# AGI vs ASI vs ANI: The Three Levels of AI, Explained

- URL: https://superintelligence-news.com/articles/agi-vs-asi-vs-ani
- Publisher: Superintelligence News
- Section: AGI
- Type: explainer
- Author: Umair
- Published: 2026-10-16
- In this story: [Google DeepMind](https://superintelligence-news.com/companies/google-deepmind), [OpenAI](https://superintelligence-news.com/companies/openai), [Nick Bostrom](https://superintelligence-news.com/people/nick-bostrom)

**In plain English:** Narrow AI is good at one job, like chess. AGI would be about as capable as a person at most thinking jobs. ASI would beat the best humans at nearly everything. Only the first is surely here today.

> ANI, or narrow AI, is skilled at one kind of task. AGI would match people at most thinking tasks. ASI would far exceed the best humans at almost everything. Narrow AI is everywhere. Whether AGI has arrived is disputed, after OpenAI's president suggested in September 2026 that an "AGI era" had begun. Google DeepMind researchers list ASI as "not yet achieved".

## Key facts

- Google DeepMind's Levels of AGI framework rates AI on two scales, how general it is and how well it performs, and lists artificial superintelligence (ASI) as "not yet achieved".
- The same framework placed ChatGPT, as of 2023, at Level 1, "Emerging AGI", and counts Deep Blue and AlphaGo as "Exceptional Narrow AI".
- OpenAI's charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work".
- After OpenAI unveiled GPT-6 Astra on September 3, 2026, the ARC Prize Foundation, which tested it, said it was "not claiming that it is AGI".
- A June 2026 Google DeepMind report sets the bar for ASI above "large human-expert collectives", not just above the best single human.

ANI, AGI and ASI are three rough levels of AI ability. Narrow AI (ANI) is skilled at one kind of task. Artificial general intelligence (AGI) would match people at most thinking tasks. Artificial superintelligence (ASI) would far exceed the best humans at almost everything.

Narrow AI is all around us. Whether AGI has arrived is now openly disputed. Google DeepMind researchers list ASI as "not yet achieved". This guide explains each term, where chatbots such as ChatGPT fit, and why experts disagree. For the wider picture, start with [what is superintelligence](/articles/what-is-superintelligence).

## What is the difference between ANI, AGI and ASI?

The difference is how broad an AI's skills are and how strong they are. ANI handles one clearly scoped task. AGI would handle a wide range of tasks about as well as a typical skilled person. ASI would beat the best people at nearly every task. One 2026 Google DeepMind report sets the bar higher still, at large teams of experts.

| Level | Also called | What it means | Examples | Exists today? |
|---|---|---|---|---|
| Artificial narrow intelligence (ANI) | Narrow AI, weak AI | Skilled at one task or a small set of tasks | Spelling and grammar checkers, chess engines, AlphaFold | Yes |
| Artificial general intelligence (AGI) | General AI, human-level AI | About as good as a person at most thinking tasks, including learning new skills | None that experts agree on | Disputed |
| Artificial superintelligence (ASI) | Superintelligence | Far beyond the best humans in almost every field | None | No |

The labels describe what a system can do, not how it is built. The same model can look strong on one task and weak on another. That is why the line between narrow AI and AGI is so hard to draw.

## What is ANI (artificial narrow intelligence)?

ANI is AI built for one job or a small set of jobs. It is the only one of the three levels that clearly exists. IBM's explainer, first published in 2023, notes that current AI is often called artificial narrow intelligence, weak AI or narrow AI. Narrow does not mean weak in skill, though.

Google DeepMind's [Levels of AGI paper](https://arxiv.org/html/2311.02462v5) rates narrow systems from "Emerging" to "Superhuman". AlphaFold predicts a protein's 3D structure from its amino acid sequence. It does this "above the level of the world's top scientists", the paper says, so it ranks as "Superhuman Narrow AI". It does that one task only.

The photo at the top shows a retired back-end rack from AlphaGo, DeepMind's Go-playing program. The paper counts AlphaGo as "Exceptional Narrow AI". That means it plays better than at least 99% of skilled adults, but at one game.

