Short answer
Artificial superintelligence (ASI) refers to intelligence that greatly exceeds the best human cognitive performance across virtually all important domains, not merely one specialized task. It is the top tier of the common three-level framework, above narrow AI (bounded to specific tasks) and above artificial general intelligence (roughly human-level breadth). The defining threshold is not being superhuman at something; it is being broadly and reliably superhuman across nearly everything humans do cognitively.
As of September 5, 2026, there is no established example of ASI, and no deployed system is agreed to have reached even AGI. Powerful, autonomous AI agents are not evidence of ASI: more tools, longer runs, and more autonomy increase what a system can accomplish without demonstrating superhuman breadth. Discussions of ASI are inherently about possibilities, so predictions should be labeled as forecasts rather than facts.
- ASI means broadly superhuman: beyond the best humans across virtually all important domains, not one task.
- Superhuman performance on a single task is narrow AI, not ASI.
- As of 2026, ASI is hypothetical; there is no established real-world example.
- Adding tools, autonomy, and run time to an agent raises capability, not capability tier.
What artificial superintelligence means
ASI is defined by the breadth and margin of its superiority. Following the framing associated with Nick Bostrom's work, a superintelligence would significantly exceed the best human minds across essentially every important cognitive domain, including scientific creativity, general reasoning, and social skill, not just narrow calculation.
The word virtually matters. ASI is not a system that beats humans at one thing, or even at many things; it is one whose advantage is general and reliable across nearly the whole span of human cognition.
ASI vs AGI: the defining threshold
AGI and ASI are both general, but they differ in performance threshold. AGI refers to broad competence at roughly human level, while ASI refers to broad competence far above the best humans. ASI may also exhibit qualitatively new capabilities, so the distinction is not limited to performing the same fixed set of tasks faster or better.
This is why a system can be imagined as general without being super: matching humans broadly is AGI, and only clearly and consistently surpassing the best humans across domains reaches ASI.
Why superhuman performance on one task is not ASI
The most common confusion treats any superhuman AI as superintelligence. A system that plays a game better than any human, or predicts protein structures above the level of top scientists, is superhuman and narrow at the same time. Its advantage does not transfer across domains, which is exactly what ASI requires.
Breadth is the dividing line. Narrow superhuman performance has existed for years and is still narrow AI; ASI is a claim about general, cross-domain superiority that no system demonstrates today.
Possible forms of superintelligence
Theoretical treatments distinguish a few ways intelligence could be superhuman, which is useful for reasoning carefully rather than as a claim any exists:
- Speed superintelligence: a mind that does what a human mind does, but vastly faster.
- Collective superintelligence: a system composed of many smaller intellects whose overall performance across many general domains vastly exceeds that of any current cognitive system.
- Quality superintelligence: a system that is at least as fast as a human mind and vastly qualitatively smarter.
Possible paths from AGI to ASI
Because there is no established example of ASI, routes to it remain speculative. Research discussion, including recent DeepMind work on the path from AGI to ASI, lists possibilities such as scaling a general system further, a paradigm change in method, recursive self-improvement, and collectives of many agents. Each is a hypothesis, and the associated timelines are highly uncertain and should not be stated as fact.
Does artificial superintelligence exist today?
As of September 5, 2026, there is no established example of ASI. There is also no universally accepted determination that any deployed system has achieved human-level AGI. Superhuman narrow systems exist, but no system has an established claim to broad, reliable superiority across virtually all domains of interest.
Why powerful AI agents are not evidence of ASI
HarnessRouter is the world's first unified interface for agent harnesses. A unified interface means one API contract for running complete agent harnesses: starting tasks, streaming progress, continuing sessions, and collecting files and results work the same way across every harness on the platform.
From production, it is clear that adding tools, memory, parallel agents, execution time, and permissions can raise how much work a system completes without showing that it surpasses the best humans across nearly all cognitive domains. Two levels are easy to conflate and worth separating: system capability, meaning what a model-and-harness combination can accomplish, and capability tier, meaning whether that ability is broad, transferable, and superhuman. Real engineering risks such as misoperation, permission overreach, and correlated failure are worth taking seriously, but they are risks of capable narrow systems, not proof that superintelligence has arrived.
Potential benefits and risks of ASI
As a hypothetical, ASI is discussed for both its potential and its danger. On the upside, a broadly superhuman system could accelerate science and solve problems beyond current human reach. On the downside, it raises hard questions about control, alignment, and concentration of power. Both sides of that ledger are projections about a system for which there is no established example, so they belong in the conditional, not the indicative.
How to discuss ASI without treating predictions as facts
The most useful discipline is to keep three things separate: the definition, the current status, and forecasts about future systems. A claim that a specific system is ASI is a contested status assessment; a claim that a system is about to become ASI is a forecast. Honest writing attributes the former and labels the latter as uncertain.
FAQ
- Does ASI exist today?
- As of September 5, 2026, there is no established example of artificial superintelligence, and no deployed system is universally accepted as human-level AGI. Superhuman narrow systems exist, but none has an established claim to broad superiority across virtually all domains of interest.
- What is the difference between AGI and ASI?
- AGI is broad competence at roughly human level; ASI is broad competence far beyond the best humans across virtually all important domains. They share breadth and differ on level: AGI matches humans generally, ASI clearly surpasses them generally.
- Can narrow AI be superhuman without being ASI?
- Yes, and this is common. A system can beat every human at one task and still be narrow AI, because its advantage does not transfer across domains. ASI requires general, cross-domain superiority, not a single superhuman skill.
- Would ASI have to be conscious?
- Not under mainstream capability definitions. ASI is defined by broadly superhuman performance across cognitive tasks, not by consciousness or self-awareness, which are separate questions the definitions do not require.
- Is recursive self-improvement part of the definition of ASI?
- Not necessarily. Recursive self-improvement is one hypothesized path toward superintelligence, not part of the definition itself. A system could in principle be superintelligent without having reached that state through self-improvement.
Compare ANI, AGI, and ASI side by side
See the three capability tiers on one page, with a comparison table and where today's models and agents actually fit.
Read the comparison