When Will Superintelligence Arrive? CEO Forecasts (2026)

Searches for "when will superintelligence arrive" spiked after September 22, 2026, when President Trump told the UN General Assembly that US government documents would refer to AI as "Super Intelligence (SI)."

Here is the short answer. As of September 23, 2026, there is no confirmed date for superintelligence. Public forecasts from the people building it range from one to two years (Anthropic) to five to twenty years (Safe Superintelligence). The Metaculus community forecast for the first announced general AI sits at a median of April 23, 2031 (retrieved September 23, 2026). And Trump's "SI" is a renaming of today's AI, not an announcement that superintelligence has arrived.

The short answer: forecasts span 2027 to the mid-2040s

Ask five lab leaders and you get a fifteen-year spread. Dario Amodei of Anthropic said in February 2026 that a "country of geniuses in a datacenter" is one to two years, maybe one to three years away. Ilya Sutskever, who left OpenAI to work only on superintelligence, said in November 2025 that it is five to twenty years away. Counting from each statement, that is 2027–2029 versus 2030–2045.

The gap is not a measurement error. It is a definition problem: AGI (human-level breadth) and ASI (decisively beyond human across essentially every domain) are different milestones, and the two words get used interchangeably in headlines. We separate them in our breakdown of AGI, ASI and SI.

Superintelligence timeline: who said what, and when

Most "when will AGI arrive" roundups quote CEOs without dates. That matters, because a 2024 quote and a 2026 quote carry very different weight — several 2024 forecasts have already expired. Every row below comes from a primary source we read in full: the speaker's own blog, essay, or a dated on-record interview.

Dario Amodei: the shortest timeline on record

On the podcast published February 13, 2026, Amodei put the "country of geniuses in a datacenter" at one to two years, maybe one to three, and said he is 90–95% confident it happens within ten years. He also volunteered how wrong he has been before: in Machines of Loving Grace (October 2024) he wrote that powerful AI "could come as early as 2026," and he now says he "wouldn't be surprised" to be off by a year or two. That caveat rarely survives into the headlines that quote him.

His January 2026 essay The Adolescence of Technology keeps both tails open: "as little as 1–2 years away, although it could also be considerably further out."

Sam Altman: no year, but "the takeoff has started"

Altman is the outlier who refuses to name a year. In Reflections (January 2025) he wrote that OpenAI is "now confident we know how to build AGI" and is "beginning to turn our aim beyond that, to superintelligence in the true sense of the word." In The Gentle Singularity (June 2025) he wrote that "humanity is close to building digital superintelligence" and predicted that 2026 would bring systems that generate novel insights.

Exactly one year ago today, Abundant Intelligence (September 23, 2025) swapped dates for supply: a "factory that can produce a gigawatt of new AI infrastructure every week." In practice, OpenAI statements are easier to read as compute commitments than as calendar commitments.

Demis Hassabis: around 2030, possibly 2029

Speaking to Axios on May 26, 2026, Hassabis said he "still broadly expects AGI around 2030, though he now sees 2029 as a possibility," and described the current agent wave as "a practice run." His is the most recent mid-range estimate in the table, which makes it the most usable single number for planning.

Elon Musk: end of 2026 for "smarter than any human"

At Davos on January 23, 2026, Musk said AI "smarter than any human" could arrive by the end of 2026, "no later than next year," and that by "2030 or 2031, AI will be smarter than all of humanity collectively." Read this alongside his track record: the "smarter than any human next year" claim has been rolled forward across multiple years, so the deadline itself keeps moving.

Ilya Sutskever: five to twenty years

Asked on the Dwarkesh Podcast published November 25, 2025 when a system that learns as well as a human would arrive, Sutskever answered "I think like 5 to 20." His reasoning is that today's recipe is not enough — the field needs to move from static pre-training to continual, on-the-job learning. The person who founded a company to build nothing but superintelligence holds the longest timeline in this table. That is the single most overlooked data point in the debate.

Mark Zuckerberg: no year, but a decisive decade

Meta's Personal Superintelligence letter (July 2025) says "developing superintelligence is now in sight" and that "the rest of this decade seems likely to be the decisive period." Like Altman, Zuckerberg declines to name a year. We compare how each lab frames the goal in our roundup of the companies building superintelligence.