## What is AGI (artificial general intelligence)?

AGI is AI that could do most thinking tasks about as well as people, and learn new ones. People do not agree on one meaning. Depending on who sets it, the bar ranges from roughly an unskilled person to beating humans at most paid work.

Three versions show the range:

- **A lower bar:** the Levels of AGI paper places the definition in a 2023 essay by Blaise Agüera y Arcas and Peter Norvig, "[Artificial General Intelligence Is Already Here](https://www.noemamag.com/artificial-general-intelligence-is-already-here/)", at its lowest level, "Emerging AGI".
- **The middle:** Google DeepMind's researchers (2023, revised 2025) describe a "Competent AGI" as one that performs at least as well as half of skilled adults on most cognitive tasks. They say this level "best corresponds to many prior conceptions of AGI". A June 2026 Google DeepMind report uses the same bar: "An AGI is a system that is roughly as intelligent as a single human."
- **A higher bar:** OpenAI's [charter](https://openai.com/charter/) (2018) defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". The DeepMind paper says this "better matches" its higher level, "Exceptional AGI".

The term dates back to 1997. According to the Levels of AGI paper, Mark Gubrud first used "artificial general intelligence" in an article about military technology that year. The paper says Shane Legg, who later co-founded DeepMind, and Ben Goertzel "popularized the term AGI among computer scientists in 2001". IBM's explainer dates its spread to a 2007 book by Goertzel with the same name.

## What is ASI (artificial superintelligence)?

ASI is AI that would be far more able than humans in nearly every field. In a [paper first written in 1997](https://nickbostrom.com/papers/how-long-before-superintelligence/), philosopher Nick Bostrom defined it as an intellect "much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills".

Google DeepMind's 2026 report, [From AGI to ASI](https://arxiv.org/html/2606.12683v2), describes an ASI as an AI that "has superhuman abilities across virtually all tasks and domains of human interest and activity". Its authors say they "set the bar for ASI high". To count, a system must beat "large human-expert collectives", meaning the best that a large, well-run group of people could do.

So the two bars differ. Bostrom compared a superintelligence with "the best human brains". The DeepMind report compares it with large groups of experts working together, which it describes as "outperforming large groups of experts instead of individual experts".

Both rule out narrow systems. The 2026 report says AlphaFold and AlphaGo, which are superhuman "in single domains", "are thus ruled out as ASIs".

## Google DeepMind's six levels, from no AI to ASI

Researchers at Google DeepMind proposed a way to place any AI on two scales at once: how general it is, and how well it performs. Their table has a Level 0 ("No AI") and five performance levels, each split into narrow and general AI. ASI is the top level of general AI, and the table lists it as not yet achieved.

| Level | Performance | Narrow AI examples | General AI |
|---|---|---|---|
| 0: No AI | Not AI | Calculator software, compilers | People doing tasks through a computer, such as Amazon Mechanical Turk |
| 1: Emerging | Equal to or somewhat better than an unskilled person | Simple rule-based systems | ChatGPT, Bard, Llama 2, Gemini (as rated in 2023) |
| 2: Competent | Better than at least 50% of skilled adults | Siri, Alexa, Google Assistant | Not yet achieved |
| 3: Expert | Better than at least 90% of skilled adults | Grammarly, image generators such as DALL-E 2 | Not yet achieved |
| 4: Exceptional | Better than at least 99% of skilled adults | Deep Blue, AlphaGo | Not yet achieved |
| 5: Superhuman | Better than all humans | AlphaFold, AlphaZero, Stockfish | ASI: not yet achieved |

Source: the [Levels of AGI paper](https://arxiv.org/html/2311.02462v5), Table 1 (version 5, September 2025). The authors call their placement of the examples "approximate".

The table makes one point clear. A system can be superhuman yet narrow, like AlphaFold. It can also be general yet weak, like the 2023 chatbots. AGI and ASI both need breadth and strength together. The paper adds that a general system at one level may still reach higher levels on a few tasks.

## Is ChatGPT ANI, AGI or ASI?

ChatGPT is not ASI. Whether it counts as narrow AI or early AGI depends on the framework. Many explainers, including IBM's, describe current AI as narrow. Google DeepMind's researchers instead rated ChatGPT, as of 2023, as "Emerging AGI": general, but at roughly the level of an unskilled person on most tasks.

