GPT-5.4 has been released, clarifying the direction of future AI iterations. The current AI field has evolved from dialogue boxes to system intelligent agents, with humans responsible for aesthetics and AI responsible for implementation, moving towards a human-machine collaborative workflow. ➤ GPT-5.4 Core Upgrade Points 1. Combining GPT-5.2's universal reasoning with GPT-5.3-Codex's top-notch programming capabilities 2. Supports 1 million token windows (approximately 5000 pages of documentation) and solves the pain point of forgetting long texts easily 3. Native computer operation, the model can directly view the screen, mouse, and keyboard like a human. In the OSWorld test, the success rate of 75.0% has exceeded the human average level 4. Introduce the function of interrupting midway. Dialogue is no longer a rigid turn based system, and users can insert new requirements at any time while the model is thinking or answering 5. Efficiency and cost optimization, introducing Tool Search mechanism. The model no longer needs to preload all tool definitions, but instead searches on demand, significantly saving 47% of token consumption. Why is this happening? Currently, top AI laboratories around the world are facing data walls. At most by 2026, the high-quality text, code, and books generated by all of humanity may be collected on a large scale by large models, and the training of text has reached a bottleneck period. Claude code, codex, and openclaw are deeply integrated with the current operating system, replacing some human operations and calling system tools, and have autonomous consciousness with the goal of completing tasks. Another thing that many people don't know is that the Codex series models are trained together with the Codex framework, which means that the Codex series models and the Codex framework are native to each other, and the models can naturally call all the development tools in Codex. ➤ In depth interpretation of the future development direction of AI 1. Shift from API stitching to operating system level native The Computer Use capability demonstrated by GPT-5.4 extends from dialog boxes to the entire operating system. Previously, the model was only coded in a restricted sandbox, and after upgrading, it will have physical hands. Not only do you understand code logic, but you can also comprehend visual feedback such as clicking, dragging, and terminal errors. The new framework layer will no longer be a bunch of preset utility functions, but a deep perception of the OS (operating system). The model learned how to observe the screen and provide feedback during training, which allows it to work like a senior engineer, modifying the code while viewing UI changes in the browser debugging window, achieving end-to-end development in a self looping manner. This has been implemented on Codex. 2. Million context+long-range task architecture design+memory system=versatile architect In the three-layer architecture of Codex, the model layer provides structured inference. The 1 million token context brought by GPT-5.4 essentially provides a broader canvas for this reasoning. OpenAI's memory system has always been far ahead, with the release of lossless memory and infinite memory. Especially when the model and framework are native to each other, the model can instantly retrieve the entire code repository (at the level of millions of tokens), and the framework can accurately apply modifications to dozens of associated files. Now in Codex, it is possible to achieve full architecture rewriting and accurately understand the meaning of the code. 3. Searchization and dynamic extension of tool calls The Tool Search mechanism introduced in GPT-5.4 enables the framework to understand the output pattern of the model, allowing the model to obtain more contextual information for precise operation. The future development direction is not to preload thousands of tool libraries (to avoid token waste), but to capture and load a data visualization component in real-time through Tool Search when the model infers that I need it. This means that the current Skills may be an intermediate transitional product, and more tools will be embedded in the model content. Large models can choose which tool to call on their own. The advantage is that it allows the large model to maintain extremely high token efficiency. It solves the paradox that the more tools there are, the dumber the model becomes, allowing the agent's skill tree to extend infinitely, automatically optimize, find the optimal path, and then train it into the next generation of models. 4. Real time interaction, from turn based to interrupt and modify at any time The interrupt function introduced in GPT-5.4 breaks the black box state generated by AI and adjusts the thinking in a timely manner. Introduce more human decision-making at the collaborative level, rather than allowing AI to operate autonomously, to achieve white box collaboration. Humans are responsible for important decisions such as aesthetics, requirement definition, and solution selection, while AI is responsible for implementation. Due to the introduction of real-time intervention capabilities, AI has transformed from a blind box state of delivering tasks at once to an engineering partner who can modify requirements at any time. A simple understanding of the new AI Native mode (Codex+GPT-5.4) is to build an F1 racing car directly from scratch, and the engine, chassis, and tires of the car are designed to work together for maximum speed from day one. In the future, we may no longer need to search for stronger models, but rather for systems that integrate more deeply with the development environment.