What Is AGI, and Are We Actually Living in the AGI Era?
AGI has no agreed definition, which is why OpenAI, Nvidia and leading researchers can disagree about the same model in the same week. This explainer lays out the three main definitions, the measurements that exist (the Hendrycks AGI score and ARC-AGI-3), what GPT-6 Astra actually demonstrated in September 2026, and gives an honest, definition-by-definition answer to whether the AGI era has begun.
On 3 September 2026, OpenAI released GPT-6 Astra and told the world to welcome the "AGI era." Three days later, Nvidia CEO Jensen Huang posted that "AGI has arrived." Within hours, cognitive scientist Gary Marcus replied that the claim came with no evidence and no definition. All three statements were made about the same model, in the same week, by people who follow this field closely.
That disagreement is not a sign that someone is lying. It is a sign that the word AGI does not yet have a shared meaning. This article separates the definitions, looks at the measurements that exist, and tries to answer the question honestly.
What AGI means
AGI stands for artificial general intelligence. The everyday AI most people use is narrow: it is very good at one family of tasks, such as recognising faces, translating text, or recommending videos. AGI is the idea of a single system that can learn, reason, and apply knowledge across virtually any intellectual task a person can, including tasks it was never specifically trained on, at roughly human level or better.
Beyond AGI sits artificial superintelligence (ASI), a hypothetical system that would outperform the best humans across every domain by a wide margin. The two are often blurred in public discussion, but most researchers treat them as distinct thresholds.
That is the simple version. The difficulty is that "any intellectual task a person can do" is not something you can put on a test paper, and different groups have filled the gap with different definitions.
Three families of definitions
Cognitive definitions. AGI matches the mental versatility of a well-educated adult. This is the framing behind the October 2025 paper "A Definition of AGI" by Dan Hendrycks, Yoshua Bengio, Gary Marcus, Max Tegmark and dozens of co-authors. It breaks intelligence into ten domains drawn from the Cattell-Horn-Carroll model of human cognition, including knowledge, reasoning, working memory, long-term memory, visual and auditory processing, and speed.
Economic definitions. OpenAI's charter describes AGI as highly autonomous systems that outperform humans at most economically valuable work. Reports in 2024 suggested OpenAI and Microsoft had at one point tied the contractual meani ng of AGI to a system capable of generating $100 billion in profit. In September 2026, MiniMax co-founder Yeyi Yun suggested a different economic bar: AI generating 1% of global GDP on its own.
Efficiency definitions. The ARC Prize Foundation, which runs the ARC-AGI benchmarks, defines AGI as a system's ability to acquire any skill a human can, as efficiently as a human can. The emphasis is on learning new things quickly, not on how much a system already knows.
Google DeepMind added a useful ladder in its 2023 "Levels of AGI" paper: Emerging (roughly equal to an unskilled human), Competent (50th percentile of skilled adults), Expert (90th), Virtuoso (99th), and Superhuman. Under that scheme, large language models of the time were classed as Emerging AGI, which is the source of the "we already have a weak form of AGI" argument.
Notice that a system could pass one of these definitions and fail another. That is exactly what is happening now.
How researchers try to measure it
The AGI score
The Hendrycks-Bengio framework produces a percentage. Under it, GPT-4 (2023) scored about 27% and GPT-5 (2025) about 57%. The authors describe the profile of current models as "jagged": strong on knowledge, reading, writing and mathematics, with the deepest deficit in long-term memory storage, meaning the ability to learn from experience and retain it across sessions without retraining. A follow-up paper argued that a simple average hides this problem, and that a stricter aggregation puts GPT-5 closer to 24%.
No published score under this framework exists yet for GPT-6 Astra as of this article's date.
ARC-AGI-3
ARC-AGI-3, released in March 2026, is an interactive benchmark. The system is dropped into unfamiliar puzzle environments with no instructions and has to explore, work out the rules, set goals and plan. Ordinary people solve 100% of the environments; at launch, frontier models scored under 1%.
On 3 September 2026, ARC Prize published independent results for GPT-6 Astra. Using its standard, provider-neutral setup, Astra scored 62.7%. Using an OpenAI-specific harness that preserves the model's hidden reasoning state between steps, it scored 99.9%. Astra also needed fewer actions than the median human on 96% of levels, and it spontaneously built compact symbolic notations to track game rules.
ARC Prize called this a "step-function change" in capability and, in the same post, said plainly that it is not claiming Astra is AGI. The environments are closed, deterministic and narrow in format, and the foundation had stated before launch that saturating the benchmark would not count as proof of AGI. Benchmark author François Chollet repeated that point publicly the same week.
