ChatGPT-5.4 can now control computers – what could happen?

OpenAI has introduced a new AI model with ChatGPT-5.4. It should not only significantly improve the chatbot in many areas, but also enable it to specifically control the mouse and keyboard. The feature seems like a big bet for the future, but it’s not quite fully developed yet. A commentary analysis.

What can ChatGPT-5.4 do?

• Strictly speaking, OpenAI even has GPT‑5.4 Pro and GPT‑5.4 Thinking two new AI models published. While the Thinking version is integrated directly into ChatGPT, the Pro version is aimed primarily at developers and companies via the programming interface (API). The promise: GPT-5.4 Thinking is primarily intended to provide better answers in chat through the model researched more specifically on the internet. GPT-5.4 Pro was designed for complex tasks and more performance. Among other things, the model is said to be better at creating tables, presentations and documents.
• The probably most important innovation is that ChatGPT can control computer interfaces for the first time with GPT-5.4. According to OpenAI, the model is able to execute mouse and keyboard commands and operate programs. In addition the AI ​​analyzes screenshots of the screen and then decides independently which actions are necessary. ChatGPT should, for example, be able to open and operate programs, fill out forms and merge data from multiple applications.
• Another innovation is clear Expansion of the context window. According to OpenAI, GPT-5.4 can process up to a million tokens at the same time. This should allow the AI ​​to analyze large documents, entire code projects or extensive data sets with a single query. With a success rate of 83 percent, AI dwarfs the competition (70 percent) in this area. The update is also intended to increase the accuracy of the answers. Individual false statements would occur around 33 percent less frequently, while complete answers should contain 18 percent fewer errors.

ChatGPT-5.4 is intended to control computers

With GPT‑5.4, OpenAI has delivered an update that actually seems to work better in many areas instead of causing new errors. The ability to control the mouse and keyboard theoretically opens up possibilities for AI that could dwarf the competition. This computer use function provides a a clear step towards autonomous agents but in which people still have to play an active role.

What is interesting is that the signature of the autonomous AI agent OpenClaw is already becoming visible. With the commitment of developer Peter Steinberger, OpenAI appears to have acquired the necessary expertise for complex and practical applications. The focus is again on coding, presentations and office tasks clearly addressed to what is currently probably the biggest competitor, Anthropic.

Despite the supposed progress, there are limitations. Because in areas such as health, memory or impossible tasks shows GPT‑5.4 weaknesses. Even OpenAI admits this bluntly. Although the AI ​​often provides very good individual answers, it occasionally drifts away from the topic.

This points to a familiar pattern that shows that progress in an area often regressions brings with it in other areas. Even though OpenAI has apparently made progress here, the company still doesn’t seem to have this problem completely under control.

Voices

• OpenAI boss Sam Altman in a post on
• A Reddit user is full of praise: “I am very happy with ChatGPT 5.4. Honestly, I haven’t experienced a version that I liked so much in terms of quality, consistency and natural interaction since version 4.0. ChatGPT 5.4 feels smoother, more stable and much better for everyday use. My main request is simple: Please don’t ruin what already works so well. Not every update has to replace the identity of what people already love. Sometimes it’s wisest to do that to preserve what works and build on it. Thank you for ChatGPT 5.4 and please keep this strong foundation.”
• Journalist David Gewirtz tested ChatGPT 5.4 for ZDNET. His conclusion: “Each answer I got was quite good in its own right. But in half of my tests, the AI didn’t answer the question asked. You can get good answers, but you have to correct the AI relentlessly to keep it on topic. This gets tiring over time. It could lead to misinterpretations. Because the answers are written so well and so confidently, it’s easy to get carried away by the AI, even if the answer doesn’t fit. Whenever I see results like this, I become increasingly worried about a world overrun by AI agents.”

There is a lot at stake for OpenAI

Only the coming weeks will show whether GPT‑5.4 remains more than just a nice upgrade across a broader user base – and beyond a controlled one Demo data and press promises. Only when thousands of real users test the model for hours will it become clear whether it can harmoniously combine security, creativity and usefulness.

There is a lot at stake for OpenAI. Ultimately, it is important to regain trust after it became known that its AI models were being made available to the US military and many users then turned their backs on the company. If we can once again let performance speak more than politics, GPT‑5.4 could be a direct hit. The computer use function represents a big bet on the future and is intended to allow ChatGPT to operate more autonomously.

But that also opens up questions. How far can and should autonomous interaction on the computer go without people losing control? How do the new skills affect everyday work? Data protection and ethical standards out of? And last but not least: Will the AI ​​master the balancing act between performance and loss of control or will OpenAI make ChatGPT even worse with such a function?

Also interesting: