Who Gets to Declare AGI? Anatomy of the Definition War Moving Trillions

THE THESIS. On September 3, 2026, OpenAI launched GPT-6 Astra, its new reasoning and agent model. Three days later, Nvidia CEO Jensen Huang posted on X that 'AGI has arrived.' OpenAI president Greg Brockman said he 'personally' believes we are in the AGI era, though he left the verdict to users. The problem is simple to state: there is no accepted definition of AGI, artificial general intelligence. Our thesis is that AGI does not work like a scientific milestone reached on a given day. It works like a label under negotiation. Three families of definitions are competing — cognitive, economic, and contractual — and each serves the interests of whoever defends it. Whoever controls the definition controls billion-dollar contracts, regulatory thresholds, and the stock-market narrative. That is why this war of words moves trillions.

THREE DEFINITIONS, THREE INTERESTS. The first family is cognitive: AGI as a machine matching human mental versatility. Google DeepMind proposed its 'Levels of AGI' in 2023, separating performance, generality, and autonomy. Even more useful is the work led by Dan Hendrycks ('A Definition of AGI,' 2025): it defines AGI as matching a well-educated adult across ten cognitive domains, and the measurement is revealing — GPT-4 scores 27% and GPT-5 scores 58%, with long-term memory near 0%, according to the paper itself. This definition benefits researchers: it makes progress measurable and the gaps visible. The second family is economic: OpenAI's charter defines AGI as 'highly autonomous systems that outperform humans at most economically valuable work.' It benefits the company: it is elastic, and the one announcing it certifies it. The third is contractual: as The Information revealed in December 2024, the OpenAI-Microsoft agreement defined AGI as the moment OpenAI generates roughly $100 billion in profits for its earliest investors. It benefits lawyers and investors: it turns a scientific question into a payment trigger.

THE NUMBER THAT MATTERS. OpenAI presented Astra with a 99.9% score on ARC-AGI-3, a test of novel puzzles designed to resist memorization and measure genuine reasoning. ARC Prize, the organization behind the test, re-ran the evaluation and found something uncomfortable: the same model scores 62.7% in its standard harness and 99.9% in OpenAI's adapter. A harness is the software wrapped around a model during the exam: which tools it can use, what it remembers between attempts, how it manages context. The most telling detail: with reasoning switched off, OpenAI's adapter still scores 96.7% — 34 points above the same model reasoning at maximum effort inside the standard harness, according to ARC Prize. Co-founder Mike Knoop was blunt: 'we lack evidence to call this AGI yet.' There is more: Fortune documented that OpenAI altered five metrics after initial publication (the ARC figure went from 98.6% in a draft to the published 99.9%; the hallucination rate dropped from 4.2% to 2% and back to 4.2%), pulling and republishing the announcement without explanation. Our reading: when the number decides the narrative, the number becomes an engineering target. ARC Prize will now publish both results side by side. That transparency is worth more than any percentage.

MOVING GOALPOSTS. We have been here before. In March 2025, a UC San Diego study showed GPT-4.5 was judged human 73% of the time in a three-party Turing test — more often than the actual humans. The Turing test, for 70 years the symbol of machine intelligence, was passed. And nothing happened: no panic, no AGI proclamation. We simply decided it did not measure what we thought. That pattern — cross the threshold, then redefine it — repeats with every benchmark, every standardized AI exam: it saturates (every model passes) or gets contaminated (answers leak into training data). Moving the goalposts is partly healthy, because definitions improve; and partly cynical, because each side moves them wherever suits its story. That is why we prefer continuous measures over binary verdicts. The best available comes from METR: the length of tasks an agent completes alone at a 50% success rate has doubled every 7 months since 2019, and every 4 months in 2024-2025, per its data. That curve does not say 'AGI yes or no.' It says how much autonomous work you can delegate today. Which is the useful question.

WHY THE LABEL MOVES TRILLIONS. The word AGI is a legal and financial trigger, not just an academic debate. Exhibit one: the contract. The OpenAI-Microsoft AGI clause — which by October 2025 was to be judged by an independent expert panel — died on April 27, 2026, in the deal's renegotiation, as Simon Willison documents: Microsoft now holds a non-exclusive license through 2032 and payments continue 'regardless of OpenAI's technological progress.' Translation: when real money was at stake, both companies removed the word AGI from the contract. Exhibit two: regulation. Europe's AI Act avoids the word and uses a training-compute threshold (10^25 operations) to presume systemic risk; export controls, as we analyzed in the Mythos 6 case, turn a model declared 'strategic' into a geopolitical asset. Exhibit three: the market. Huang declared AGI the same week Nvidia announced 400,000 additional GPUs, with the stock up 23.7% for the year and an average price target of $325, according to Seeking Alpha. We are not alleging bad faith; we are pointing at the incentive: whoever proclaims AGI sells the chips that train it. Every capability claim should be weighted by the balance sheet of whoever makes it.

OUR READING AND THE IMPLICATIONS. AGI will not arrive on a date; it is a gradient we will recognize in hindsight, the way we recognized the internet. In the short term, the definition war has real costs: contracts broken or armored, regulation that answers panic rather than demonstrated capability, narrative bubbles, and workers who read 'AGI has arrived' and fear for jobs that will not change this year. In the long term, the direction matters more than the label: systems doubling their autonomous work horizon every few months point toward accelerated science, medicine, and abundance — whether we call them AGI or not. Four practical implications. One: demand that every result be published with its harness, as ARC Prize will now do. Two: prefer continuous measures (METR's time horizon, Hendrycks' cognitive profile) over binary proclamations. Three: regulate demonstrated capabilities, not words — the dead clause proves that even the best lawyers could not define AGI in writing. And four: discount every announcement by the announcer's incentive. 'Have we reached AGI?' is the wrong question. The right one: what can this machine do today without supervision — and who answers when it fails?