Sam Altman's AGI predictions: 5 of 19 came true, 1 missed, 10 still open

Five of Sam Altman's 19 AGI predictions came true, and ten are still open

Altman's deadlines have shrunk as his forecasts piled up. His 2021 essay spoke in windows of five and ten years. His 2024 essay said "a few thousand days". By 2026 he was naming the end of this year. The chart puts each call at its deadline, or at the day it was settled, in a lane for its status.

The pattern is hard to miss. Every call that came true is about what models can do or what they cost: cheaper intelligence, new results in physics and math, an AI research intern inside OpenAI. The one clear miss is about how fast companies would absorb it. Altman agrees. On 26 August 2026 he wrote on X: "i don't think i was wrong on capability timelines; actually quite good on those predictions." The economy was the part he missed: "society and the economy are taking longer to integrate the capabilities than i had guessed."

The two orange dots are calls about OpenAI's own systems in 2026. They're the hardest to check, because only OpenAI can see them.

Every Sam Altman AGI prediction, with its deadline and status

Here are all 19, oldest first, in Altman's own words. "Came true" means a source shows it happened by the deadline, and for the research intern that source is OpenAI itself. "Unclear" means there's no date or no way to measure it. Each status is as of 22 October 2026.

Two choices shape the count. We left out "We are now confident we know how to build AGI as we have traditionally understood it," from January 2025, because it describes what OpenAI knows, not what will happen. And we kept the Y Combinator answer only because so many headlines turned it into "AGI in 2025". Garry Tan asked what Altman was excited about in 2025. Altman named AGI. That's a wish, not a date.

What does Sam Altman say about AGI in 2026?

Sam Altman says OpenAI isn't at AGI yet, but will have a system he'd call AGI by the end of 2026. He told TIME's Alex Heath, in an interview published on 26 August 2026, that the company was "not quite yet" there. TIME reported that by the end of the year, OpenAI "would have an internal system he would call AGI." In the same piece, chief research officer Mark Chen put OpenAI "80% of the way" to AGI.

Nobody outside OpenAI can score that one. The system is internal, and the definition is his. The Decoder put the catch in its headline: AGI by the end of 2026 "if you accept his definition".

OpenAI did publish numbers for the claim it could measure. In September 2026 it said it had hit the research intern goal Altman set the October before. It defines that intern as "a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days." By mid August, the same post says, its research staff ran 3.1 agent workdays for every human workday. It also says this: "We do not yet know how to safely get all the way to aligned, full RSI." RSI is recursive self improvement, the loop where AI builds better AI.

Which of Sam Altman's predictions came true?

Five of Sam Altman's predictions came true, and all five are about cost or capability. The oldest is the cleanest. In March 2021 he wrote that within five years, "computer programs that can think will read legal documents and give medical advice." GPT-4 passed a simulated bar exam two years later. By January 2026, OpenAI said, more than 230 million people a week were asking ChatGPT health questions.

The cost call held up too, and then some. Altman's February 2025 rule was a 10x price drop every 12 months for the same level of AI. In March 2025 Epoch AI measured the drop across six benchmarks at 9x to 900x a year, with a median of 50x. He undersold it.

The 2026 science calls landed early, after one false start. In October 2025 an OpenAI executive posted that GPT-5 had solved 10 open Erdős problems. It had found existing answers in the literature, and Demis Hassabis of Google DeepMind called the episode embarrassing, TechCrunch reported. The real results came in 2026. On 13 February OpenAI published a physics preprint, written with researchers from Harvard, Cambridge and the Institute for Advanced Study, in which GPT-5.2 Pro first conjectured the key formula. On 20 May an OpenAI model disproved an Erdős geometry conjecture from 1946. Thomas Bloom, who runs the Erdős problems site, backed the proof.

The research intern is the fifth. It's also the weakest, since OpenAI is the only one that measured it.

Where has Sam Altman been wrong about AI?

Sam Altman has been wrong about speed outside the lab, and he says so. On David Senra's Founders podcast, released on 23 August 2026, he said: "I think I was wrong about a few things, but in terms of the speed, one of them is the economy just has so much inertia." The forecast he was owning up to dates from 2023. After GPT-4, he said, he thought "very quickly after that there was going to be much more disruption in software businesses being up for grabs right away, than turned out to be," Moneywise reported.

His 2025 agents call sits right on that line. Agents did ship: OpenAI launched Codex in May 2025 and ChatGPT agent in July. But "materially change the output of companies" is a claim about the economy. MIT's NANDA project found in August 2025 that only about 5% of company generative AI pilots were extracting measurable value. We score it unclear, not missed. He wrote "may", and nobody measured company output. Our AI predictions scorecard is harsher and gives it a C minus.

When does Sam Altman think superintelligence will arrive?

Sam Altman puts superintelligence somewhere between about 2028 and 2032. In September 2024 he wrote that it might come "in a few thousand days". Counting 2,000 to 3,000 days from that post lands between March 2030 and December 2032.

He pulled the early end closer in February 2026. At the India AI Impact Summit in New Delhi he said: "On our current trajectory, we believe we may be only a couple of years away from early versions of true superintelligence." Fortune reported the line, along with the data center forecast for the end of 2028.

He's also tied it to 2030. Speaking to Die Welt in September 2025, he said he'd be very surprised if models that can do what people can't didn't exist by then. "To me, that'll start to feel like something we could properly call superintelligence," he added. Every one of these comes hedged: "it is possible", "we may be", "if we are right". None names a test.

How does Sam Altman define AGI?

Sam Altman defines AGI loosely, and he says the term is loose. In February 2025 he wrote that AGI "is a weakly defined term, but generally speaking we mean it to be a system that can tackle increasingly complex problems, at human level, in many fields." A month earlier he told Bloomberg: "I think “AGI” has become a very sloppy term."

OpenAI's charter sets a higher bar: "highly autonomous systems that outperform humans at most economically valuable work." That gap is why the AGI rows are the hardest to score. A deadline only counts if the finish line stays put. The cost, science and intern calls named something you can measure, so they're the ones we could mark true.

Dario Amodei runs on a different clock, and we scored his calls the same way in Dario Amodei's predictions. Sam Altman vs Dario Amodei sets the two CEOs side by side, and why Dario Amodei left OpenAI covers how their split began.