In September 2023, the paper said such models were already "Competent" at some tasks, such as "short essay writing, simple coding". They were still "Emerging" at most tasks, including math and tasks that depend on facts.

Models have got better since then, but not evenly. The [International AI Safety Report 2026](https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026) says capabilities remain "jagged": "leading systems may excel at some difficult tasks while failing at other, simpler ones." So the same chatbot can look like AGI on one test and like narrow AI on another.

## Has AGI been achieved yet?

There is no agreement. In September 2026, OpenAI's president suggested its new GPT-6 Astra model marked the start of an "AGI era". Nvidia's chief executive wrote that "AGI has arrived". The group behind a test OpenAI cited, and two experts quoted by Information Age, said the evidence does not show it.

OpenAI unveiled GPT-6 Astra on September 3, 2026, and began rolling it out to customers. In a briefing for reporters before the launch, OpenAI president Greg Brockman said: "I think it's not unreasonable to feel that we are now in the AGI era," [Fox Business](https://www.foxbusiness.com/technology/nvidia-ceo-jensen-huang-declares-agi-has-arrived-after-openai-unveils-gpt-6-astra) and [Information Age](https://ia.acs.org.au/article/2026/openai-says-the-agi-era-is-here-experts-disagree.html) reported. Jensen Huang, chief executive of Nvidia, a major OpenAI partner, wrote on X that "AGI has arrived". Fox Business noted that Nvidia's own test for AGI is top scores on professional exams in many fields.

Others disagreed:

- **The ARC Prize Foundation**, which runs the ARC-AGI-3 test that Brockman pointed to, [wrote](https://arcprize.org/blog/astra): "while we believe Astra represents meaningful progress towards generalization, we are not claiming that it is AGI." It reported 62.7% on its standard setup, where the model keeps only the notes it chooses to write, and 99.9% on a setup that keeps the model's hidden reasoning between steps. ARC Prize says a future AGI "should be able to solve ARC-AGI-3" under the standard conditions.
- **Toby Walsh**, chief scientist of the UNSW AI Institute in Australia, told Information Age: "We really don't have a single test yet for AGI."
- **Rebecca Johnson**, an AI evaluation expert at the University of Sydney, said: "AGI has a philosophy-of-science problem masquerading as a benchmark problem."

OpenAI's chief executive, Sam Altman, has himself called AGI an "irrelevant marketing term", Information Age reported, while suggesting it is "sort of" here, or "close, at least".

No public body decides whether AGI has arrived. In October 2025, [Microsoft said](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/) that under its deal with OpenAI, "once AGI is declared by OpenAI, that declaration will now be verified by an independent expert panel." In April 2026, the two companies [changed their deal again](https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/). OpenAI's revenue-share payments to Microsoft continue through 2030, "independent of OpenAI's technology progress", up to a total cap.

## How big is the gap between AGI and ASI?

In ability, the gap is large by the yardstick in Google DeepMind's 2026 report. How long the jump would take is far less clear. The report pictures AGI as "roughly as intelligent as a single human" and ASI as better than large teams of top experts. It also expects the first AGI to be well ahead of people in some areas: "the first AGI will already be superhuman on many tasks".

The report describes four possible paths from AGI to ASI. They are scaling up, new kinds of AI, AI improving AI, and many AI systems working together. We explain them in our guide to [what is superintelligence](/articles/what-is-superintelligence).

The report also lists six possible brakes, and calls the size of each one "an open research question". Two of the six concern how AI is built:

- **The neural paradigm may not be enough.** Today's large neural networks, even with extra training and tools, might not reach AGI at all. The report adds that today's methods may instead keep improving smoothly, in which case this would be "at best a friction but not a fundamental blocker".
- **An "abstraction barrier".** Systems trained mainly on human ideas may struggle to form new concepts from raw data, which the report calls "a major factor in human scientific & cultural progress".

