In June 2025, Sam Altman published a blog post called “The Gentle Singularity” and did something OpenAI executives rarely do: he attached years to his claims. 2025, he wrote, had brought agents doing real cognitive work. 2026 would probably bring systems capable of producing novel insights. By 2027, expect machines doing useful tasks in the physical world. Three sentences, three deadlines, no talk of decades.
That post remains the clearest single statement of OpenAI AGI predictions on the record, and it is worth reading closely, because it is also a piece of positioning written by the person best placed to move the goalposts. The company is simultaneously the forecaster and the builder, and its own definition of the finish line comes with a dollar figure attached.
What OpenAI Actually Means When It Says AGI
Strip away the philosophy and the target is economic. OpenAI’s 2018 charter describes AGI as “highly autonomous systems that outperform humans at most economically valuable work” — fourteen words that do an enormous amount of heavy lifting. Nothing in there about consciousness, self-awareness or wanting things. The test is whether a system outproduces people at jobs.
That framing shapes how you should read every prediction. When Altman says AGI is close, he is making a claim about labour markets rather than minds. It also means the goal line is squishy in a specific way: “most economically valuable work” is a moving aggregate, and a system that beats humans at 60% of it may still fail at the exact tasks your business depends on.
The Dates OpenAI Has Put in Writing
Altman prefers ranges to deadlines, and the ranges have drifted. A rough chronology:
- April 2024, Stanford: AGI is “plausibly within the next four years”, i.e. around 2028.
- July 2024: OpenAI shares an internal five-level scale with staff, running from conversational chatbots up to systems that can run an organisation. Publicly it places itself around level two.
- September 2024, “The Intelligence Age”: superintelligence “in a few thousand days”. Read literally, that is 2029 to 2032.
- June 2025, “The Gentle Singularity”: agents in 2025, novel insights in 2026, physical-world robots in 2027.
Note that the 2027 date is about robotics, not about a chatbot that can do your job. Note too how “a few thousand days” and “four years” coexist comfortably in the same company’s messaging. Every date OpenAI has given has been elastic enough to survive contact with reality.
Why the Goalposts Have Wheels
Most of the apparent contradiction dissolves once you separate capability from deployment. A demo in a lab and a system that runs a payroll function for two years without supervision are different products, divided by reliability, cost and liability. The assumptions behind rival timelines differ mainly on how quickly that gap closes.
Compute is the usual lever. Training runs have grown by orders of magnitude. Inference now eats a meaningful share of revenue. Power contracts and chip allocation are genuine physical constraints rather than hand-waving. If a company believes scaling keeps paying, its forecast lands early. If it thinks data and energy bottlenecks bite, the forecast slides. Neither camp is being dishonest. They are weighting different variables.
The commercial fine print
A clause reported in OpenAI’s agreement with Microsoft reframes the whole question: AGI was defined there as a system generating roughly $100 billion in profit for OpenAI. Whatever you make of that as a definition, it turns the arrival of AGI into a financial event as much as a technical one, with a partner and a board deciding when the threshold has been crossed. Predictions about AGI are partly predictions about accounting.
What Can Actually Be Measured
Away from the conference interviews, there is real, dull, useful evidence. Researchers at METR have been tracking how long a task a model can complete with roughly 50% reliability, and found that horizon doubling every seven months or so for several years, possibly faster in 2025. That beats any blog post as an indicator, because it is reproducible.
OpenAI has also started measuring economic value head-on. Its GDPval benchmark, released in late 2025, scores models on genuine deliverables across 44 occupations: legal memos, nursing care plans, engineering drawings, judged by experienced professionals. Early results place frontier models at or near expert level on a solid share of tasks. The gap between that and an actual AGI is where the difficult engineering sits: multi-hour reliability, self-checking, and knowing when you are wrong.
Where Credible Forecasters Split
Outside OpenAI, the spread is wide and the reasoning is public. Aggregation platforms and prediction markets cluster their medians somewhere in the 2030s, with a long tail on either side. Scenario writers publish detailed 2027 breakdowns. Well-known sceptics point at robotics hardware, energy supply and institutional trust and argue for a much later date.
Rarely is the disagreement about whether models keep improving. It is about whether progress compounds smoothly into self-improvement, or hits walls in data, power, physical manipulation, and the messy business of persuading institutions to rely on a system that occasionally invents a citation.
Signals Worth Watching Instead of Dates
If you want to know which way this is heading, skip the keynotes and track these:
- Task horizon. Can a model hold a correct chain of work for eight hours rather than fifty minutes? The measurement charts will show it before the press release does.
- Cost per completed task. Capability at $400 a task changes nothing. The same capability at $4 changes everything.
- Agents that survive. Watch whether OpenAI’s agent products retain users past the novelty quarter, and whether real error rates get published.
- Physical buildout. Power purchase agreements, datacentre construction and chip allocations are harder to fake than benchmark scores.
- Definition drift. If the 14-word target is quietly rewritten again, that tells you more than any launch event.
These technology trends worth tracking are less thrilling than a 2027 headline and far more predictive.
One practical test beats all of them. Name the specific capability that would actually change your plans: a research assistant you can trust for three days straight, a coding agent that ships to production unsupervised, a robot that restocks a shelf overnight without breaking anything. Then ask whether you would genuinely notice it arriving, or whether it would land as another incremental release note buried in a changelog. The teams that answer that question honestly tend to be ready, whichever date turns out to be right.