As reported by Cnbeta, on March 4, an engineer at OpenAI accidentally included the yet-to-be-released 'gpt-5.4' in a version check condition when submitting code to Codex's GitHub repository. The pertinent content was swiftly removed and revised. Despite this, two revolutionary features of GPT-5.4 have garnered considerable attention: firstly, it supports a context window of up to 2 million Tokens and introduces genuine stateful AI, which allows for the continuous preservation of workflow, development environment, and tool invocation states across various sessions; secondly, it facilitates the direct reading of full-resolution images, circumventing traditional compression techniques to achieve visual analysis at the pixel level. According to sources, the model may be officially launched as soon as next week.
