OpenAI Unveils GPT-5.4, First Mainline Model with Native Computer Control and 1M-Token Context

Embedding an AI agent directly inside Excel in 2026 echoes the way VisiCalc (1979) and Lotus 1-2-3 (1983) transformed the PC from hobby hardware into an indispensable business tool; both moments shifted the locus of work from specialized operators to mass office adoption. GPT-5.4’s leap from text generation to autonomous software operation signals the maturing of a decades-long trend—from rule-based assistants like Microsoft’s 1997 Clippy, through narrow voice agents (Siri 2011), to today’s large-context, tool-using models—toward fully delegated digital labor. If the trajectory holds, the marginal cost of many white-collar tasks will approach the plummeting cost curve we saw in manufacturing after the introduction of industrial robots in the 1960s. Yet claims of “human-level” proficiency rest largely on benchmarks either designed or curated by OpenAI and its partners, inviting the same scrutiny that followed IBM’s 1997 Deep Blue chess win or Google’s 2016 AlphaGo victory—impressive feats that masked remaining gaps in general intelligence. On a 100-year horizon, GPT-5.4 may be remembered less for its raw IQ points than for normalizing AI agents that can read, write, and now click—an inflection that could redraw boundaries between human and machine agency across the knowledge economy.