12 of 12
AGI
AGI is a term that has been used by people for a machine that can do many kinds of intellectual work at the level of a person, without being rebuilt for each new job.
OpenAI's Codex can operate a computer in the background: it sees the screen, clicks, and types with its own cursor while you keep working. In September 2026 the company opened that loop to other developers, so a long job can run in a hosted sandbox, write files, and keep going after you step away.
The same month, Anthropic let its assistant open apps, click, and type in background windows, in Cowork and in its coding product. It can compile a program, launch it, click through the buttons, and take a screenshot, in the same sitting.
How close are we to AGI?
In January 2026, Nature published Humanity's Last Exam, put together by Dan Hendrycks at the Center for AI Safety and a team at Scale. About a thousand experts wrote 2,500 questions meant to stump a machine that has already read the internet. When the exam appeared in 2025, the best systems scored in the single digits. By September 2026 the best score on Scale's official board is about 55 percent, and some independent reruns sit a few points higher. Hendrycks has pointed to an earlier math contest he released in 2021, which went from under 10 percent to over 90 in three years.
François Chollet's ARC Prize asks whether a system can invent the rules of a situation it has not seen, the way a person sits down to an unfamiliar game. In March 2026 the third version put systems into tiny worlds with no instructions. People solved every environment. The best machines scored under 1 percent. By September 2026 one lab's system scored about 63 percent when it could only keep notes it chose to write down, and about 100 percent when the setup kept private notes between turns.
In 2025 a Berkeley group called METR began measuring how long a software job can be before an agent fails it, timing each job by how long a human expert would need. By September 2026 the jobs a system finishes half the time have stretched from minutes into hours, and some measurements now reach 16 hours. Finishing those jobs eight times out of ten still happens on much shorter tasks. METR says numbers above 16 hours are not trustworthy on the current tasks, which are clean programming problems.
Is that AGI?
That is not AGI. Dan Hendrycks's paper says a high score on Humanity's Last Exam is expert answers to closed questions. After a system reached about 100 percent on François Chollet's games, when the setup kept its notes between turns, the foundation that runs the prize said the jump was real and they were still not calling it AGI.