Close-up of a white humanoid robot head and shoulders against a plain background

Artificial General Intelligence (AGI): What It Is in 2026

Updated November 21, 2026.

Artificial general intelligence (AGI) is a hypothetical kind of AI that could learn and do almost any intellectual task a person can, instead of being built for one job. No one agrees on exactly where the line is, so no one can prove it has been crossed. In September 2026 OpenAI said its GPT-6 Astra model may mark the start of an "AGI era," but independent researchers and rival labs say today's systems still fall short.

This guide explains what AGI means in plain English, how it differs from narrow AI, strong AI and superintelligence, what the 2026 debate is really about, and how to judge the next "we've reached AGI" headline for yourself. If you are new to the basics, start with what AI is.

What is artificial general intelligence in simple terms?

Think of the AI you already use. A spam filter only sorts email. A chess engine only plays chess. A chatbot such as ChatGPT is far more flexible, but it still makes basic mistakes, forgets context and needs people to check its work.

AGI is the idea of a system that is general: it could pick up a new skill it was never trained on, carry knowledge from one area to another, and work through unfamiliar problems about as well as a capable adult. IBM describes it as a stage where an AI "can match or exceed the cognitive abilities of human beings across any task," and stresses that it is still hypothetical.

The word "general" is the whole point. Today's best models are broad, but broad is not the same as general. A model can write code, translate and summarize, yet still fail at a simple task a 10-year-old could learn in minutes.

What is the difference between narrow AI, AGI and super AI?

People use a lot of overlapping labels. Here is how they fit together.

Term

What it means

Exists today?

Narrow AI (weak AI)

AI built for specific tasks, like face unlock or route planning

Yes, almost all AI

AGI (general AI)

Human-level ability across most intellectual tasks

Disputed; no agreed proof

Strong AI

AI with a real mind or understanding, not just behavior

No evidence

Superintelligence (super AI, ASI)

Far beyond the best humans at nearly everything

No

Sentient or self-aware AI

AI that has feelings or experiences

No evidence

A few points that trip people up:

  • Narrow AI and weak AI are usually treated as the same thing. Every app on our list of real examples of AI is narrow AI.

  • Strong AI is an older philosophy term linked to John Searle. It is about whether a machine truly understands, not just how well it performs. AGI is about performance, so a system could count as AGI without being "strong AI."

  • Sentient AI is a separate question again. Being clever at tasks does not show that a system feels anything, and there is no scientific test that settles it.

How is AGI defined by the companies building it?

There is no single official definition, which is why AGI arguments go in circles. Three you will see most often:

  • OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work." That is a jobs-based test, not a brain-based one.

  • Google DeepMind's "Levels of AGI" paper (first published 2023, revised September 2025) rates systems on how well they perform and how general they are, from "Emerging" up to "Superhuman." It treats AGI as a ladder rather than an on/off switch.

  • Academic definitions, like the one in IBM's explainer, focus on matching human thinking across any task, including learning new skills without being retrained.

Contracts have used the term too. When Microsoft and OpenAI rewrote their deal in October 2025, Microsoft said that once OpenAI declared AGI, the claim "will now be verified by an independent expert panel." In April 2026 the companies changed terms again, and OpenAI said its revenue-share payments to Microsoft would continue through 2030 "independent of OpenAI's technology progress." Commentators, including The Verge, read that as the end of the old "AGI clause." Neither company has published the full contract, so treat that reading as interpretation.

Has AGI already been achieved in 2026?

This is the live debate, and the honest answer is: it depends on whose definition you use.

The case that we are close or there. On September 3, 2026, OpenAI released GPT-6 Astra, which can operate browsers, spreadsheets and desktop software. OpenAI president Greg Brockman told reporters, "I think it might be about this model," and closed the briefing with "Welcome to the AGI era." Google DeepMind CEO Demis Hassabis, speaking in May 2026, said he still broadly expects AGI around 2030 and now sees 2029 as possible.

The case against. Critics point to how the claim was measured. OpenAI's headline score on the ARC-AGI-3 reasoning benchmark was 99.9%, but Live Science reported in October 2026 that the figure relied on OpenAI's own test harness; under the benchmark's neutral Standard harness it was 62.7%. The ARC Prize Foundation's president, Greg Kamradt, said topping the benchmark is not "proof of achieving AGI." Brockman's framing also sits awkwardly with OpenAI's own economic-work definition, since its launch materials did not lead with real-world work results.

Others argue the whole approach is wrong. Yann LeCun, who left Meta in late 2025, says large language models are "Impressive, yes. Intelligent, no." His new company, AMI Labs, raised about $1.03 billion in March 2026 to build "world models" that learn how the physical world works.

