# AGI vs ASI: what is the difference?

> AGI matches people across most tasks; ASI would far exceed them. The definitions labs use, the levels in between, 2026 benchmarks and who is building ASI.

By Tautvydas Kalvaitis · Updated 2026-10-11 · https://whyworry.si/guides/agi-vs-asi

**Key facts**

- AGI (artificial general intelligence) is AI that matches skilled humans across most tasks; OpenAI's 2018 Charter defines it by outperforming humans at most economically valuable work.
- ASI (artificial superintelligence) is AI that far exceeds humans in virtually every field, in Nick Bostrom's 2014 definition.
- Google DeepMind's "Levels of AGI" framework (November 2023) puts ASI at Level 5, a system that beats every human across a wide range of tasks.
- On 24 July 2026 the ARC Prize listed Claude Opus 5 at 90.4% on ARC-AGI-2 and 30.2% on the harder ARC-AGI-3.
- The US executive order of 29 September 2026 renamed AI "Super Intelligence" in federal communications; it is a new label for today's AI, not a claim that ASI exists.


AGI is AI that can do most intellectual work as well as a skilled person. ASI is AI that would do almost all of it far better than any person. As of October 2026, no lab has declared AGI, and ASI remains hypothetical. Below: the definitions in the words of the people who set them, the levels in between, the 2026 evidence and who says they are building ASI.

**AGI**

- Artificial general intelligence
- Matches skilled humans at most tasks
- DeepMind Level 2 and up
- Not declared by any lab as of Oct 2026

**ASI**

- Artificial superintelligence
- Far exceeds all humans in almost every field
- DeepMind Level 5
- Hypothetical as of Oct 2026


## AGI vs ASI in one table

| | AGI | ASI |
| --- | --- | --- |
| Short definition | Matches or beats humans at most economically valuable work | Greatly exceeds humans in virtually all domains |
| Who set it | OpenAI Charter, April 2018 | Nick Bostrom, *Superintelligence*, 2014 |
| DeepMind level | Level 2 "Competent" (at least 50th percentile of skilled adults) and up | Level 5 "Superhuman" (better than every human) |
| How it improves | Mostly with human researchers in the loop | Often linked to recursive self-improvement, where the system improves itself |
| Status, Oct 2026 | Not declared; OpenAI's claim would go to an independent expert panel | Hypothetical |

Sources: [OpenAI Charter](https://openai.com/charter/), [Nextgov](https://www.nextgov.com/artificial-intelligence/2026/09/super-intelligence-presidents-new-term-ai-explained/416164/), [DeepMind, Levels of AGI](https://arxiv.org/html/2311.02462v5), [IBM](https://www.ibm.com/think/topics/artificial-superintelligence), [Microsoft](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/).

OpenAI's Charter defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". That is a jobs test: could the system do most paid knowledge work? Bostrom defines superintelligence as "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest" ([Nextgov](https://www.nextgov.com/artificial-intelligence/2026/09/super-intelligence-presidents-new-term-ai-explained/416164/)). That is a different bar: not equal to the best people, but far beyond them.

The gap matters. A system at AGI level could replace a lot of work. A system at ASI level could, in principle, out-think every scientist, strategist and engineer at once. IBM describes ASI as a system that would keep improving itself as it learns ([IBM](https://www.ibm.com/think/topics/artificial-superintelligence)).

Who decides when AGI is here is also now written down. Under the Microsoft and OpenAI agreement of 28 October 2025, an OpenAI declaration of AGI would be checked by an independent expert panel ([Microsoft](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/)).

## The ladder: ANI, AGI, ASI

1. **Level 1: Emerging**: Equal to or somewhat better than an unskilled human; where the paper placed ChatGPT, Bard, Llama 2 and Gemini in 2023
2. **Level 2: Competent**: At least the 50th percentile of skilled adults across general tasks
3. **Level 3: Expert**: At least the 90th percentile of skilled adults
4. **Level 4: Exceptional**: At least the 99th percentile of skilled adults (first called "Virtuoso")
5. **Level 5: Superhuman**: Outperforms every human; the paper calls this ASI


Google DeepMind's "Levels of AGI" paper (first posted 4 November 2023, revised 24 September 2025) turns the vague debate into a ladder ([arXiv](https://arxiv.org/html/2311.02462v5)). It rates systems on two axes: how well they perform and how general they are. Level 0 is no AI at all. In the general column, the paper marks Levels 2 to 5 as not yet reached.

