What AGI means

Few terms in technology carry as much weight with as little agreement. Artificial general intelligence, AGI for short, is the stated goal of the largest AI labs, the subject of a famous contract clause and the end point of many forecasts.

Here is what the term means, where the systems stand in autumn 2026, and what we think a company should do with it.

What people mean by AGI

Start with three plain terms. Narrow AI does one job, often better than any person: a chess engine, a spam filter, a model that predicts how proteins fold. General AI would do most of the thinking work people do, across fields, and learn new work as a person does. Superintelligence would outperform every person at nearly everything. Narrow AI is everywhere. The arguments are about the other two.

The most quoted definition is in OpenAI's charter of 2018: highly autonomous systems that outperform humans at most economically valuable work. It measures work, not minds.

In 2023, Meredith Ringel Morris and colleagues at Google DeepMind proposed levels instead of a single finish line, rating a system on how well it performs, from emerging to superhuman, and on how general it is. AlphaFold is superhuman, but narrow. They placed ChatGPT, Bard, Llama 2 and Gemini at the first level of general AI, equal to or somewhat better than an unskilled person. Autonomy, they noted, is a separate question.

In October 2025, Dan Hendrycks and 32 coauthors defined AGI as matching the cognitive versatility and proficiency of a well educated adult, measured across ten areas such as reasoning, memory and perception. GPT-4 scored 27 percent, GPT-5 57 percent. The profile was jagged: strong where knowledge helps, weak at basics such as storing what it learns over time.

The ARC Prize Foundation defines AGI as a system that can acquire any skill a human can, as efficiently as a human can, and builds its tests around that.

Where things stand in 2026

The sharpest movement this year came on the foundation's newest test. When ARC-AGI-3 launched in March, a set of interactive games without instructions, people solved them all and the best AI systems scored 0.51 percent. In September, OpenAI's GPT-6 Astra scored 62.7 percent in the standard setup, and 99.9 percent in a setup that keeps the model's hidden reasoning between steps. The foundation called it a step change, and added that it is not claiming the result is AGI: the test is tightly bounded and does not capture how open and complex the real world is.

Real work moves more slowly. METR times AI agents on software tasks against skilled people. In May it reported that the best public models succeeded half the time on tasks that take a person about 12 hours. To succeed four times in five, the tasks had to shrink to about an hour and a half. Long tasks can be attempted. Short ones are finished dependably.

The leaders of the field disagree on how close the rest is. In August, Sam Altman told TIME that OpenAI was not quite there yet, but would have an internal system he would call AGI by the end of the year. At Davos in January, Demis Hassabis of Google DeepMind said today's systems are nowhere near human level AGI and that one or two more breakthroughs may be needed. Yann LeCun, who left Meta to found AMI Labs, argued that models trained on text will never reach human level intelligence, because they lack a model of cause and effect in the real world.

Taken together, today's systems are superhuman at some tasks, expert at many and still weak at remembering, learning on the job and staying reliable over long stretches. Whether that counts as AGI depends on the definition you hold.

Why the word matters anyway

For years, the word had a price. Under the original partnership, Microsoft would lose its rights to OpenAI's technology once OpenAI reached AGI. In December 2024, The Information reported a financial test: AGI would count as reached when OpenAI's systems generate at least 100 billion dollars in profits. In October 2025, the two agreed that an independent panel of experts would have to verify any declaration, and extended Microsoft's rights to 2032, including models after AGI.

In April 2026 the clause went. OpenAI's revenue share payments to Microsoft now continue to 2030, subject to a total cap and independent of OpenAI's technology progress, and Microsoft's licence to 2032 is no longer exclusive. As The Verge put it, there is no panel any more, and no rule for what happens if or when AGI is declared.

The word also carries promises. OpenAI's charter commits the company to stop competing with, and start assisting, a value aligned and safety conscious project that comes close to building AGI first. A promise like that only works if someone can tell when close has been reached.

Lawmakers avoid the word. The EU AI Act regulates general purpose AI models, with duties for their providers since 2 August 2025, and presumes systemic risk for a model trained with more than 10 to the power of 25 floating point operations. The law counts what it can measure. And a forecast of AGI shapes investment, hiring and the patience of boards long before anyone agrees on what has arrived.

What it means for a company

For a company, waiting for a declaration is the wrong plan. It may never come, and the contract shows it can stop mattering even to those who negotiated over it. The jagged profile is the useful part: the same system can be superhuman at one task and unreliable at the next. So the question is not when AGI arrives. It is which work can be handed over now, and on what evidence.

List the work your team repeats: the offer, the month end, the reply to a standard request, the first draft of a page. Ask three things of each. Can a machine do it well enough today? Can a person check the result in a few minutes? Can a mistake be undone at little cost? Start where all three answers are yes.

Then hand it over step by step. First the machine drafts and a person decides. When the drafts stop needing corrections over enough real work to count, the machine runs the routine and a person approves at fixed gates. Only then does it run on its own, inside written rules, with every step logged. Throughout, people set the direction and check the result. What changes is how much of the work in between the machine carries.

None of this waits for a definition. Even a system that passed every test for AGI would need a company's knowledge, its rules and a person who answers for the result. Those can be built now, whatever the next model is called.

Where crue stands

Our line reads: Human led. AI powered. Autonomous next. It describes a path, not a forecast about AGI.

Today every crue product is human led. People bring the idea, the experience and the taste, AI does the production, and people review the result and give it the final touch. Next, the products become agentic: agents do most of the work, and people set the goals and approve at the gates. After that comes autonomous, one routine at a time, as each one earns it.

We do not build AGI, and we do not wait for it. Better models bring each stage closer, but the stages do not depend on what a lab calls its system. Helm, our vision of a company run by its founder with agents, is where we take the path furthest. Twin gives a company its AI brain, and Eigna a private model of its own.

If a system one day meets every definition of AGI, a company will face the questions it faces today: which work to hand over, who sets the direction and who answers for the result. The word will keep moving. Someone still has to stand at the helm.