OpenAI has unveiled its latest artificial intelligence model, GPT-5.4. This marks the third major upgrade in the GPT-5 family. The new model brings improvements in reasoning, coding performance, and AI agent workflows. It also introduces something far more ambitious — computer-use capabilities.
Unlike GPT-5.3-Codex, which focused mainly on coding within the Codex platform, GPT-5.4 is designed as a full general-purpose model. It is being rolled out across ChatGPT as well as the OpenAI API. The company has also launched a more advanced version called GPT-5.4 Pro. The goal is clear. OpenAI is pushing its AI models closer to becoming autonomous digital assistants capable of performing real tasks on computers.
Availability And Access
More Read
GPT-5.4 is now available inside ChatGPT as “GPT-5.4 Thinking.” It replaces the previous GPT-5.2 Thinking model. At launch, access is limited to ChatGPT Plus, Team, and Pro subscribers. Developers can also use the model through OpenAI’s API and Codex platform.
The more advanced GPT-5.4 Pro model is currently available only to Pro and Enterprise subscribers. This tier is designed for users who require higher reasoning capabilities, stronger reliability, and deeper enterprise integration.
Improved Reasoning And Coding
OpenAI says GPT-5.4 improves reasoning accuracy and coding performance. The model integrates the coding strengths that were introduced with GPT-5.3-Codex. This means better code generation, debugging, and development assistance.
The model also performs better in general productivity tasks. It can work with spreadsheets, documents, and presentations more effectively. This is important for real-world workflows where AI is expected to assist with everyday office tasks.
The company says GPT-5.4 has been optimised to handle complex instructions more reliably. This includes tasks that require multiple steps or tool integrations.
Agentic Workflows And Computer-Use
The biggest highlight of GPT-5.4 is its agentic computer-use capability. In simple terms, the model can deploy AI agents that interact with a virtual computer environment. These agents can browse the web, operate software, and perform tasks that usually require human input.
GPT-5.4 can generate code that interacts with a computer system. It can issue keyboard commands, move a mouse cursor, and perform structured workflows.
For example, an AI agent could open a browser, search for information, fill forms, or place orders. These tasks are executed inside a controlled virtual machine environment. OpenAI describes this as a major step toward task-performing AI systems.
Performance Benchmarks
OpenAI has shared internal evaluation benchmarks to highlight GPT-5.4’s improvements.
On the OSWorld-Verified benchmark, which tests a model’s ability to operate a desktop interface through screenshots and keyboard or mouse commands, GPT-5.4 achieved a 75 percent success rate. This benchmark specifically measures how well an AI model navigates a computer environment.
The model also performed strongly on visual reasoning tasks. On the MMMU-Pro benchmark, GPT-5.4 recorded an 81.2 percent success rate. This benchmark measures how well AI systems understand visual information. These improvements are largely attributed to enhanced visual perception and reasoning capabilities.
Safety And Security Measures
OpenAI has also addressed safety concerns surrounding advanced AI capabilities. The company has classified GPT-5.4 as having “High cyber capability.” Because of this, the model is being deployed with a more advanced cyber safety framework.
The safety stack includes monitoring systems, trusted access controls, and asynchronous blocking. These measures are designed to prevent misuse of the technology. OpenAI has also introduced a new open-source evaluation method called Chain-of-Thought controllability. This test measures whether AI models attempt to hide or manipulate their reasoning processes to bypass monitoring systems.
According to OpenAI, these safeguards are necessary as AI systems become more capable and autonomous.
Conclusion
GPT-5.4 represents another step toward AI systems that can actively perform tasks rather than simply generate text.
The addition of agentic computer-use capabilities signals a shift in how AI may be used in the future. Instead of acting only as a conversational assistant, models like GPT-5.4 could eventually automate complex workflows across software and digital environments.
At the same time, OpenAI is clearly aware of the risks. That is why the company has expanded its safety systems and monitoring frameworks.
For developers, enterprises, and AI enthusiasts, GPT-5.4 could become one of the most significant updates in the GPT-5 family so far.
Follow Us: Facebook | X | Instagram | YouTube | Pinterest
