OpenAI has taken a significant step in the evolution of artificial intelligence with the presentation of GPT-5.4, their first model designed to operate computers autonomouslyFar from being limited to generating text or code, this system is designed to operate within real-world applications and carry out complete processes almost as if it were a person in front of a keyboard and mouse.
The company led by Sam Altman positions this launch as their most advanced proposal for professional work environmentsWith a particular focus on workflow automation, scheduling, and lengthy tasks requiring phased planning, OpenAI, without much fanfare, proposes a shift in the role of AI with GPT-5.4: from conversational assistant to digital operator capable of directly intervening in corporate software.
A model focused on controlling the computer
The main novelty of GPT-5.4 is its ability to interact with the operating system and applications without constant supervisionThe model can open programs, navigate websites, move through menus, and enter commands using actions equivalent to using a keyboard and mouse, allowing it to complete complex workflows within desktop environments.
This ability was previously limited to external integrations or specific scripts, but with GPT-5.4 Computer control becomes a native function of the modelIn this way, AI agents can execute tasks from start to finish in office tools, development platforms, business applications, or web dashboards, without the user having to intervene at each step.
To make this viable, the system incorporates a context window that supports up to one million tokensIn practice, this allows you to handle extensive documentation, multi-step processes, and long-term projects without losing track between one phase and the next, something especially relevant in technical or administrative fields where large amounts of data are handled.
In addition, OpenAI has added a feature to automatic tool search This helps the model determine which application or resource it needs at any given time. In this way, the agent can decide for itself whether to open a code editor, a spreadsheet, a project manager, or a browser, and chain actions across different platforms to complete the same task.
This approach is aligned with the company's strategy to promote what are known as AI agentssystems that not only answer questions, but also act as a kind of "digital employee" within an organization's software infrastructure.
Enhanced reasoning and greater programming ability
Alongside the computer's control capabilities, OpenAI has introduced substantial improvements in reasoning and programmingAccording to the company, GPT-5.4 outperforms its predecessor, which specialized in coding. GPT-5.3-Code, both in accuracy and performance in development environments and professional tools.
This translates into greater solvency for write, review and debug code in different languages, as well as for integration into continuous integration workflows, test automation, or repository analysis. The model is designed to operate in demanding scenarios, such as software engineering projects, advanced data analysis, or script automation in enterprise systems.
Alongside the standard version, OpenAI has released GPT-5.4 ProThis version is designed for those who need extra capacity and stability for intensive workloads. Available both in ChatGPT and via API, it targets sectors such as finance, engineering, technology consulting, and business analytics, where consistent responses and the handling of large volumes of information are essential.
In practice, GPT-5.4 Pro offers better performance in complex tasks, with more headroom for lengthy processes and more predictable performance in operations that demand high levels of computation or context management.
Thinking function: view the AI ​​plan and change it on the fly
One of the most striking elements of this generation is the integration of the function ThinkingThis feature, inherited from GPT-4.5 but extended to work with GPT-5.4, allows the ChatGPT interface to... visualize the reasoning scheme in advance that the model intends to follow to solve a task, especially useful in long or delicate processes.
Instead of simply receiving a final result, the user can see how the AI ​​breaks down the problem, what steps it plans to take, and what tools it plans to use at each stage. In this way, it is possible stop execution mid-flow, correct instructions, refine objectives, or change priorities before the system completes the entire process.
This ability to "open up" reasoning is especially relevant in complex consultations or technical investigationsFor example, in advanced searches, analysis of large databases, or reviews of legal and scientific documentation, the user gains flexibility to tailor the results to what they actually need, without having to repeat the request from scratch.
OpenAI also notes that Thinking mode incorporates improvements in tasks of in-depth research and investigationMaintaining context across different related queries helps build more continuous workflows, where AI doesn't lose track of what has been done in previous steps.
More efficient, fewer errors, and better navigation than the human average.
Another point that OpenAI emphasizes is the efficiency of GPT-5.4 when consuming tokensThe model, according to the company, needs less textual context to arrive at a valid solution compared to previous versions such as GPT-5.2, which translates into a tighter use of resources, something relevant for companies that depend on the API on a large scale.
In terms of response quality, OpenAI states that GPT-5.4 is 33% less likely to generate incorrect statements and that complete responses show 18% fewer errors compared to the previous generation. Although these figures come from internal testing, they point to a significant reduction in hallucinations and inaccuracies, a key aspect in professional use.
The company has also highlighted that GPT-5.4 has surpassed average human performance for the first time in desktop navigation tasks. In the OSWorld benchmark, the model achieved a 75% success rate when performing actions in operating system environments, above the 72,4% assigned to the human average and well ahead of the 47,3% recorded by GPT-5.2.
These types of tests measure the AI's ability to handle windows, search for options in menus, complete forms, open and close applications, or make changes to system settings, among other common actions in the day-to-day use of a computer.
From the perspective of organizations, these advances represent greater reliability when delegating critical tasks in AI agents, both in programming and in spreadsheet manipulation, document management or systems administration, reducing the risk of costly errors.
Plans, variations and access in Europe
GPT-5.4 is integrated into OpenAI's product offering as its reference model for professional workIt is available to those with a ChatGPT subscription on the Plus, Pro and Team plans, opening the door to its use by freelancers, SMEs and work teams distributed in Spain and the rest of Europe.
Furthermore, both GPT-5.4 and GPT-5.4 Pro can be used via the OpenAI API for direct integrations into applications and servicesThis allows European companies to incorporate agents capable of controlling computers, automating internal workflows, or assisting employees within their own business tools, without needing to develop a model from scratch.
The company positions this launch as the centerpiece of its strategy for automate complex work processesFrom software project management to financial operations, systems administration, and report generation, the combination of desktop control, a large context window, and improved reasoning aims to cover a wide range of use cases across the European business landscape.
With GPT-5.4, OpenAI consolidates the idea of ​​agents that not only respond to what they are asked, but also take the initiative Within the digital environment, they coordinate with multiple tools and complete tasks from start to finish. For organizations and professionals who work daily with computers, this means the possibility of delegating a large part of the repetitive or low-complexity tasks to AI, while maintaining control over key decisions.
Although there is still a way to go to see how these types of systems are regulated and integrated across all sectors, the new model marks a turning point: artificial intelligence is no longer just a content generator and is establishing itself as autonomous operator within the desktop, with direct implications for the way companies and professionals work in Spain and the rest of Europe.
