Recently, an engineer from OpenAI submitted a code pull request in the public GitHub repository of Codex, inadvertently including the yet-to-be-announced model &#34;gpt-5.4&#34; in the version judgment condition.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d6beccc6-2a66-4c68-a6be-ed22171fb6dd.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d6beccc6-2a66-4c68-a6be-ed22171fb6dd.jpeg"/>
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2f8845e0-c468-4d6a-802d-4ed81a8b0189.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/2f8845e0-c468-4d6a-802d-4ed81a8b0189.jpeg"/>
Almost simultaneously, public model endpoints labeled &#34;alpha-gpt-5.4&#34; and screenshots of dropdown menus went viral on social platform X.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/0da00c7d-8dd0-4397-95c1-13824dd5277b.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/0da00c7d-8dd0-4397-95c1-13824dd5277b.jpeg"/>
What happened next was extremely dramatic; this information seemed to trigger some internal alarm, leading to the rapid deletion of the original post, and the involved code was forcibly pushed to be overwritten, quietly changed to &#34;gpt-5.3-codex.&#34;
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d221eacb-28a3-4424-98da-199a84b5bcaf.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/d221eacb-28a3-4424-98da-199a84b5bcaf.jpeg"/>
This retraction operation, which seemed to be an attempt to cover up, actually dispelled external doubts about the &#34;misuse of placeholders,&#34; making the speculation about the early leak of the new version more credible.
Various signs indicate that OpenAI is preparing to skip version 5.3 and is brewing a surprise that could reset the industry landscape.
Rumors suggest that this generational leap could land as soon as next week.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/832ec8b3-8206-49d4-b783-a15358cb122f.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/832ec8b3-8206-49d4-b783-a15358cb122f.jpeg"/>
It aims to end the recent incremental updates in the large model field and directly throw a trump card at its competitors.
Combining the various intelligence that has surfaced, the core killer feature of this major version is already emerging.
It has abandoned the route of grappling with peers in conventional reasoning benchmark tests and shifted the main battlefield to memory and context architecture.
With a context window of up to 2 million tokens paired with true Stateful AI, the model is finally free from the &#34;goldfish-like memory.&#34;
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/66f2704c-3a16-47a1-98c7-4f6b8789ffac.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/66f2704c-3a16-47a1-98c7-4f6b8789ffac.jpeg"/>
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b0775688-20ca-4c20-abb9-2ead8ef8bd51.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/b0775688-20ca-4c20-abb9-2ead8ef8bd51.jpeg"/>
It can fully retain your workflow, development environment, and even tool invocation state across different sessions.
Workers no longer have to repeat lengthy project backgrounds like a parrot every time they start a new conversation.
The model will inherently possess this persistent cognitive continuity, truly integrating into the user's daily development rhythm.
The significant leap in visual capabilities is also exciting for developers.
The leaked information explicitly mentioned a feature switch specifically for &#34;gpt-5.4 and higher,&#34; which allows the model to bypass traditional image compression mechanisms and directly read full-resolution raw bytes.
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/0c4692d2-76f5-47e1-aff5-cedd2fccca04.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/0c4692d2-76f5-47e1-aff5-cedd2fccca04.jpeg"/>
This means that front-end engineers and designers can directly provide it with extremely detailed UI design diagrams or complex engineering schematics, completely bidding farewell to the past experience of AI making serious nonsense about blurry compressed packages, achieving pixel-level visual analysis.
While Gemini 3.1 Pro and Claude 4.6 are still competing for advantages in the decimal rankings, GPT-5.4's ambition lies in completing the transformation from &#34;chatbot&#34; to &#34;fully automated agent employee.&#34;
<img src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/06d7e74e-691b-4faf-b239-10e0047e4da4.jpeg?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" alt="Image" original-src="https://imageproxy.pbkrs.com/https://wpimg-wscn.awtmt.com/06d7e74e-691b-4faf-b239-10e0047e4da4.jpeg"/>
It can reliably execute multi-step complex tasks in the background, making those so-called advanced competitors look like high-end calculators with dialogue boxes.
Of course, this level of context and state retention has also directly ignited the &#34;memory war&#34; at the hardware level.
The explosive growth of massive KV caches poses extreme challenges for the allocation of high-bandwidth memory and SRAM, and the introduction of optical interconnect technology has shifted from theoretical concepts to real-world needs.
OpenAI has clearly prepared its underlying computing architecture to face this storm.
Source: New Intelligence

OpenAI

AIYY

OpenAI accidentally leaked GPT-5.4! The new version achieves cross-session persistent memory with a context window of 2 million tokens and supports full-resolution visual direct reading. GPT will evolve from a chat tool to a "fully automated agent," completely reshaping workflows and igniting a battle for underlying hardware memory

- An OpenAI engineer accidentally revealed the unannounced "gpt-5.4" version in a code pull request, coinciding with the spread of "alpha-gpt-5.4" images on social media.  
- Following this, the original post was deleted, and the code was overwritten, suggesting that OpenAI may be skipping version 5.3 to launch a groundbreaking update, possibly next week.  
- The new version aims to transition from conventional updates to a stateful AI model with a 2 million Token context window, enhancing user experience by preserving workflow continuity and enabling advanced visual capabilities.

GPT-5.4 is rumored to launch next week! 2 million context window + persistent state, saying goodbye to frequent forgetfulness