OpenAI and GPT-5.4

OpenAI continues to accelerate the pace of artificial intelligence development and has introduced a new model, GPT-5.4, with a context window of up to 1 million tokens and expanded capabilities for working with a computer. The update came unexpectedly quickly: the release took place just two days after the launch of version GPT-5.3 Instant.

The new model is already available in the ChatGPT service, through the API for developers, and is also integrated into the programming tool Codex. For users with a Pro subscription, the GPT-5.4 Pro version is available, designed for more complex tasks and intensive computations.

According to the developers, the main goal of the update is to increase the practical efficiency of the model in solving real work tasks: from document analysis and programming to preparing analytics and generating text.

Key technological changes

One of the most notable innovations is the support for a native computer-use mode. The model can now analyze interface screenshots, understand what is happening on the screen, and execute commands on a computer. In essence, this is a step toward full-fledged AI agents capable of working with software in the same way a human does.

Another important update is the significant expansion of the context window. As mentioned above, the model can process up to 1 million tokens, which allows it to retain enormous amounts of information within a single dialogue.

This is especially important for complex workflows such as analyzing large documents, multi-page reports, software code, or long research projects. The model can now “remember” much more data within a single session and use it when constructing responses.

According to OpenAI, the new version demonstrates noticeable performance growth in key tests. Particularly strong progress is observed in tasks related to agent-based scenarios – when the model must independently plan actions and solve complex multi-step tasks.

On the OSWorld-Verified benchmark, the model shows a result of about 75%, while the previous version reached approximately 47.3%. In the BrowseComp test, which evaluates the model’s ability to search and analyze information, the score increased to 82.7% compared to 65.8% previously. In the GDPval test, which evaluates the quality of logical analysis and reasoning, the result reached 83%, while the previous version achieved 70.9%. More modest growth is observed on the SWE-Bench Pro benchmark related to programming tasks. Here the new model shows 57.7% compared to 55.6% for the previous version. Despite the small difference, developers note that the model has become more stable in complex engineering tasks.

GPT-5.4 also introduces a new reasoning level called xhigh. This mode is designed for particularly complex tasks and uses significantly more computing resources.

In practice, the model spends more time analyzing the problem, builds deeper chains of reasoning, and checks multiple solution options before producing a final answer. Such an approach can be useful for scientific tasks, complex calculations, programming, or strategic analysis.

Developers also highlight improvements in model accuracy. According to OpenAI’s internal tests, the number of false statements has decreased by about 33%, while the overall number of responses containing factual errors has dropped by 18%.

This is an important step for corporate AI usage, as many companies view such models as tools for analytics, document processing, and decision support.

Another technical innovation is native support for context compaction. This is a mechanism that allows long dialogues and working sessions to be automatically compressed while preserving key information. As a result, the model does not “lose the thread” of the conversation even in very long chains of tasks. This function is particularly important for agent systems, where a single task may include dozens of sequential steps.

Changes also affect the interface of working with the model itself. In Thinking mode, the model can now display a plan for solving the task before executing the request. The user sees the proposed sequence of actions and can adjust the direction of the model’s work. This makes interaction more transparent and controllable: instead of waiting for the final answer, the user can intervene in the process and change the strategy.

What this means for users

The main focus of GPT-5.4 is practical application. The model has become better at handling everyday work tasks:

• document and report analysis
• information structuring
• article and content preparation
• programming
• creation of analytical materials
• research assistance

Thanks to the large context window and improved logic, responses become not only more accurate but also more useful for real work. This is especially important for fields such as marketing, analytics, software development, and media, where AI models are increasingly used as full-fledged working tools.

To obtain the most structured and useful result, users can apply a universal prompt formula. First, ask the model to briefly formulate the objective of the task. Then request a solution in the following format: short conclusion → step-by-step plan → example of application → possible errors and limitations.

If several solutions exist, it is useful to ask the model to compare them and identify the most effective one. This approach helps generate responses that can be immediately applied in practice.

The release of GPT-5.4 shows that the pace of artificial intelligence development continues to accelerate. Not long ago, model updates were released every few months, but now new versions appear literally within days.

For users, this means one thing: the capabilities of AI tools will grow faster and faster, and the boundary between a “chatbot” and a full-fledged digital assistant is gradually beginning to disappear.

A video fragment of the presentation of the new product can be watched in our Telegram channel.