What prediction markets say (as of September 23, 2026)

Lab leaders have fundraising and sales incentives attached to short timelines, so it is worth checking forecasters with money on the line. All figures below were retrieved on September 23, 2026 and move daily.

Metaculus: median April 23, 2031

Metaculus's "When Will the First General AI Be Announced?" is one of the longest-running AGI questions anywhere. On September 23, 2026 the community forecast was a median of April 23, 2031, with a 25–75% interval of May 2, 2028 to December 2037.

Resolution requires clearing four separate criteria, including a Turing-test-style interview and a robotic manipulation task. So it measures an announced AGI, not the arrival of superintelligence.

Polymarket: 17.5% that OpenAI declares AGI before 2027

On September 23, 2026 the only live AGI-timing market on Polymarket with meaningful volume was "OpenAI announces it has achieved AGI before 2027?" It was trading at 17.5% Yes on roughly $271,000 of volume.

Note what it actually asks: whether OpenAI itself declares that it has reached AGI. It is a corporate-announcement market, not a capability market. Quoting 17.5% as "the odds superintelligence arrives" would be a misreading.

Three ways these numbers get misread

• The resolution criteria drift from the question. Both markets resolve on an announcement of AGI, not on superintelligence existing.
• Short deadlines look bearish. Polymarket's 17.5% covers only "before 2027." Ask about the 2030s and the number would be far higher.
• Thin liquidity. At roughly $271,000 of volume, a handful of large traders move the price. Do not treat it like an election market.

Trump's "SI" rename is not superintelligence arriving

In his September 22, 2026 UN General Assembly address, President Trump said AI would henceforth be called "Super Intelligence." The White House release quotes him saying it is "hereinafter officially called 'Super Intelligence.'"

This is a naming change for existing systems, not a capability announcement. The scope is US government documents; it carries no force over private companies, other governments, or academic usage. As of September 23, 2026 we could not confirm any executive order implementing it.

In the research literature, superintelligence follows Nick Bostrom's definition — intellect that greatly exceeds human cognitive performance in virtually all domains of interest. Today's ChatGPT, Claude, and Gemini do not meet it. We cover the announcement itself in our explainer on Trump renaming AI to SI and the terminology in what "SI" and superintelligence actually mean.

Why the forecasts disagree by fifteen years

In practice, the spread comes down to three things.

1. Different targets. Amodei's "powerful AI," Hassabis's "AGI," and Sutskever's "human-level learning efficiency" are not the same milestone. Align the target before comparing the dates.
2. No accepted test. There is no agreed exam that declares superintelligence has arrived. Metaculus stacking four criteria is a symptom of that, not a solution.
3. Incentives. Short timelines help with fundraising and recruiting. It is telling that Sutskever, whose company had not released a public model as of September 23, 2026, gives the longest range.

A three-question test for business leaders

For most companies, being right about the date is worth very little. These three questions are worth more.

1. What exactly is being forecast? AGI or ASI, a research result or a shipped product. If the "what they are forecasting" column differs, the dates were never comparable.
2. Would you be fine if the forecast is wrong? Anything that pays back inside twelve months — document drafting, support triage, internal search — can be decided today, independent of any superintelligence timeline.
3. Are you deciding on "when" or on "what today's models can do"? Current capability is easier to verify than any forecast. Our comparison of ChatGPT, Claude and Gemini is built entirely from what shipped.

The honest caveat: if Amodei's one-to-three years is right, the impact lands on headcount planning and process design, not on tool procurement. So the durable move is to map which of your workflows would be displaced first. That work pays off no matter which forecast wins.

Summary

There is no confirmed date for superintelligence as of September 23, 2026. Lab leaders' public forecasts run from one to two years out to five to twenty, and the Metaculus median for an announced general AI is April 23, 2031. Trump's "SI" is a change of vocabulary, not a change of capability.

Mihata runs monthly AI strategy sessions where we help teams separate confirmed facts from forecasts and decide what is actually worth adopting now. We are not trying to sell you on a timeline — sorting the signal from the noise is simply hard to do alone, and we think this is a useful service. If that sounds relevant, we would be glad to talk.