AGI vs ASI vs 'SI': The Difference, in One Table (2026)

AGI, ASI and "SI" get thrown into the same sentence constantly, but they are not the same kind of word. AGI and ASI are capability tiers. "SI" is a renaming of the word "AI" itself, announced by President Trump at the UN General Assembly on September 22, 2026. It is a label, not a tier.

One number makes the hierarchy concrete. In Google DeepMind's 2023 framework "Levels of AGI," AGI is split into five levels, and ASI sits at the top as Level 5 — defined as outperforming 100% of humans. Everything from Level 2 upward is marked not yet achieved, and frontier models like ChatGPT were placed at Level 1, "Emerging AGI" (arXiv:2311.02462, ICML 2024). So by the research framing, today's AI is at the entry level of AGI, and nobody has built ASI.

Below is a single table covering the term, its source definition, who proposed it, and whether current AI qualifies — plus a four-question test for deciding which term belongs in your own writing.

AGI vs ASI vs SI: one comparison table

Only definitions traceable to a primary source are included.

That last column is the whole confusion in one line. AGI and ASI are yardsticks, so "does X qualify" is a valid question. SI is a relabeling, so asking whether today's AI "is SI yet" makes no sense. Trump's SI did not inherit the meaning of the existing academic term ASI, despite the overlapping words.

What AGI means: the OpenAI Charter threshold

The most-cited definition of AGI comes from OpenAI's Charter:

OpenAI's mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity. (OpenAI Charter)

Note the threshold: "most economically valuable work." It is an economic bar, not a benchmark score. DeepMind's paper maps this OpenAI threshold onto Level 4, "Exceptional AGI" — at least the 99th percentile of skilled adults — which is considerably higher than where most casual "AGI is near" claims land.

In practice, this is where arguments go sideways. A writer using OpenAI's economic bar and a writer using DeepMind's Level 2 (50th percentile of skilled adults) are describing capability levels two tiers apart while using the same three letters. If you cite an AGI definition, name the source and year in the same sentence. That one habit resolves most of these disputes before they start.

What ASI means: Bostrom's bar and DeepMind's Level 5

ASI traces back to philosopher Nick Bostrom's definition of superintelligence:

A superintelligence is any intellect that … vastly outperforms the best human brains in practically every field, including scientific creativity, general wisdom, and social skills. (Nick Bostrom, Ethical Issues in Advanced Artificial Intelligence)

The comparison class is the best human brains, across practically every field — not the average person, and not one domain. Bostrom expanded the concept into the 2014 book Superintelligence: Paths, Dangers, Strategies, which is the foundation for most ASI discussion today.

DeepMind operationalized it: Level 5 "Superhuman" means outperforming 100% of humans. Crucially, the paper distinguishes narrow from general at that level. Level 5 Narrow AI already exists — AlphaFold for protein structure prediction, AlphaZero and Stockfish for games. Level 5 General AI, which is what ASI means, is marked "not yet achieved."

Keep that pairing handy: AI that beats every human at one task exists; AI that beats every human at everything does not. When someone claims ASI has arrived, the first question is which of the two they mean.

The five Levels of AGI, and where models sit today

DeepMind classifies systems on two axes — performance (depth) and generality (breadth). Here is the general-purpose column.

The narrow column tells the other half of the story: it has real systems at every level including Level 5, while the general column is empty from Level 2 down. The hard part is not raising performance — it is raising performance without losing breadth.

One versioning note: the paper renamed Level 4 from "Virtuoso" to "Exceptional" in v5 (September 2025). Older articles still say "Virtuoso AGI," so check which version a citation refers to.

Where Trump's "SI" fits: a rename, not a tier

On September 22, 2026, at the 81st UN General Assembly, President Trump announced that AI would henceforth be called "Super Intelligence (SI)." The White House release quotes the line directly:

The United States totally rejects any attempt to construct a globalist scheme of control for the Artificial Intelligence being spoken of so much now — hereinafter officially called "Super Intelligence." (The White House)

The operative phrase is hereinafter officially called. No new capability threshold is being defined; an existing word is being swapped out. The stated reasoning in the speech was that "artificial" sounds like "fake."

So SI does not belong as a row beneath ASI in a capability ladder. The correct mapping is:

• ASI — an academic term for a future milestone (Bostrom; DeepMind Level 5)
• SI — a new name for AI in general, including today's Level 1 systems
• SI is not an abbreviation of ASI, nor a subdivision of it

Scope matters too. The speech referred to "all United States documents." It carries no force over private companies, other governments, or academic terminology. As of September 23, 2026, we have not been able to confirm any executive order or statute formalizing the change — worth stating plainly rather than assuming it is settled. For the announcement itself, see our breakdown of Trump's UN speech renaming AI to SI, and for the term's meaning, what "Super Intelligence (SI)" actually refers to.

Four questions before you use any of these terms

Whether you are writing a proposal, an internal policy, an investor update or a blog post, running these four checks prevents most of the rework we see.

1. Are you describing a tier or a label?

Use AI → AGI → ASI for capability progression. Use SI only when discussing US government document terminology. Never mix the two in one paragraph — that single rule prevents most of the confusion.

2. Whose definition are you citing?

OpenAI's Charter, DeepMind's Level 2, and Bostrom's "best human brains" bar sit at very different heights. Name the source and the year alongside the claim.

3. Have you checked whether current AI qualifies?

The most common internal error is treating ASI as "today's AI, but better." General ASI is unachieved by every primary source available. Don't use ASI to describe a shipping product; if you use it, flag it explicitly as a hypothetical future stage.

4. Who is reading it?

For documents filed with or read by US agencies, matching their terminology may be reasonable. Elsewhere, "SI" collides with established meanings — system integration and the International System of Units (SI units) among them. Spell it out on first use: "SI (Super Intelligence)."

In practice, deciding which tier of AI belongs in which workflow matters far more than what the technology is called. Mihata builds custom AI solutions and runs monthly AI strategy sessions for teams working through exactly this kind of question. If that sounds useful, we would be glad for you to take a look.

Going further

The natural follow-up question is when superintelligence might actually arrive. We collected CEO forecasts and prediction-market numbers in our look at when superintelligence is expected.

Summary

AGI is the tier where a system handles most human work at human level. ASI is the tier where it surpasses every human across essentially every domain. SI is a new name for AI as a whole. The first two are yardsticks; the third is a label. Hold that distinction and any mixed-up article or press release becomes readable.

If you are working through AI terminology for your own documents, or figuring out where AI fits in your operations, feel free to get in touch — we are happy to start with a conversation about your current setup.