Everything below the general column is narrow AI, also called artificial narrow intelligence (ANI): a system built for one kind of task. Siri, spam filters and chess engines are ANI. A chess engine can be superhuman at chess and still be narrow.

ChatGPT and similar chatbots are generative AI. They are far more general than Siri, but most researchers still class them as below AGI: the DeepMind paper put 2023 chatbots at Level 1, competent on some tasks and emerging on most.

![A small single-purpose robot, a human-sized figure juggling many tools and a giant galaxy-bodied figure stand in a row while Lumen measures them.](https://whyworry.si/img/guides/agi-vs-asi-fig.webp)


## Where the ideas come from

- **1965**: Statistician I. J. Good describes an "ultraintelligent machine" and an intelligence explosion ([Quote Investigator](https://quoteinvestigator.com/2022/01/04/ultraintelligent/))
- **2014**: Nick Bostrom publishes *Superintelligence: Paths, Dangers, Strategies*
- **Apr 2018**: OpenAI Charter sets its AGI definition ([OpenAI](https://openai.com/charter/))
- **4 Nov 2023**: Google DeepMind publishes "Levels of AGI" ([arXiv](https://arxiv.org/html/2311.02462v5))
- **10 Jun 2025**: Sam Altman, "The Gentle Singularity" ([blog](https://blog.samaltman.com/the-gentle-singularity))
- **22 Oct 2025**: Future of Life Institute Statement on Superintelligence ([FLI](https://superintelligence-statement.org/))
- **28 Oct 2025**: Microsoft and OpenAI agree an expert panel will verify any AGI declaration ([Microsoft](https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/))
- **12 Jun 2026**: Google DeepMind publishes "From AGI to ASI" ([DeepMind](https://deepmind.google/research/publications/239142/))
- **6 Sep 2026**: OpenAI says it reached its "automated AI research intern" goal ([TechRadar](https://www.techradar.com/pro/openai-says-it-has-built-an-automated-research-intern-to-carry-out-menial-tasks-and-its-only-just-getting-started))
- **29 Sep 2026**: US Executive Order 14434 renames AI "Super Intelligence" ([whitehouse.gov](https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/))


The core idea is 60 years old. Good argued that a machine able to out-think humans could design even better machines, setting off an intelligence explosion. He called such a machine the last invention people would need to make, with the caveat that it would have to stay under human control ([Quote Investigator](https://quoteinvestigator.com/2022/01/04/ultraintelligent/)). The popular name for that runaway point is the singularity, which Altman's 2025 essay title borrows.

The newest academic map is DeepMind's June 2026 report. It describes four possible paths from AGI to ASI: scaling AGI systems up, a shift to a new AI paradigm, recursive improvement, and ASI emerging from large collectives of AI agents working together ([DeepMind](https://deepmind.google/research/publications/239142/)).

## How close is AGI? The 2026 evidence

- **90.4%**: Claude Opus 5 on ARC-AGI-2 (Max effort)
- **30.2%**: Claude Opus 5 on ARC-AGI-3 (High effort)
- **Sep 2026**: OpenAI says its "research intern" goal is met
- **Mar 2028**: OpenAI's target for an automated AI researcher

Source: [ARC Prize, 24 Jul 2026](https://arcprize.org/results/anthropic-claude-opus-5), [Unite.AI](https://unite.ai/openai-hits-goal-of-building-an-automated-research-intern)


Benchmarks built to test general reasoning are being beaten fast. ARC-AGI-2 was designed to be easy for people and hard for AI; on 24 July 2026 the ARC Prize listed Claude Opus 5 at 90.4% on it. The newer ARC-AGI-3 is much harder, and the same model topped it at 30.2% ([ARC Prize](https://arcprize.org/results/anthropic-claude-opus-5)).

On 6 September 2026 OpenAI said it had reached the goal it set in October 2025: an automated research intern that can do well-defined research tasks under human direction. Its next target is an automated AI researcher by March 2028 ([Unite.AI](https://unite.ai/openai-hits-goal-of-building-an-automated-research-intern)). The evidence is OpenAI's own measurement, and reporters noted that agents still needed human steering on longer tasks.

So, is it AGI? No lab has said so, and a high score on one benchmark is not the same as doing most jobs.

