OpenAI Rolls Out GPT-5.4, First General-Purpose LLM That Can Operate a Personal Computer

When VisiCalc debuted on the Apple II in 1979 it turned computers from curiosities into office necessities; GPT-5.4’s ability to click, type and browse without human hands echoes that inflection point, but for cognitive labor. The launch sits atop decades of automation pushing upward in the task stack—from 1960s numerical control of machine tools, to 1990s ERP systems that automated clerical workflows, to 2010s RPA bots mimicking GUI actions. What’s new is that a single statistical model, pretrained on internet text, is now steering arbitrary software in real time, collapsing the historical divide between “logic hard-coded by engineers” and “pattern recognition learned from data.” Over a 100-year horizon this could shift computers from tools we operate to semi-autonomous colleagues that operate other tools, amplifying both productivity and the surface area for error or misuse—much as the internal-combustion tractor of the 1920s multiplied farm output but also reshaped labor markets and ecosystems. Whether GPT-5.4’s agentic turn proves a transient gimmick or the Gutenberg moment for software agents will depend on societal governance, cost curves, and the still-unsolved problem of aligning open-ended machine goals with human intent.