OpenAI unveils GPT-5.4, an AI model that can operate computers and software

The new upgrade brings forth six key improvements in areas such as coding, image perception and multimodal tasks, long-running task execution and multi-step agent workflows, token efficiency and end-to-end performance on tool-heavy workloads, agentic web search and multi-source synthesis, document-heavy workflows in analytics, etc.

While a persistent concern around AI models has been their propensity for hallucinations, with GPT-5.4, OpenAI claims that it is 33 per cent less likely to be false than earlier models. The model has been designed for professional work, as it demonstrates strong performance in tasks like legal analysis, financial modelling, slide decks, and coding. Besides, developers can build AI agents that can plan tasks, execute them, and adjust in case something goes wrong. 

This marks a new shift in AI models. The early ChatGPT answered questions; it was followed by the GPT-4 era, which made it capable of writing essays, code, and summaries. With GPT-5, AI models can now reason through problems, and now with GPT-5.4 they can actually perform tasks on a computer. 

When it comes to real-world use cases, GPT-5.4 can work inside tools like spreadsheets and documents, which are key to office productivity. Some of the tasks it can perform are analysing financial data in Excel, building dashboards automatically, and drafting reports from raw data. It is capable of processing huge contracts and legal documents. For software development, GPT-5.4 can write large codebases, debug programs and run software testing automatically, and it can even control browsers using automation tools. 

OpenAI’s latest model comes at a time when rivals like Anthropic are increasingly making headway in complex enterprise tasks. The Dario Amodei-led company had introduced Claude Opus 4.6 and Claude Sonnet 4.6 last month, and both models have been dubbed as faster and more efficient for everyday workloads. While both the upgrades introduced by Anthropic and OpenAI emphasise different strengths, they demonstrate the growing competition to build AI models that function as practical digital workers.