Ask five people at OpenAI when AGI arrives and you’ll get five answers, none of them binding. That isn’t quite evasion. It’s what happens when a company is simultaneously trying to build the thing, sell the story of the thing, and keep its safety staff from quitting over the story of the thing.
Still, the public record is thicker than most people assume. Executives have named years. The charter has a definition. Internal capability scales have leaked. If you know where to look, the OpenAI AGI timeline is less of a mystery than a set of competing claims made by people with different incentives.
The dates OpenAI has actually put on the record
None of these are deadlines. All of them are data points.
- 2026. Kevin Weil, then OpenAI’s chief product officer, said in a 2024 interview that the company would reach AGI “by 2026, maybe.” He framed it as a guess, and it was reported everywhere as a promise.
- “A few thousand days.” Sam Altman’s September 2024 essay The Intelligence Age put superintelligence roughly that far out. Do the arithmetic and you land somewhere between 2029 and 2035, depending on whether you count calendar days or workdays.
- “The very near future.” At Davos in January 2025, Altman said AGI was coming soon and would be “less disruptive than people think,” which is a strange pair of sentences to hold in the same mouth.
- No date at all. OpenAI’s official posture, in blog posts and policy filings, has consistently been that AGI is uncertain and that timelines are speculative. This is the position its lawyers prefer.
Notice the pattern. The people closest to the product give the shortest timelines. The people closest to the charter give the longest. Both groups are describing the same roadmap.
The 14 words doing all the work
OpenAI’s charter defines AGI as “highly autonomous systems that outperform humans at most economically valuable work.” That’s it. No benchmark, no threshold, no test you could run on a Tuesday and get a pass or fail. The full breakdown of OpenAI’s 14-word AGI definition is worth reading closely, because every timeline claim inherits its ambiguity.
“Most economically valuable work” could mean the median job, or the aggregate output of the economy. “Outperform humans” could mean a single skilled worker, or the best human in the field. Those are wildly different bars, separated by perhaps a decade of research. When an OpenAI executive says 2026, they are usually pointing at the lower bar. When a critic says 2040, they are usually pointing at the upper one.
The five-level ladder, and where the rungs actually sit
In July 2024, Bloomberg reported an internal OpenAI scale with five levels: Chatbots, Reasoners, Agents, Innovators, Organizations. The company told reporters it sat at level one and was about to reach level two.
What each rung means in practice
Level one is a conversational model. Level two is something that can reason through a problem rather than pattern-match an answer. That reasoning shift arrived with the o-series models in late 2024, which is why OpenAI quietly stopped talking about the ladder afterward. Level three is agents that can carry out multi-step work unsupervised for extended periods. Level four is systems that generate new scientific or technical ideas. Level five is an entity that can run an entire operation.
The honest read: the first two rungs had public benchmarks attached. The last three don’t. Nobody can tell you what score on what test counts as “Innovator.” So the ladder is useful for tracking direction and useless for dating arrival.
Why the OpenAI AGI timeline keeps sliding
Three forces move the number, and only one of them is science.
Compute and power are the pacing item
Stargate, the $500 billion infrastructure venture announced in January 2025, is a useful reality check. A single large data centre campus takes three to five years from groundbreaking to full load. Grid connections, turbine orders, and transformer lead times are measured in years, not quarters. If the model research finished tomorrow, the physical capacity to train and serve the next generation of systems would still arrive on a construction schedule. That’s the quiet argument for timelines stretching past 2027. The assumptions behind the 2027 and 2100 predictions mostly come down to how fast you think concrete gets poured.
The target moves when you approach it
Every capability that gets solved stops counting. Reading comprehension, grade-school math, code generation, competitive programming: each one was a milestone until it wasn’t. Models now saturate the benchmarks they were built to beat, so researchers build harder ones, and the finish line relocates. This is a genuine measurement problem, not a conspiracy, but it does mean the goalposts have moved in one direction for a decade.
Incentives pull in both directions
A lab that says AGI is twenty years off struggles to raise capital. A lab that says eighteen months struggles with regulators, enterprise buyers, and its own safety team. The result is language calibrated to survive both audiences, which is why executives reach for phrases like “a few thousand days” instead of a quarter and a year. The pattern of expert dates drifting isn’t unique to OpenAI, but OpenAI is the loudest about it.
How to read the next announcement without being taken in
Roadmap posts are marketing documents with technical footnotes. Read them like an investor reading an earnings call.
- Look for named evaluations, not adjectives. “PhD-level reasoning” means nothing without a benchmark, a version number, and a disclosed testing harness.
- Check whether a demo was one run or a median. Best-of-thirty is a different product from pass-at-one.
- Watch what stops being mentioned. Quietly retired milestones are more informative than new ones.
- Follow the money, not the blog. Compute contracts, power purchase agreements, and hiring velocity lag public claims by six to eighteen months and are much harder to spin.
There’s a fuller checklist in this guide to reading AGI roadmap updates without getting fooled, and it’s worth bookmarking before the next model launch.
OpenAI’s dates are partly a function of everyone else’s
Anthropic’s Dario Amodei has talked about powerful AI arriving around 2026 or 2027. Google DeepMind’s Shane Legg has given 2028 as a coin-flip year. xAI’s public claims are shorter still. Nobody in this race wants to be the lab that says it will take longer, because the narrative of inevitability is worth billions in valuation and talent recruitment. The real contest underneath the AGI race is for compute, power, and the small pool of researchers who can train frontier models, not for the calendar date.
The signals that would actually settle it
Forget the forecasts. Watch for these, because they’re hard to fake and easy to verify after the fact.
Sustained autonomous work. Not a demo agent booking a flight, but a system left running for a week on a real project with a real budget, producing output a paying client accepts. That’s the bridge from level two to level three, and nobody has crossed it publicly.
Revenue per employee at the labs themselves. If OpenAI’s own headcount flattens while revenue compounds, the models are doing economically valuable work inside the company. That’s the cheapest available proxy for the charter’s definition.
Macroeconomic traces. AGI as defined would show up in productivity statistics eventually, with a lag of several years. If the timeline claims of 2026 are right, we should see unusual movement in services-sector output per hour before the end of the decade. We don’t yet, and that absence is more informative than any interview.
A change in how OpenAI talks about safety. The charter ties AGI to a governance trigger: once the board determines AGI has arrived, Microsoft’s commercial rights change and the mission clause activates. That determination is a legal event, not a philosophical one. When it happens, you won’t need a blog post to tell you. There will be a filing.