So: one major lab says the AGI era has begun, another expects it around 2030, and some leading researchers think current methods will not get there at all.

Rows of server racks with blinking lights inside a data center

When do experts think AGI will arrive?

Forecasts have moved earlier fast. In the largest survey of its kind, Grace and colleagues asked 2,778 researchers who publish at top AI venues (survey run in late 2023, published January 2024). They gave a 50% chance that machines would outperform humans at every task by 2047, which was 13 years earlier than the same group's 2022 survey. They put a 10% chance on 2027.

Two cautions. That survey predates the 2025–2026 agent models, and "outperform humans at every task" is a stricter bar than OpenAI's "most economically valuable work." Different definitions produce very different dates.

Why does AGI matter, and what are the risks?

If a system really could do most knowledge work, the effects would be large: faster science and medicine, cheaper expert help, and big changes to jobs. Our guide to what AI is used for shows where narrow AI is already making a difference.

The risks people raise fall into a few groups:

  • Jobs and income. If AI can do most paid desk work, who earns what becomes a political question, not just a tech one.

  • Misuse. More capable systems can help attackers as well as defenders. OpenAI rated GPT-6 Astra as the first model to hit its "critical" level for cybersecurity capability, and it slowed the release to add more safety testing.

  • Control and oversight. OpenAI acknowledged that Astra was harder to monitor in tests designed to see whether a model could evade oversight. That is exactly the kind of problem safety researchers worry will grow.

  • Extreme outcomes. In the Grace survey, between 38% and 51% of respondents gave at least a 10% chance that advanced AI could lead to outcomes as bad as human extinction. That is not a prediction, but it shows the concern is mainstream among researchers.

  • Hype. Overstated AGI claims can push people to trust AI with decisions it cannot reliably make.

How can you check an "AGI is here" claim yourself?

Here is a six-question checklist we use when a new model is billed as AGI.

  1. Which definition? Human-level at every task, or "most economically valuable work"? Ask which one is being claimed.

  2. Who measured it? Company benchmarks are a start. Look for independent groups re-running the tests.

  3. Same test, same rules? Check whether the score used a special setup that others cannot reproduce.

  4. Can it learn something new? General intelligence means picking up unfamiliar skills, not just scoring well on known tests.

  5. How reliable is it? A system that is brilliant 90% of the time but confidently wrong the rest still needs a human checker.

  6. What do outside experts say? Read at least one skeptic and one supporter before deciding.

Worked example: GPT-6 Astra. (1) OpenAI's definition is economic work, but the headline was a reasoning benchmark. (2) Independent groups did re-run it. (3) The 99.9% used a proprietary harness; the neutral score was 62.7%. (4) ARC-AGI-3 does test new skills, a point in its favor. (5) It still makes errors, like every model. (6) Supporters and skeptics are split. Verdict: a big step, not settled proof.

Close-up of black and white chess pieces on a chess board

Is ChatGPT or generative AI the same as AGI?

No. Generative AI is a type of AI that creates text, images, audio or code. Tools like ChatGPT and other AI chatbots are built on it. Some people argue the newest models are early forms of AGI; most definitions say they are not there yet, because they still struggle with reliable reasoning, long-term memory and learning new skills on the fly.

Frequently asked questions

Is AGI the same as AI?

No. AI is the whole field, including simple tools like spam filters. AGI is one hypothetical goal within it: AI that is general enough to handle most intellectual tasks a human can.

Does AGI exist yet?

There is no agreed proof. OpenAI's president says GPT-6 Astra may mark the start of the AGI era, while independent benchmark groups and many researchers say it does not meet the bar.

What is the difference between AGI and superintelligence?

AGI means roughly human-level ability across many tasks. Superintelligence means far beyond the best humans at nearly everything. Superintelligence would come after AGI, if it ever comes.

Would AGI be self-aware?

Not necessarily. AGI is about what a system can do. Whether a machine could have feelings or awareness is a separate question, and there is no evidence any current AI does.

3 tips before you go

  1. When you see "AGI" in a headline, find the definition being used before you read the rest.

  2. Look for an independent benchmark result, such as one from the ARC Prize Foundation, before trusting a company's own score.

  3. Keep checking AI answers on anything that matters, such as health, money or legal questions, no matter how capable the model is said to be.

Related reading

Sources

How we write: see how we test and our editorial policy.

AICYBO is an independent site. We are not affiliated with, endorsed by or sponsored by OpenAI, Microsoft, Google, Google DeepMind, IBM, Meta, AMI Labs or the ARC Prize Foundation. Product names belong to their owners.

Written and checked under our editorial policy. Spotted a mistake? Tell us and we will fix it.