Request an invitation ↗
Journal

/

Agentic AI

AGI vs ASI: what superintelligence means, and who says they are building it

AGI and artificial superintelligence defined side by side from primary texts, and a dated record of the labs that now say they are building superintelligence.

November 12, 2026

8 min read

The Journal

In short

  • Since 2023, OpenAI, Meta and Microsoft have each said in writing that they are working towards superintelligence, and a company was set up in 2024 with it as its only product.
  • AGI (artificial general intelligence) usually means AI that performs at the level of skilled people across most cognitive work. ASI (artificial superintelligence) means AI that performs beyond the best people across that same range.
  • A Google DeepMind team put the line in numbers: its "Competent AGI" level starts at the 50th percentile of skilled adults, while ASI is a general system that "outperforms 100% of humans".
  • The word means different things in different statements, so check which definition a claim uses and how progress would be measured.

AGI vs ASI: the definitions side by side

AGI. OpenAI's Charter, published in 2018, defines AGI as "highly autonomous systems that outperform humans at most economically valuable work". A 2023 paper by Meredith Ringel Morris and seven Google DeepMind co-authors, Levels of AGI, grades systems on depth (how well they perform against skilled people) and breadth (how many tasks they can handle). Its "Competent AGI" level, which the authors call "probably the best catch-all for many existing definitions of AGI", requires performance "at least 50th percentile of skilled adults" on most cognitive tasks. The paper places OpenAI's threshold higher, at Level 4, "Exceptional AGI", the 99th percentile.

ASI. The most cited definition comes from the philosopher Nick Bostrom. In a paper first published in 1998, he described a superintelligence as "an intellect that is much smarter than the best human brains in practically every field, including scientific creativity, general wisdom and social skills". His 2014 book shortened it to "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest", as quoted in a survey of expert opinion he co-wrote.

The top of the DeepMind scale matches Bostrom's idea. Level 5, "Superhuman", means a system that "outperforms 100% of humans". A narrow system can already sit there: the authors class AlphaFold, which predicts protein structures, as "Superhuman Narrow AI". ASI is the general version, able to do "a wide range of tasks at a level that no human can match". In the paper's table it is marked "not yet achieved".

Recent definitions from labs and a campaign group:

  • Microsoft AI, November 2025. Mustafa Suleyman, who leads its superintelligence team: "If AGI is often seen as the point at which an AI can match human performance at all tasks, then superintelligence is when it can go far beyond that performance."
  • OpenAI, April 2026. A policy paper describes superintelligence as "AI systems capable of outperforming the smartest humans even when they are assisted by AI".
  • Future of Life Institute, autumn 2025. The group, which campaigns against building it, uses "artificial intelligence capable of outperforming all humans at most cognitive tasks".

Where the difference lies

Every definition above treats both terms as general, so a program that beats every human at one task, such as chess or protein folding, stays narrow however strong it is. Bostrom's 1998 paper also rules out organisations: "Entities such as companies or the scientific community are not superintelligences according to this definition." The two differ in depth: AGI is measured against skilled people, and ASI against the best of them in nearly every field, three levels higher on the DeepMind scale.

Three of the authors quoted here point to self-improvement as the route from one level to the next. Bostrom wrote in 2009 that once machines reach roughly human level, "the step from there to superintelligence is likely to be much quicker", in a scenario where a "seed AI" "applies its wits to create a smarter version of itself". Suleyman expects superintelligence to have an "open-ended ability of 'learning to learn'". Meta's chief executive, Mark Zuckerberg, wrote in July 2025: "Over the last few months we have begun to see glimpses of our AI systems improving themselves."

The agentic half of the debate, about what AI systems do on their own, is covered in AI agents vs agentic AI.

