GPT-5.4 Details Emerge, Hinting at Permanent Memory and Enhanced Reasoning

OpenAI's upcoming GPT-5.4 model is reportedly undergoing testing, with early indications suggesting significant advancements in code generation, reasoning capabilities, and context window size. Foreign media reports and internal leaks point to features such as an "extreme reasoning model" and the potential for "permanent memory."

If these details prove accurate, GPT-5.4 could represent a substantial leap in AI model capabilities rather than a minor iteration.

Hand interacting with holographic code interface, symbolizing advanced AI code generation.

Early Testing and Code Generation

Evidence of GPT-5.4's development has surfaced through various channels, including OpenAI's own internal systems. Developer Corey Noles reportedly encountered a long model name, "gpt-5.4-ab-arm1-1020-1p-codexswic-ev3," in an error log while using OpenAI Codex, which suggests an experimental version of GPT-5.4 is actively being tested.

Further leaks from OpenAI's official Codex repository included Pull Requests (PRs) referencing "minimum model version = (5, 4)" and a "toggle Fast mode for GPT-5.4." These PRs were subsequently removed. Additionally, an OpenAI Codex team employee, Tibo, briefly posted a screenshot on social media showing GPT-5.4 in a model selector before deleting the post. Other developers have also reported seeing similar model strings in Codex error messages, reinforcing the idea that GPT-5.4 is deployed internally for A/B testing.

Testers have noted a significant increase in processing speed and the ability to generate over 6,000 lines of code from a single prompt, a task previously considered nearly impossible. The emergence of a "Fast Mode" feature suggests OpenAI may be exploring new inference architectures, such as multi-level latency pipelines or models with varying speed settings. Some users have also reported seeing "like or dislike" buttons next to Chain-of-Thought summaries, which could indicate their accounts are assigned to the GPT-5.4 test model.

Software engineers collaborating in a modern office, discussing complex code and data.

Extreme Reasoning and Extended Context

Reports from The Information align with these observations, highlighting an "Extreme Reasoning Mode" for GPT-5.4. This mode is designed to allocate more computational resources and time for deeper reasoning on complex problems, contrasting with traditional models that have limited thinking time. While reasoning capabilities may not be a primary concern for all general ChatGPT users or businesses seeking rapid answers, this feature could be highly beneficial for scientific research and enterprise applications requiring extensive problem-solving.

The context window for GPT-5.4 is also expected to increase significantly, from 400,000 tokens to one million tokens. This expansion would allow the model to process hundreds of thousands of words, analyze entire books, or handle large codebases and datasets. This development would bring OpenAI's models closer to the long-context capabilities already offered by competitors like Google's Gemini and Anthropic's Claude. Some rumors suggest the context window could even reach two million tokens.

Vast digital library with glowing data points, symbolizing an expanded AI context window.

Long-Duration Tasks and Persistent Memory

GPT-5.4 is also anticipated to perform better on tasks requiring several hours, indicating improved ability to retain user request details and adhere to instructions across multiple steps, reducing error rates. This capability would be particularly useful for OpenAI's Codex programming tool, which automates complex, long-duration tasks, and for AI agents that need to manage multi-step processes without constant human intervention.

Perhaps the most notable rumor surrounding GPT-5.4 is the potential for "permanent memory." An engineer's post on X, reposted by investor Garry Tan, described a "persistent state" for the model. This concept, reportedly discussed by figures like Jeff Dean, suggests major AI labs are exploring this direction. Speculation points to OpenAI potentially combining State Space Models (SSM) and Transformers on a large scale. SSMs continuously pass hidden states and exhibit linear computational complexity, unlike Transformers, which have quadratic complexity with context length.

A persistent state would enable AI models to retain information over extended periods, providing true long-term memory capabilities. This would mark a significant technological breakthrough if successfully implemented.

Stylized glowing circuit brain, symbolizing AI with permanent memory.

Accelerated Development Cycle

OpenAI appears to have adopted a more rapid update cycle for its models, with new versions emerging almost monthly since GPT-5. This accelerated pace, with GPT-5.1, GPT-5.2, and now GPT-5.4, suggests OpenAI is responding to competitive pressures. While ChatGPT's weekly active users have reached 910 million, the company has yet to achieve its goal of one billion. Competitors like Google and Anthropic are actively advancing their long-context, agent, and reasoning capabilities, prompting OpenAI to intensify its development efforts.

If GPT-5.4 delivers on its rumored features, including a one-million-token context, extreme reasoning, and persistent state, it could transform AI into continuously operating intelligent agents and represent a critical turning point in the history of large language models.