“Highly autonomous systems that outperform humans at most economically valuable work.” That single sentence, tucked into OpenAI’s charter, has shaped more of the AGI debate than any paper the company has published. It is the definition investors argue over, the one regulators quote back at executives, and the one that decides whether a multi-billion-dollar licensing deal covers a given model or not.
It is also deliberately slippery. Understanding why OpenAI chose those words — and what it has built around them — tells you more about the race to AGI than any benchmark chart.
What OpenAI Actually Means by AGI
The charter language is broad on purpose. Academic definitions tend to be either impossibly strict (an agent that can do every task a human can, in any environment) or so loose that a decent chatbot qualifies. OpenAI’s version sits in between: most economically valuable work, not all of it. That “most” is where the arguments live.
Compare it with how the rest of the field talks. Google DeepMind researchers lean on generality and transfer — can a system pick up a novel task without retraining? Anthropic tends to describe frontier capabilities in terms of risk thresholds rather than a finish line. OpenAI’s framing is unapologetically economic. It is a definition written by people who expected to sell the output.
That matters because the definition is not just marketing. Inside OpenAI’s corporate structure, it functions like a contractual switch.
The Clause That Forced the Capped-Profit Structure
OpenAI started in December 2015 as a nonprofit with a pledge to develop AGI “for the benefit of humanity.” By 2019 it needed compute money that donations could not supply, so it created OpenAI LP — a capped-profit entity where investors’ returns are limited (originally 100x) and where the nonprofit’s mission stays in charge.
The Microsoft agreement that followed contains a detail most people skim past: Microsoft’s license to OpenAI’s technology covers everything except AGI. Once the board determines that AGI has been achieved, those systems fall outside the commercial deal, and the determination is meant to be verified by an independent panel of experts rather than by Microsoft or OpenAI alone.
Why the 2023 Board Fight Was Really About AGI
When the board fired Sam Altman in November 2023 and reinstated him five days later under a restructured board, the public argument was about governance and communication. Underneath sat a question with enormous financial weight: who holds the authority to declare that AGI has arrived, and what happens to the company’s obligations when they do? The current structure — a nonprofit foundation controlling a public benefit corporation — answers that question in a way that keeps the trigger inside the mission side of the house, at least on paper.
OpenAI’s Five Levels, and Where It Says It Is
In 2024 OpenAI shared an internal scale that gets cited constantly:
- Level 1: Conversational chatbots (ChatGPT, 2022)
- Level 2: Reasoners that can think through problems before answering (the o-series models)
- Level 3: Agents that can execute multi-step tasks over hours or days
- Level 4: Innovators capable of original research
- Level 5: Organisations — systems that can run an entire operation
Altman has said publicly that the company sits somewhere around level 2, edging into 3. That tracks with what shipped: reasoning models that spend more compute deliberating, then agentic tools that browse, book, code and file things on your behalf. The ladder is genuinely useful shorthand, though it is also a communication device — each rung sounds inevitable once the previous one is named, which is not how research usually works. If you want a sharper eye for these staged reveals, reading AGI roadmap updates critically is a skill worth building now.
The Timeline: “A Few Thousand Days”
In a September 2024 essay called The Intelligence Age, Altman wrote that superintelligence could arrive “in a few thousand days” — roughly five to ten years. In interviews since, he has shortened the horizon, describing AGI as plausible within the next few years and predicting that agents will meaningfully enter the workforce before the decade turns. The company’s stated mission page is blunter still, describing its aim as ensuring that AGI benefits all of humanity and warning that the transition may be abrupt.
Treat those dates as directional, not predictive. The gap between 2027 and 2100 is almost entirely a gap in assumptions about compute scaling, algorithmic progress, and how expensive each additional percentage point of capability turns out to be. That’s the argument laid out in why credible AGI timelines differ so wildly, and it explains why the same company can sound confident on a podcast and cautious in a filing.
The Compute Bill Behind the Promise
AGI claims are cheap. Training runs are not. OpenAI has committed to infrastructure at a scale no software company has attempted: the Stargate venture announced in January 2025 with SoftBank, Oracle and MGX, framed as up to $500 billion over four years, with the first sites breaking ground in Texas. Add roughly $13 billion of Microsoft investment, a multi-year Azure agreement, and custom silicon work, and you get a company betting that capability follows wattage.
That bet is why the AGI question is inseparable from money. Whoever reaches a system that qualifies as economically transformative first gets to set prices, define safety norms, and write the licensing terms everyone else signs. The dynamics of that scramble — who is genuinely ahead and what the real prize is — are worth understanding separately from the headline dates; the AGI competition is not a single race, and the leaders keep changing depending on which axis you measure.
The Safety Record Is Where the Story Gets Uneven
OpenAI’s safety apparatus has been reorganised repeatedly. The Superalignment team, announced in 2023 to work on controlling systems smarter than humans, was effectively dissolved in May 2024 after its leaders, Ilya Sutskever and Jan Leike, left — Leike going on to say that safety culture had taken a back seat to shipping product. A Safety and Security Committee was formed in its place, and the Preparedness Framework assigns risk ratings to new models before release.
None of that is scandalous on its own; research organisations restructure constantly. But it does mean that when OpenAI talks about AGI readiness, the verification is largely internal. External evaluations exist, and third-party red-teaming has grown, yet the people deciding whether a model is safe enough to release are employed by the company releasing it. Distinguishing real safety signals from corporate noise is now a discipline in itself — this guide to reading AGI safety news is a reasonable place to calibrate.
How to Read the Next Big OpenAI Announcement
When the next model lands with a launch video and a benchmark table, here’s a short checklist that cuts through most of it:
- Which benchmarks, and who ran them? Vendor-reported scores on saturated tests mean little; independent evaluations on unfamiliar tasks mean a lot.
- What does “economically valuable” mean in practice? Ask what percentage of a real job the system completed, over how many hours, with what human supervision.
- Was it an agent or a demo? Long-horizon reliability degrades fast. A five-minute task that works 90% of the time is very different from a five-hour task at 90%.
- Did the board actually declare anything? A blog post saying “AGI is near” is not the same as the governance trigger being pulled.
That last point is the one most coverage misses. There is a difference between a company claiming AGI and a company declaring it, and only the second has legal and financial consequences. For a closer look at what separates a genuine leap from a polished reveal, what to watch beyond the demo reel covers the gap between stagecraft and capability.
The Real Deciding Vote Isn’t Technical
Somewhere in OpenAI’s future is a meeting — probably a small one, probably not filmed — where a group of people decides whether a system in front of them meets the standard written into the charter in 2018. That meeting will be judged on benchmark results, internal evaluations, deployment plans and external pressure, and it will be enormously difficult to hold to any objective standard, because “most economically valuable work” was never defined precisely enough to be a test.
The honest read on OpenAI’s AGI project is that the science is remarkable, the timelines are marketing as much as forecast, and the governance question — who gets to say it happened — remains the weakest joint in the whole structure. Watch the structure, not the press release. When that meeting finally happens, the interesting story won’t be the model they show. It will be whether anyone outside the room believes the verdict.