Just now, there was an explosive spoiler of the core details of GPT-5.4. It may have permanent memory and extremely powerful reasoning capabilities.
[Introduction] GPT-5.4 is coming soon! The testing has already begun, and it can write 6,000 lines of code at once. Foreign media has reported that it has an extreme reasoning model, and there are even rumors that it has permanent memory.
If you think the AI circle has been a bit quiet and unexciting lately, it might just be the calm before the storm.
According to multiple sources, GPT-5.4 is on the verge of release!
The figure of GPT-5.4 has appeared on LMArena.
Just now, foreign media The Information has revealed many core secrets about GPT-5.4.
The news is explosive: it has a longer context, a more extreme reasoning model, and even, permanent memory might appear!
If the rumors are true, this generation of models is likely not just a simple upgrade, but a huge leap in capability form.
Early Testing of GPT-5.4, Code Explosively Exposed
Lately, GPT-5.4 has been leaking everywhere.
From Codex error logs, GitHub PRs, to an employee accidentally sending a screenshot, in just a few days, GPT-5.4 has been "exposed" at least three times.
Moreover, all these were accidentally leaked by OpenAI itself.
Initially, developer Corey Noles triggered a network security restriction when using OpenAI Codex. A very long model name appeared in the error log returned by the system:
- gpt-5.4-ab-arm1-1020-1p-codexswic-ev3
The most crucial part of this string of characters is the first three - gpt-5.4.
In short, this string of characters can be regarded as OpenAI's internal deployment ID, which translates to "a real, deployed, and currently tested experimental version of GPT-5.4."
Then, two Pull Requests appeared in OpenAI's official Codex repository:
One PR states:
- minimum model version = (5, 4)
The other PR is even more direct:
- toggle Fast mode for GPT-5.4
It seems that OpenAI has added a "Fast Mode" switch for GPT-5.4. A few hours later, the two PRs were force pushed and deleted.
Next, a more dramatic scene occurred: An employee Tibo of the OpenAI Codex team posted a screenshot on a social platform. In the picture, GPT-5.4 is clearly written in the model selector.
Soon, this post was deleted.
Subsequently, another developer reported seeing a similar model string in the Codex error message.
This further indicates that GPT-5.4 has been deployed on the internal server and is undergoing real A/B testing.
Generate 6,000 Lines of Code with One Prompt?
Moreover, some developers have already reported an obvious change - speed!
Some testers say that the model's reasoning is significantly faster, and it can generate longer code. It can even generate over 6,000 lines of code with one prompt!
This was almost impossible before.
Some people have also discovered a new feature - Fast Mode.
This might mean that OpenAI is trying out a new reasoning architecture, such as a multi-level latency pipeline or models with different speed levels.
In addition, new features of the model interface have also been discovered: Some users say that if they see a like or dislike button next to the reasoning summary (Chain-of-Thought summary), it might mean that their account has been assigned to the test model of GPT-5.4.
Extreme Reasoning Mode, Performance Skyrockets
The revelation from The Information perfectly matches the above information.
The most eye - catching part of this revelation is the Extreme Reasoning Mode.
Traditional models have limited thinking time, but this extreme reasoning mode takes the model's performance to the extreme. When encountering difficult problems, it can spend more time and allocate more computing resources for deeper reasoning.
Interestingly, according to a survey, many ordinary ChatGPT users are not very interested in the reasoning function.
From a business application perspective, this function is not very suitable either, because enterprises want AI to give an answer as soon as possible.
Therefore, OpenAI's continued emphasis on reasoning ability seems rather pure - minded.
However, this news is great for the scientific research field and some enterprise customers.
Obviously, in the scientific research field, many users are willing to let a model run for hours or even days on valuable research questions.
At the same time, some enterprises also need GPT-5.4's stronger reasoning ability and long - term task performance to build AI agents that can automate more complex business processes.
According to various revelations, this extreme reasoning mode will be extremely powerful, which has raised everyone's expectations.
