Who Gets to Say AI Has Reached AGI? OpenAI Says It Is 80 Percent There

AGI is the point where artificial intelligence outperforms humans at most economically valuable work — and OpenAI now grades itself 80 percent of the way there.

OpenAI grades itself 80 percent of the way to AGI

On 26 August 2026, TIME published “Inside OpenAI’s Reboot,” a profile by Alex Heath built on two weeks inside OpenAI’s offices and dozens of hours of interviews with more than 20 leaders, employees, investors and rivals. The headline claim: OpenAI is not quite at AGI yet, but by the end of 2026 the company expects an internal system Sam Altman would call artificial general intelligence.

The number came from OpenAI’s own research chief. Chief Research Officer Mark Chen says OpenAI is “80% of the way” to AGI, the-decoder reported, echoing the TIME account. The bar OpenAI grades against is OpenAI’s charter definition of AGI: “highly autonomous systems that outperform humans at most economically valuable work.”

OpenAI is 80% of the way to AGI.

— — Mark Chen, Chief Research Officer, OpenAI

What the 80 percent rests on: the Astra model

The confidence rests on a forthcoming model family called Astra. Chief Scientist Jakub Pachocki describes Astra as able to take a research idea, “implement it inside OpenAI’s code base, run the experiment, and return results.” In a customer preview, Altman went further, casting Astra as the moment machine research starts to compound.

I expect this will be the first model where the model actually invents new things in a way that matters. That’s a very AGI-like thing.

— — Sam Altman, OpenAI

President Greg Brockman framed the stakes in retrospective terms, saying the moment, viewed from two years out, may be “remembered as the moment AGI was created.” Reporting from BigGo Finance tracked the same year-end internal-AGI target and Astra’s autonomous-research pitch.

Who gets to hold the referee’s whistle?

Here is the load-bearing problem, and capability is not the crux. The milestone, the definition, and the grade all sit inside one company. OpenAI wrote the charter definition of AGI. OpenAI’s Chief Research Officer supplies the 80 percent. And the “internal system” would be declared AGI internally — the threshold reached and marked by the same party, with no external referee checking the work. When one lab owns the exam, the marking scheme and the score, “AGI” stops being a scientific finding and becomes a corporate announcement.

No scientific consensus exists in 2026 on what the AGI threshold is or how to measure the threshold. Altman knows the problem better than most. Altman has called AGI “not a super useful term” because everyone defines the word differently — a point CNBC documented when experts agreed the label had gone slippery. The chief executive promising year-end AGI is on record that the phrase is nearly meaningless.

Two definitions grading different exams

The competing definitions do not measure the same thing. OpenAI’s bar is economic: outperform humans at most economically valuable work. DeepMind co-founder Demis Hassabis sets a cognitive bar — a system that can do “pretty much any cognitive task that humans can do” — as MindStudio’s survey of the disagreement lays out. An economic exam and a cognitive exam produce different pass marks, so “80 percent” answers a question only OpenAI has asked.

Timelines diverge as sharply as definitions. Andrej Karpathy places AGI roughly a decade out — a “decade of agents” — against OpenAI’s year-end claim. Melanie Mitchell and other researchers question whether the term still carries meaning at all. The goalpost-moving critique cuts both ways, as an essay on the milestone that disappears and a Science analysis both note: sceptics say the bar keeps rising, while boosters say a permissive definition means we are already inside an early-AGI period. An academic survey of the AGI question reaches the same verdict — no agreed test, no agreed date.

Set the corporate scorecard beside other frames on this site and the contrast sharpens. Forecaster Daniel Kokotajlo reaches a 70 percent probability of AI transforming society by argument and evidence, not by self-assessment. A separate case holds AI needs a body to reach general intelligence at all — a definitional objection OpenAI’s economic bar sidesteps entirely.

The dignity question at the centre of AGI

Who holds the authority to declare a species-level threshold crossed? Ubuntu asks for clear eyes, and the honest answer is simple: no single company should. A claim so large about machine capability — and, by implication, about the worth of the human work AGI is meant to exceed, the “most economically valuable work” in OpenAI’s own phrase — needs an external, plural, contestable standard. Not one lab’s internal benchmark and a round number from its research chief.

Here is where Emergent Intelligence (EI) — the dignity-first frame I use for what is more commonly called AI — parts company with the announcement. EI treats a threshold so consequential as a civic question, decided in the open, rather than a launch metric. OpenAI’s own charter language about benefiting everyone only sharpens the point: a benefit owed to everyone cannot be certified by one party grading its own homework. The same lab’s recorded walk-backs on risk are reason enough to want a referee outside the building.

When the definition, the metric and the scorecard all sit inside one lab, “AGI” stops being a discovery and becomes a press release.

Frequently Asked Questions

These are the questions people are asking about the AI and AGI debate. Short answers follow, drawn from the August 2026 TIME profile and the wider research literature on artificial general intelligence.

What is AGI?

In short, AGI is artificial general intelligence — and no single definition commands agreement. OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work,” while research from DeepMind frames AGI as matching human cognition across tasks. Analysis shows the two bars measure different things.

How does OpenAI measure being 80 percent to AGI?

Simply put, OpenAI measures progress against OpenAI’s own economic definition, and Chief Research Officer Mark Chen supplies the estimate. According to the TIME profile, the 80 percent figure rests on the forthcoming Astra model, which data and demos suggest can run research experiments inside OpenAI’s codebase autonomously.

Why is the AGI definition debate significant?

The key is authority. Analysis shows no external referee certifies an AGI claim, so when one company owns the definition, the metric and the grade, evidence of a milestone becomes indistinguishable from marketing. Research across the field finds no scientific consensus on the AGI threshold.

Who is declaring AGI, and who should?

In other words, OpenAI proposes to declare an internal AGI system by the end of 2026, graded against OpenAI’s own bar. Evidence and precedent suggest a plural, external standard should hold such authority instead — the argument at the heart of Emergent Intelligence.

What are the risks of a company grading its own AGI?

The answer is captured standards. Data and reporting reveal a company can set a permissive economic bar, meet the bar, and announce a species-level milestone with no independent check. According to the debate literature, competing definitions and shifting goalposts make any self-certified AGI claim contestable by design.

Sources:

TIME — Inside OpenAI’s Reboot · the-decoder — AGI by end of 2026 on OpenAI’s definition · Forbes — AGI by year-end amid a safety crisis · BigGo Finance — internal AGI and Astra · CNBC — Altman: AGI a pointless term · MindStudio — Hassabis, Altman and LeCun disagree · Lviv Herald — the milestone that disappears · Science — measuring general intelligence · arXiv — what AGI actually means · Related on this site: AI as a new species: Kokotajlo’s 70% warning · The body gap: why AI needs a body to reach AGI · OpenAI’s ‘benefit everyone’ AGI framing · The lab split on existential risk