Key takeaways
- OpenAI's charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work, while Google DeepMind grades systems on five levels of performance and generality.
- ARC Prize defines AGI as matching human learning efficiency; top models now score 93% to 95% on ARC-AGI-2, and the best ARC-AGI-3 score is 81.2%.
- Anthropic said in March 2025 it expects powerful AI systems in late 2026 or early 2027, and DeepMind said in April 2025 that AGI could arrive within the coming years.
AGI, short for artificial general intelligence, means an AI system that can do most of the thinking work people do, at least as well as people do it. That is the core idea. What it doesn't come with is one agreed definition. OpenAI's charter ties it to economically valuable work, Google DeepMind splits it into levels, and the ARC Prize defines it by how fast a system learns.
The disagreement matters for a practical reason. Whether "AGI has arrived" depends entirely on which definition you use, and the labs that use the word most also have a stake in when it gets declared.
What AGI means: four definitions
OpenAI. The company's charter, published in 2018, states its mission as making sure AGI benefits all of humanity, and defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." The definition has contract weight. Under the Microsoft and OpenAI agreement of October 28, 2025, once OpenAI declares AGI, "that declaration will now be verified by an independent expert panel."
Google DeepMind. In the paper "Levels of AGI for Operationalizing Progress on the Path to AGI" (Morris, Sohl-Dickstein, Legg and others, first posted November 2023, presented at ICML 2024), DeepMind researchers describe AGI as "an AI system that is at least as capable as a human at most tasks." Their argument is that AGI is a path, not a single finish line, so they grade systems on two axes: how well they perform (depth) and on how wide a range of tasks (breadth). DeepMind's April 2, 2025 post on AGI safety uses the same short definition: "AI that's at least as capable as humans at most cognitive tasks."
Anthropic. Anthropic mostly avoids the term. CEO Dario Amodei wrote in his October 2024 essay Machines of Loving Grace that he finds AGI "an imprecise term that has gathered a lot of sci-fi baggage and hype." He uses "powerful AI" instead: a model "smarter than a Nobel Prize winner across most relevant fields" that can use a computer like a remote worker, run on its own for days, and work as many copies at once. He sums it up as a "country of geniuses in a datacenter."
ARC Prize. The nonprofit behind the ARC-AGI benchmark, built on François Chollet's 2019 paper "On the Measure of Intelligence," says plainly: "AGI is a system that can match the learning efficiency of humans." It rejects the economic definition outright, calling automation of most valuable work "an incorrect measure of intelligence." For ARC, knowing a lot is not the point. Picking up a new skill from a few examples, as a person does, is.
The DeepMind levels of AGI
The DeepMind paper's table is the most widely cited way to place a system on the map. Each level sets the minimum performance on most tasks, compared with skilled adults.
| Level | Performance | Narrow example in the paper | General (AGI) status in the paper |
|---|---|---|---|
| 1: Emerging | Equal to or somewhat better than an unskilled human | Simple rule-based systems | ChatGPT, Bard, Llama 2, Gemini |
| 2: Competent | At least the 50th percentile of skilled adults | Siri, Alexa | Not yet achieved |
| 3: Expert | At least the 90th percentile | Grammarly, Imagen | Not yet achieved |
| 4: Exceptional | At least the 99th percentile | Deep Blue, AlphaGo | Not yet achieved |
| 5: Superhuman | Outperforms 100% of humans | AlphaFold, AlphaZero | Artificial superintelligence (ASI), not yet achieved |
The authors placed the frontier chatbots of September 2023 at "Emerging AGI" and said "Competent AGI," the level that best matches most older definitions, had not been reached by any public system. They also map other definitions onto the scale: OpenAI's labor-replacement threshold fits "Exceptional AGI," in their reading. The paper's latest version, from September 2025, renamed Level 4 from "Virtuoso" to "Exceptional" but did not publish a new placement of current models.
How anyone would tell AGI had arrived
There is no single test. Two public measurements get the most attention.
ARC-AGI. These are puzzles that are "easy for humans, hard for AI": a few example grids, a hidden rule, and a new grid to solve. On ARC-AGI-2, every task was solved by at least two people in a live study of more than 400 members of the public. The ARC Prize leaderboard lists its human panel at 100% on ARC-AGI-2. The leaderboard also shows that this test, which was hard for AI when it launched in 2025, is now close to saturated: GPT-6 Astra scores 95.0%, Gemini 4 Carbon 94.2% and Claude Fable 5.2 93.8%, all well past the 85% the prize set as its goal.