## When might ASI arrive? What forecasters say

| Forecaster | What they predict | Median date | As of |
| --- | --- | --- | --- |
| Metaculus community, about 1,900 forecasters | First general AI announced | 5 Jan 2031 | 11 Oct 2026 |
| Daniel Kokotajlo, AI Futures Project | Superintelligence | Mar 2029 | 16 Aug 2026 |
| Eli Lifland, AI Futures Project | Superintelligence | Jul 2033 | 16 Aug 2026 |
| AI Impacts survey of AI researchers | 50% chance of human-level machine intelligence | 2047 | 2023 survey |

Sources: [Metaculus](https://www.metaculus.com/questions/5121/), [AI Futures Project](https://blog.aifutures.org/p/q25-2026-timelines-update-uplift), [AI Impacts](https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_ai).

Forecasts are opinions with wide ranges. On Metaculus, the middle half of forecasts runs from April 2028 to April 2037. The AI Futures Project, the team behind [the AI 2027 scenario](/guides/ai-2027), shows a gap of more than four years between two of its own authors.

Academic surveys run later than forecasting sites. In the 2023 AI Impacts survey, researchers gave even odds of human-level machine intelligence by 2047, 13 years earlier than the same survey found in 2022. ASI, by most accounts, comes after AGI, but some forecasters expect the gap to be short if AI starts to speed up AI research itself.

## Who says they are building superintelligence

| Organisation | Unit or stated aim | Since | Note |
| --- | --- | --- | --- |
| Safe Superintelligence Inc. | A lab with one goal: safe superintelligence | Jun 2024 | Co-founded by Ilya Sutskever; reported $32B valuation in Apr 2025 |
| Meta | Meta Superintelligence Labs | 30 Jun 2025 | Led by Alexandr Wang with Nat Friedman, after Meta's Scale AI investment |
| OpenAI | Altman wrote that humanity is close to building digital superintelligence | 10 Jun 2025 | Essay "The Gentle Singularity" |
| Microsoft AI | MAI Superintelligence Team, aiming at "humanist" superintelligence | 6 Nov 2025 | Led by Mustafa Suleyman |
| Google DeepMind | Research on paths from AGI to ASI | 12 Jun 2026 | Report "From AGI to ASI" |

Sources: [TechCrunch](https://techcrunch.com/2025/04/12/openai-co-founder-ilya-sutskevers-safe-superintelligence-reportedly-valued-at-32b), [NBC New York](https://www.nbcnewyork.com/news/business/money-report/mark-zuckerberg-announces-creation-of-meta-superintelligence-labs-read-the-memo/6320911/), [Sam Altman](https://blog.samaltman.com/the-gentle-singularity), [Microsoft AI](https://microsoft.ai/blog/towards-humanist-superintelligence), [DeepMind](https://deepmind.google/research/publications/239142/).

A few years ago, "superintelligence" was a word for philosophers. Now it is a team name at Meta and Microsoft, and the stated goal of a startup valued in the tens of billions. SSI's valuation was reported by the press and not confirmed by the company.

The aims differ. Microsoft frames its goal as superintelligence that stays under human control and serves specific human needs. Altman's essay argues that systems able to produce new scientific insight are close. Governments are also building their own compute and models; see [sovereign AI](/guides/what-is-sovereign-ai).

## The safety debate

> **A call to stop** The Future of Life Institute's Statement on Superintelligence, launched on 22 October 2025, calls for a ban on developing superintelligence until there is broad scientific consensus that it can be done safely and controllably, and strong public support. The site showed 77,280 signatures on 11 October 2026, about 5,000 of them via an Ekō petition.


Two positions dominate. Supporters, Altman among them, argue that systems able to produce new insights could speed up scientific discovery and make many things far cheaper ([Altman](https://blog.samaltman.com/the-gentle-singularity)). Critics, including many signers of the FLI statement, argue that nobody yet knows how to keep a system smarter than its makers under control, so the race itself is the risk ([FLI](https://superintelligence-statement.org/)).

The worry is not new. Good's 1965 paper already made the payoff conditional on the machine staying controllable. Bostrom's 2014 book is mostly about that control problem. This guide takes no side; the dates and sources are here so you can weigh them.

## "Super Intelligence" in Washington is not ASI

**US "Super Intelligence" (SI)**

- A new federal name for today's AI systems
- Defined by the existing AI definition in 15 U.S.C. 9401(3)
- Set by Executive Order 14434, 29 Sep 2026
- Describes systems already in use

**ASI**

- A hypothetical capability level
- AI far beyond all humans in almost every field
- Defined by researchers, from Bostrom to DeepMind
- Not achieved by any system as of Oct 2026


Since 29 September 2026, US federal agencies call AI "Super Intelligence", or SI. The order behind this ties SI to the same legal definition that AI already had ([whitehouse.gov](https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/)). It is a renaming of AI. It does not claim that ASI exists, and it does not change what any system can do.