The authors stress that their own view comes with low confidence. They write that it is "implausible that AI progress would stall exactly at human-level intelligence", and that "the possibility of cruising past AGI and into ASI territory within the next decade or two cannot easily be dismissed".

Forecasts for when AGI or ASI might arrive range from a few years to several decades. We track them, with dates and sources, in [When Will Superintelligence Arrive?](/articles/when-will-superintelligence-arrive)

## Why do the labels matter?

The words now matter in law and business. A US Senate bill, the [Ban Artificial Superintelligence Act of 2026](https://www.govinfo.gov/app/details/BILLS-119s5493is), would forbid anyone to develop, deploy or fund ASI. AGI appeared in the Microsoft and OpenAI contract. Getting the level wrong can make a story sound bigger or smaller than it is.

Names can mislead, too. Executive Order 14434 tells US federal agencies to call ordinary AI "Super Intelligence". That name does not mean ASI. We explain that change in [Trump's "Super Intelligence" executive order](/articles/trump-rename-ai-super-intelligence) and the wording question in [super intelligence vs artificial intelligence](/articles/super-intelligence-vs-artificial-intelligence).

## Frequently asked questions

### What are the three types of AI?

The three types most often named are artificial narrow intelligence (ANI), artificial general intelligence (AGI) and artificial superintelligence (ASI). They describe how broad and how strong an AI's abilities are, not how it is built.

### Is ChatGPT an AGI?

There is no agreement. Google DeepMind researchers rated ChatGPT, as of 2023, as "Emerging AGI", the lowest of their five AGI levels. In 2026, OpenAI's president suggested its newest model marked an "AGI era", while outside experts said the evidence does not show it.

### Who coined the term AGI?

Mark Gubrud used "artificial general intelligence" in a 1997 article about military technology, according to Google DeepMind's Levels of AGI paper. The same paper says Shane Legg and Ben Goertzel popularized the term among computer scientists in 2001.

### What comes after AGI?

Artificial superintelligence, if it is ever built. A June 2026 Google DeepMind report describes four possible paths from AGI to ASI and lists bottlenecks that could slow or stop each one.

## Sources

- [Morris et al. (Google DeepMind), Levels of AGI for Operationalizing Progress on the Path to AGI (arXiv, version 5, September 2025)](https://arxiv.org/html/2311.02462v5)
- [Genewein et al. (Google DeepMind), From AGI to ASI (arXiv, version 2, August 2026; first posted June 2026)](https://arxiv.org/html/2606.12683v2)
- [Blaise Agüera y Arcas and Peter Norvig, Artificial General Intelligence Is Already Here (Noema, October 2023)](https://www.noemamag.com/artificial-general-intelligence-is-already-here/)
- [OpenAI Charter (April 9, 2018)](https://openai.com/charter/)
- [Nick Bostrom, How Long Before Superintelligence? (first written 1997)](https://nickbostrom.com/papers/how-long-before-superintelligence/)
- [International AI Safety Report 2026](https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026)
- [ARC Prize, OpenAI's GPT-6 Astra on ARC-AGI-3 (September 3, 2026)](https://arcprize.org/blog/astra)
- [Microsoft, The next chapter of the Microsoft–OpenAI partnership (October 28, 2025)](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/)
- [Microsoft, The next phase of the Microsoft-OpenAI partnership (April 27, 2026)](https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/)
- [IBM, What is artificial superintelligence? (first published December 2023)](https://www.ibm.com/think/topics/artificial-superintelligence)
- [Information Age (Australian Computer Society), OpenAI says 'the AGI era' is here. Experts disagree (September 2026)](https://ia.acs.org.au/article/2026/openai-says-the-agi-era-is-here-experts-disagree.html)
- [Fox Business, Nvidia CEO Jensen Huang declares 'AGI has arrived' after OpenAI unveils GPT-6 Astra (September 2026)](https://www.foxbusiness.com/technology/nvidia-ceo-jensen-huang-declares-agi-has-arrived-after-openai-unveils-gpt-6-astra)
- [The Next Web, OpenAI's AGI number came from a harness, not the model (September 2026)](https://thenextweb.com/news/openai-astra-arc-agi-3-harness-62-7-vs-99-9-benchmark-revisions)