That is why the newer ARC-AGI-3 exists. It drops agents into interactive games with no instructions, where they have to explore, work out the goal and learn as they go. A score of 100% means an agent "can beat every game as efficiently as humans." The top listed score is Claude Fable 5.2 at 81.2%, followed by Gemini 4 Carbon at 65.1% and GPT-6 Astra at 62.7%. ARC Prize says results of its 2026 competition will be announced on December 4, 2026.
METR time horizons. The research group METR measures how long a task an AI agent can finish on its own, timed by how long it takes a skilled human, with a 50% success rate. In its original March 2025 study, METR found this horizon had doubled about every seven months. Its current results (version 1.1, last updated September 8, 2026) put GPT-2 at about 3 seconds, GPT-4 at about 4 minutes, Claude Opus 4.6 at about 12 hours, and an early version of Claude Mythos Preview at about 17 hours. METR warns that "measurements above 16 hrs are unreliable with our current task suite," so the top of the chart is now past what the test can measure well. Its data file puts the doubling time since 2023 at about 129 days.
Neither test is a finish line. ARC Prize does not say that a high score equals AGI, and METR measures software and research tasks, not all of human work. Both show the same pattern, though. ARC-AGI-2 went from hard for AI to nearly solved in under two years, and METR had to warn that its task suite tops out.
How AGI differs from today's models and from superintelligence
Today's large language models (systems trained to predict text, explained in what is an LLM) are already broad. They write, code and answer questions across most fields. What the definitions above ask for on top of that is reliability across most tasks, the ability to work for long stretches without a person stepping in, and, for ARC, learning new skills as fast as a human. The debate over whether more scale gets there is covered in System 1, System 2 and scaling to AGI.
Superintelligence, or ASI, is the step past that. In DeepMind's table it is Level 5 general AI: a system that outperforms every human across a wide range of tasks. AlphaFold is superhuman at one task, predicting protein structures, which makes it narrow. ASI would be superhuman at almost everything.
Where the main labs say they stand
Each of these is the lab's own claim, not a measurement.
- OpenAI. In his January 2025 post Reflections, CEO Sam Altman wrote: "We are now confident we know how to build AGI as we have traditionally understood it," and that OpenAI was turning its aim to superintelligence. In The Gentle Singularity he wrote that "2026 will likely see the arrival of systems that can figure out novel insights."
- Google DeepMind. Its April 2025 safety post says AGI "could be here within the coming years," and lays out how the lab plans to handle four risk areas: misuse, misalignment, accidents and structural risks.
- Anthropic. In its March 6, 2025 recommendations to the White House, Anthropic said it expects "powerful AI systems will emerge in late 2026 or early 2027." In his January 2026 essay The Adolescence of Technology, Amodei wrote that powerful AI "could be as little as 1–2 years away, although it could also be considerably further out."
None of these statements says AGI has been built. If OpenAI does declare it, its contract with Microsoft now sends that claim to an independent panel before it counts.
Frequently asked questions
Is AGI here?
No lab statement cited here says AGI has been built. Frontier models now score 93% to 95% on ARC-AGI-2, but the harder ARC-AGI-3 tops out at 81.2%, and under OpenAI's deal with Microsoft any AGI declaration must be verified by an independent expert panel.
What is ASI?
ASI, or artificial superintelligence, is AI that outperforms every human across a wide range of tasks. In Google DeepMind's Levels of AGI paper it is Level 5 general AI, the step beyond AGI.
What is the difference between AGI and AI?
AI covers any system that does a task that takes intelligence, including narrow ones such as a chess engine or a spell checker. AGI is AI that matches or beats humans across most cognitive tasks, not just one.
What does ARC-AGI measure?
ARC-AGI measures how well a system solves new puzzles from a few examples, tasks that are easy for people but hard for AI. ARC Prize treats that learning efficiency as the core of general intelligence.
Sources
- OpenAI CharterOpenAI
- Levels of AGI for Operationalizing Progress on the Path to AGIGoogle DeepMind (Morris et al.), arXiv
- Taking a responsible path to AGIGoogle DeepMind
- What is ARC-AGI?ARC Prize
- ARC-AGI-2ARC Prize
- ARC-AGI-3ARC Prize
- ARC-AGI leaderboardARC Prize
- Measuring AI ability to complete long tasksMETR
- Time horizonsMETR
- Machines of Loving GraceDario Amodei
- The Adolescence of TechnologyDario Amodei
- Anthropic's recommendations to OSTP for the U.S. AI Action PlanAnthropic
- ReflectionsSam Altman
- The Gentle SingularitySam Altman
- The next chapter of the Microsoft-OpenAI partnershipMicrosoft