From AGI to superintelligence: the record, 2023 to 2026

  • May 2023, OpenAI. In Governance of superintelligence, Sam Altman, Greg Brockman and Ilya Sutskever called superintelligence "future AI systems dramatically more capable than even AGI". They wrote that it was conceivable that "within the next ten years, AI systems will exceed expert skill level in most domains".
  • July 2023, OpenAI. Announcing its Superalignment team, OpenAI explained the choice of term in a footnote: "Here we focus on superintelligence rather than AGI to stress a much higher capability level." The same post said "we believe it could arrive this decade".
  • 2024, Safe Superintelligence Inc. A new company opened its website with "Superintelligence is within reach." and described itself as "the world's first straight-shot SSI lab".
  • January 2025, OpenAI. Altman wrote in Reflections: "We are now confident we know how to build AGI as we have traditionally understood it." The next paragraph began: "We are beginning to turn our aim beyond that, to superintelligence in the true sense of the word."
  • July 2025, Meta. In a letter titled Personal Superintelligence, Zuckerberg wrote: "Developing superintelligence is now in sight."
  • Autumn 2025, Future of Life Institute. Its Statement on Superintelligence called for a prohibition on developing superintelligence until there is broad scientific consensus that it will be done "safely and controllably" and "strong public buy-in". Its context note says "many leading AI companies have the stated goal of building superintelligence in the coming decade".
  • November 2025, Microsoft. In Towards Humanist Superintelligence, Suleyman announced "the MAI Superintelligence Team" and described its goal as "systems that are problem-oriented and tend towards the domain specific".
  • April 2026, OpenAI. A policy paper, Industrial Policy for the Intelligence Age, says: "Now, we're beginning a transition toward superintelligence". In the same month Altman's Our principles kept the mission wording, "Our mission is to ensure that AGI benefits all of humanity", and warned that power could be held "by a small handful of companies using and controlling superintelligence".
  • August 2026, Meta. A second letter signed by Zuckerberg, The Future is for Everyone, says "In the next few years, people will be able to use superintelligence beyond human capacity" and notes that "Meta Superintelligence Labs are up and running".

OpenAI's 2023 footnote states the reason for the word directly: to stress a higher capability level than AGI, which is the sense the later texts use too.

How to read a superintelligence claim

Bostrom and the DeepMind authors mean a general system ahead of every human in nearly every field. Microsoft expects "medical superintelligence in the next few years", a single domain that the DeepMind scale would class as narrow. Meta's "personal superintelligence" is a vision of assistants for individuals, and its 2026 letter argues that "There is no such thing as a singular benevolent superintelligence." OpenAI's 2026 definition sets the bar against the smartest humans working with AI.

When a company says it is building superintelligence, start with the definition it uses: general ability across almost all tasks, or superhuman results in one field. Then ask how anyone would check: the DeepMind authors write that unambiguous classification "will require a standardized benchmark of tasks", and give a section of their paper to what such benchmarks would need. Finally, ask who decides what gets released. OpenAI's 2023 paper suggested "something like an IAEA for superintelligence efforts". Meta's 2026 letter says its independent board will have "the power to approve the safety criteria for releasing models". The Future of Life Institute wants a prohibition until its two conditions are met.

FAQ

What is the difference between AGI and ASI?

AGI describes AI that performs at the level of skilled people across most cognitive tasks. ASI describes AI that performs beyond the best people across the same breadth. In Google DeepMind's Levels of AGI framework, "Competent AGI" starts at the 50th percentile of skilled adults, and ASI sits at Level 5, outperforming every human.

What is ASI?

ASI stands for artificial superintelligence: AI that exceeds the best humans across nearly all cognitive tasks. Bostrom's 1998 definition asks for an intellect much smarter than the best human brains "in practically every field". In Google DeepMind's framework it is the top level, marked "not yet achieved".

Does ASI come after AGI?

In the frameworks above, yes: ASI is the level above AGI on the same scale. How long the step would take is disputed. In a survey of four groups of experts in 2012 and 2013, the median estimate was a 10% chance that superintelligence would follow human-level machine intelligence within two years, and a 75% chance within 30 years.

Is ChatGPT an AGI?

Under the DeepMind framework, written in 2023, ChatGPT and similar chatbots counted as Level 1, "Emerging AGI", and Competent AGI was "not yet achieved". In January 2025 Sam Altman wrote that OpenAI was confident it knew how to build AGI, in a post that opened with "As we get closer to AGI". That is a claim about method, and whether a system qualifies depends on the definition and the tests used.

How close are we to superintelligence?

The answers on record are dated claims, made without an agreed test. OpenAI wrote in July 2023 that superintelligence "could arrive this decade". Zuckerberg wrote in July 2025 that it was "now in sight". The Future of Life Institute, which opposes the effort, says leading companies aim to build it "in the coming decade".

The word in the summit's name

Super intelligence is in the name of the Davos Super Intelligence Summit 2027, and the first of its four keynote-led panels is on AGI and agentic AI. The summit is on 19 January 2027, from 16:00 to 20:00 CET, at the Grandhotel Belvedere in Davos. Access is by invitation only. The four panels are set out in the programme, and Why an AI summit in Davos explains how the programme turns the word into four practical questions.

Sources

An evening in Davos, by invitation

  • Date19 January 2027
  • Time16:00 – 20:00 CET
  • VenueGrandhotel Belvedere