Two other details matter for reading the headline number. First, the 99.9% figure depends on the harness; the same model under the neutral setup scored 62.7%. Second, OpenAI's own launch chart compared Astra against competitors evaluated under different configurations, which analysts at VentureBeat and The New Stack flagged as an apples-to-oranges comparison.
What actually happened this week
OpenAI described GPT-6 Astra as its most intelligent and aligned model. In a press briefing, OpenAI President Greg Brockman said AGI remains a "gray, fuzzy thing," that it "might be about this model," and that it is "not unreasonable to feel that we are now in the AGI era." He closed with "Welcome to the AGI era." Asked whether the company was formally declaring AGI, he said the term is no longer tied to any contractual trigger and described it as a "mission concept or spiritual concept." So OpenAI's leadership has said it believes the era has begun, while stopping short of a formal declaration.
Jensen Huang's post went further, declaring AGI arrived and crediting the roughly 100,000 Nvidia systems used to train the model. He had made a similar claim on the Lex Fridman podcast in March 2026, though in that case in answer to a narrower question about whether AI could build a billion-dollar business. Nvidia sells the hardware that trains these models, which is context readers can weigh for themselves.
Gary Marcus, a long-standing critic of AGI claims and a co-author of the cognitive definition above, argued that by conventional definitions Astra still falls short, that a true AGI would be a quantum leap ahead of rivals rather than a step along an existing trend, and that declaring victory without a definition "simply muddies the waters."
Where the leading voices stand
Dario Amodei (Anthropic) has been among the most aggressive forecasters, saying at Davos in January 2026 that AI models would replace the work of all software developers within a year and reach Nobel-level scientific research in multiple fields within two. He generally prefers the phrase "powerful AI" to AGI.
Demis Hassabis (Google DeepMind) put the odds of AGI at roughly 50% by the end of this decade, and said at the same event that he does not expect it to arrive through models built exactly like today's systems.
Yann LeCun has argued for years that large language models are the wrong architecture for human-level intelligence because they lack a grounded model of the physical and social world. He does not say AGI is impossible; he says today's path will not get there.
Forecasting communities sit in between. As of mid-2026, the Metaculus community median for a "weakly general" AI was around 2028, and for a fully general AI system around 2033, with the exact figures moving month to month.
One pattern is worth noting: timelines from lab leaders have shortened sharply between 2023 and 2026, while the sceptics' core objection, that current systems cannot learn continuously from experience, has not been answered by any published evaluation.
So, are we in the AGI era?
It depends entirely on which definition you hold, so here is the honest answer under each one.
By the efficiency definition (ARC): closer than ever, but the benchmark's own authors say no. Astra clears the interactive test that no model could touch six months ago, yet ARC Prize describes the result as progress toward generalisation, not arrival.
By the cognitive definition (Hendrycks et al.): no published evidence yet. The last measured frontier model sat around 57%, with a near-total gap in long-term memory. Astra has not been scored, and nothing in its launch material addresses continual learning.
By economic definitions: no. Nobody, including the people proposing these bars, claims AI is generating 1% of world GDP autonomously or outperforming humans at most economically valuable work.
By the DeepMind ladder: arguably yes, at the lowest rungs. If Emerging or Competent AGI counts as "the AGI era," then the era began somewhere between 2023 and 2026. If you mean Expert-level generality across the board, it has not.
Even the one place where AGI had a legal price tag could not pin it down. The Microsoft-OpenAI partnership carried an AGI clause since 2019. In October 2025 the companies agreed any AGI declaration would need sign-off from an independent expert panel; in April 2026 they removed the commercial consequences tied to AGI altogether. Nobody ever ruled on whether the threshold was crossed.
The most defensible reading of the evidence is this: the systems available today are general in a way that was science fiction five years ago, they now beat ordinary humans on some tests designed specifically to expose the gap, and they still cannot do the thing that every serious definition of general intelligence requires, which is to keep learning after training ends. Whether you call that "the AGI era" is a choice about words. The capabilities themselves are not in dispute.
What this means in practice
For students, teachers and working professionals, the label matters less than the trajectory. Models that can enter an unfamiliar environment, work out its rules and act efficiently are already useful for research, coding, analysis and multi-step office work, regardless of whether they meet a philosophical bar. The skills that hold their value are the ones the benchmarks keep exposing as gaps: setting goals, judging what matters, and learning from experience over time. Those remain human strengths, and they are also the skills that make a person effective at directing these tools.
Treat confident AGI announcements, from any company, the way you would treat any unverified claim: ask which definition is being used and what independent evidence supports it. On 7 September 2026, the answer to the first question varies by speaker, and the answer to the second is "some, and not enough."