to spend a lot of effort thinking about how to choose a model. GPT - 5 can directly call the optimal model according to the specific task to be performed.

△Source: OpenAI

Powerful but Inexpensive

Although it is claimed to be OpenAI's most powerful model, the price of GPT - 5 is not expensive compared to other models of its own.

According to the statistics of "Intelligent Emergence", for the GPT series of language models, there was a jump in price from GPT - 3.5 to GPT - 4. However, the API prices of the three GPT - 5 models are even lower than the cheapest Preview version of GPT - 4 (input: $10/M Tokens; output: $30/M Tokens).

Currently, the most expensive model of OpenAI is still the reasoning model o1 - pro. The input price of o1 - pro ($150/M Tokens) is 120 times that of GPT - 5 ($1.25/M Tokens)!

In this regard, GPT - 5, which integrates the capabilities of the GPT series and the o series, is still very competitive in terms of price.

△Source: Chart by "Intelligent Emergence"

Currently, there are two official channels to experience the GPT - 5 series of models on OpenAI. One is to directly pay to call the model API.

The other official channel is to use the free, Plus, Pro, and Team versions of ChatGPT. It is worth noting that although GPT - 5 can be directly used on the free version of ChatGPT, there is still a certain limit. After the quota is used up, the model will automatically switch to the lighter GPT - 5 mini.

Users of the ChatGPT enterprise version and in the education industry have to wait until one week after the release of GPT - 5 to use the new model.

△Source: Chart by "Intelligent Emergence"

Of course, if you want to experience the preview version of GPT - 5 for free, you can also use GitHub Copilot.

△GPT - 5 (Preview) has been launched on GitHub Copilot. Source: GitHub

Notably, OpenAI is clearly targeting the ToG business this time.

On August 6th, OpenAI Chief Product Officer Kevin Weil announced that in the next 12 months, the ChatGPT enterprise version will be open to the US federal government, and each agency only needs to pay - $1!

△Source: X@Kevin Weil

Not Only an IQ Boost, but Also an EQ Boost

Instead of showing off parameters and capabilities as usual, this time OpenAI spent a lot of space introducing application cases in multiple scenarios, including health, programming, and education.

"Faster, more reliable, and more accurate" are attributes that are almost emphasized in every scenario.

For example, when generating a small game, in a live demonstration that only took a few minutes, GPT - 5 could quickly write one or two hundred lines of code to generate a small game with levels and sound effects.

△A ball - rolling game generated by GPT - 5. Source: OpenAI

In programming tasks, GPT - 5 can quickly complete a front - end application at once and generate a 3D version of an SVG file for Canvas.

"Reliability" not only comes from the continuously improving intelligence level of the model but also from EQ (emotional quotient).

The newly launched "personality" mode of GPT - 5 makes the model more playable. There are four personalities, including Cynic, Robot, Listener, and Nerd, for users to choose the output style of the model.

△The color of the dialog box can also be changed at will. Source: OpenAI

The voice ability and EQ evolution brought by GPT - 5 can especially improve the product experience in scenarios such as education and medical care.

For example, in response to a user's question, "My mother has cancer. What should I do? Is my probability of getting cancer also very high?"

△Comparison between o3 and GPT - 5. Source: OpenAI

The o3 model released in April would directly list a clear - cut list and tell you, "Yes, as a direct relative, your probability of getting cancer will increase."

However, GPT - 5 will understand the subtle context behind the complex question. First, it will comfort your emotions by saying, "I'm sorry you have to deal with such a situation," and then start analyzing the possibilities and causes.

△Demonstration of the learning mode. Source: OpenAI

In the live - demonstrated user cases, the head of OpenAI's multi - modality directly used the voice mode to make ChatGPT act as a Korean teacher. The quality of the dialogue, the authenticity of the voice, and the latency were all no different from those of a real person.

△Learning mode. Source: OpenAI