That creates real confusion. When a federal document says "SI", it means the AI you already use. When a researcher says "superintelligence", they mean Level 5. Read [the US Super Intelligence executive order](/guides/super-intelligence-executive-order) for the full text and timeline. One more source of noise: searches for "ASI AI" also return crypto projects and software products that use the acronym; this guide means the concept.

If you are naming a company in this space, the renaming has also turned .si, Slovenia's domain, into shorthand for superintelligence. Our guide to [.si, the superintelligence domain](/guides/si-domains) covers how it works. Disclosure: this site sells .si domain names.

## FAQ

### What is ASI in AI?

ASI stands for artificial superintelligence: AI that would far exceed human ability in virtually every field, from science to strategy to social skill. Philosopher Nick Bostrom gave the standard definition in his 2014 book *Superintelligence*. Google DeepMind's "Levels of AGI" framework puts ASI at Level 5, above every human on a wide range of tasks. As of October 2026, ASI is hypothetical; no system has reached it.

### Is ChatGPT AGI?

No, not by most definitions. ChatGPT is generative AI built on a large language model. It handles many tasks, but it does not reliably match skilled humans at most economically valuable work, which is OpenAI's own AGI test. Google DeepMind's 2023 framework placed ChatGPT-era chatbots at Level 1, "Emerging AGI". OpenAI has not declared AGI as of October 2026, and any declaration would go to an independent expert panel.

### Is Siri ANI or AGI?

Siri is ANI, artificial narrow intelligence. It is built for a set range of jobs: answering questions, setting reminders, controlling apps. It cannot learn a new kind of work on its own or perform across most tasks the way a skilled person can, which is what AGI would need. Most AI in daily use, including spam filters, recommendation engines and voice assistants, is narrow AI.

### Is ASI possible?

Nobody knows. Many lab leaders say yes and are hiring for it: Meta, Microsoft and Safe Superintelligence Inc. all name superintelligence as a goal. Some researchers doubt the concept holds together; the authors of the 2025 paper "AI as Normal Technology" argue AI is not usefully smarter than people who work with AI. Forecasters who expect it give dates ranging from 2029 to well after 2033, and Google DeepMind's 2026 report maps four possible paths to it.

### Is the US government's "Super Intelligence" the same as ASI?

No. Executive Order 14434, signed on 29 September 2026, tells federal agencies to call AI "Super Intelligence" or "SI". The order defines SI as the same technologies already covered by the legal definition of AI in 15 U.S.C. 9401(3). It is a renaming of today's AI for government use, not a claim that artificial superintelligence exists. ASI remains a hypothetical level beyond all humans.

## Sources

- https://www.nextgov.com/artificial-intelligence/2026/09/super-intelligence-presidents-new-term-ai-explained/416164/
- https://openai.com/charter/
- https://arxiv.org/html/2311.02462v5
- https://deepmind.google/research/publications/239142/
- https://www.ibm.com/think/topics/artificial-superintelligence
- https://quoteinvestigator.com/2022/01/04/ultraintelligent/
- https://blog.samaltman.com/the-gentle-singularity
- https://www.nbcnewyork.com/news/business/money-report/mark-zuckerberg-announces-creation-of-meta-superintelligence-labs-read-the-memo/6320911/
- https://techcrunch.com/2025/04/12/openai-co-founder-ilya-sutskevers-safe-superintelligence-reportedly-valued-at-32b
- https://microsoft.ai/blog/towards-humanist-superintelligence
- https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/
- https://arcprize.org/results/anthropic-claude-opus-5
- https://www.techradar.com/pro/openai-says-it-has-built-an-automated-research-intern-to-carry-out-menial-tasks-and-its-only-just-getting-started
- https://unite.ai/openai-hits-goal-of-building-an-automated-research-intern
- https://www.metaculus.com/questions/5121/
- https://blog.aifutures.org/p/q25-2026-timelines-update-uplift
- https://wiki.aiimpacts.org/ai_timelines/predictions_of_human-level_ai_timelines/ai_timeline_surveys/2023_expert_survey_on_progress_in_ai
- https://superintelligence-statement.org/
- https://www.whitehouse.gov/presidential-actions/2026/09/inaugurating-the-era-of-super-intelligence/

## Names that fit this guide

We own and sell these names; prices are on each page.

- aligners.si: https://aligners.si/
- posthumans.si: https://posthumans.si/
- speeds.si: https://speeds.si/