Context Doubles Directly to 1M Tokens
For GPT-5.4, this is another very significant upgrade - the context window has increased from 400,000 tokens to one million tokens.
In this way, the context window of GPT-5.4 will be more than twice that of the current GPT-5.2. (Of course, some OpenAI models support one million tokens, such as GPT-4.1, but 5.2 does not.)
Now, GPT-5.4 can directly handle documents of hundreds of thousands of words, analyze an entire book, or process long code libraries and data.
This finally allows this model to catch up with Google's Gemini and Anthropic's Claude in terms of long - context ability.
After all, other models have long reached a context window of one million tokens, and now GPT-5.4 has finally made up for this shortcoming.
Rumors are even more exaggerated than The Information's report: the context has reached two million tokens!
GPT-5.4, Starting to Excel at "Long - Term Tasks"
Another point in the revelation is that GPT-5.4 performs better on tasks that last for hours.
That is to say, it can better remember the details of user requests and what it is allowed or not allowed to do in multiple steps, and it is also less likely to make mistakes.
Obviously, this will be extremely helpful for OpenAI's Codex programming tool, as Codex uses AI to automate complex long - term tasks.
In addition, this long - term task ability is also very crucial for AI Agents.
Agents can read requirements, search for information, write code, and fix bugs on their own, without requiring users to provide new prompts at every step.
Does GPT-5.4 Have Permanent Memory?
Next, the craziest rumor has emerged - GPT 5.4 might have permanent memory!
After an engineer made this revelation on X, it immediately caused a stir in the AI community. Silicon Valley investment tycoon and YC founder Garry Tan quickly reposted it.
In this post, the revealer introduced the "persistent state" of GPT-5.4 like this.
Jeff Dean mentioned this point when participating in the latentspacepod podcast. It can be seen that major AI labs are now thinking in this direction.
Some people speculate that OpenAI might have discovered how to effectively combine state - space models (SSM) and Transformers on a large scale.
The key is that the design of SSM is to continuously transfer the hidden state in each step of calculation, and its computational complexity grows linearly, unlike the quadratic growth of Transformers with the increase of context length.
This also echoes a rumor: it is said that GPT-5.4 might have a context window of two million tokens.
The so - called persistent state essentially means that an AI model will transform from a character like Guy Pearce in the movie Memento (who can only rely on short - term memory) to a character like Dustin Hoffman in Rain Man (who has stable and long - lasting memory) overnight.
In other words, this will endow the AI model with real long - term memory ability.
If it can really be achieved, it will be a huge technological breakthrough!
Under Great Pressure, OpenAI Forced to Update Monthly
Obviously, after GPT-5, OpenAI made an obvious change, and the models started to be updated monthly at a high frequency.
Since this year, we have seen GPT-5.1, GPT-5.2, and soon 5.4 will also arrive. The update rhythm is even approaching once a month.
Obviously, OpenAI has been cornered by its competitors.
Currently, the weekly active users of ChatGPT have reached 910 million. Although this number seems astonishing, it still hasn't reached OpenAI's goal of one billion WAU (weekly active users).
Competitors like Google and Anthropic are closely pressing, continuously increasing their efforts in long - context, agents, and reasoning ability.
If GPT-5.4 really has a long context of one million tokens, extreme reasoning, and a persistent state, the AI form might become a continuously working intelligent agent.
Moreover, if GPT-5.4 can really start "remembering things," this might be a crucial node in the history of large models.
The singularity has arrived, and we are accelerating rapidly. Are you ready?
Reference materials:
https://x.com/marmaduke091/status/2028604854143176958
https://www.theneurondaily.com/p/openai-leaked-gpt-5-4-three-times
https://x.com/daniel_mac8/status/2028849150368862416
https://www.theinformation.com/newsletters/ai-agenda/openais-next-ai-model-will-extreme-reasoning?rc=bfliih
This article is from the WeChat official account "New Intelligence Yuan", editor: Aeneas, published by 3