When is AGI coming? Altman, Hassabis and Amodei disagree
Altman, Amodei and Hassabis give AGI dates a decade apart. Sorting their forecasts by which question they answer, plus the one measured trend line that isn't an opinion.
Priya Raman is a staff writer at Stanford Tech Review covering AI, semiconductors, and emerging technologies across Silicon Valley.
Ask five people who build frontier AI systems when artificial general intelligence arrives and you will get five dates spread across fifteen years. That is not because four of them are wrong. It is because "AGI" is doing the work of at least three different questions, and each question has a different answer.
The first question is when a model will beat humans on the benchmarks we currently use to measure intelligence. On most of those, it already has. The second is when AI can do economically valuable work without a human watching. That one is measurable, it is moving fast, and it has a number attached. The third is when a machine matches the full range of human cognition, including the parts nobody knows how to test. That one is a research question dressed up as a date.
Sorting the predictions by which question they answer makes the disagreement much smaller than it looks.
The definitions do not agree, and that is most of the disagreement
OpenAI's own charter defines AGI as highly autonomous systems that outperform humans at most economically valuable work, which is a labour-market test rather than a cognitive one. A system could satisfy it while being unable to learn anything new after deployment. DeepMind's published framework goes the other way and grades generality in levels, separating narrow competence from broad competence and expert performance from superhuman performance, so that a model can be superhuman at one thing and below a novice at another without the label collapsing.
Benchmark designers have pulled in a third direction, building tests specifically for the skills that scaling does not obviously deliver: abstraction from a handful of examples, novel puzzle solving, holding a goal across many steps. Those tests exist precisely because saturating the older exams stopped telling anyone much. When a lab announces it has cleared one, that is evidence about a capability, not about a threshold.
The practical consequence is that two people can look at the same model and disagree about whether it counts, without either of them being confused about the facts. Before comparing any two AGI predictions, it is worth checking which of the three definitions each one is using. Most of the apparent gap between a 2027 forecast and a 2033 forecast closes once you do.
It also explains why the term is quietly falling out of use inside the labs that popularised it. Altman now talks about superintelligence, DeepMind talks about levels, and Anthropic talks about powerful AI. The phrase AGI survives mainly in headlines and in funding announcements, where its vagueness is a feature.
What the lab heads actually say
Dario Amodei, Anthropic's chief executive, has been the most aggressive of the major lab leaders, describing human-level systems as plausible in the 2026 to 2027 window and arguing that progress compounds on itself once models start meaningfully accelerating their own development.
Sam Altman has stopped defending a specific AGI date and moved the conversation past it. His public position is that systems capable of doing most remote knowledge work arrive within this presidential term, which puts the useful part of his forecast somewhere in 2026 to 2028, and that the more interesting threshold after that is superintelligence rather than parity with a human.
Demis Hassabis, who runs Google DeepMind, has held the same estimate for years and has not moved it much: roughly even odds by 2030. His stated reason for the gap is specific rather than vague. Today's models do not learn continuously, they do not hold a coherent world model across long stretches of work, and they fail in ways a competent human would not, which he treats as missing capabilities rather than missing scale.
Notice that Amodei and Hassabis are not really contradicting each other. Amodei is answering the economic question. Hassabis is answering the cognitive one. A system that automates a large share of remote work in 2027 and still cannot learn on the job is consistent with both statements.
What the forecasters say
Outside the labs, the people who bet on this for a living are considerably more conservative, and they moved the wrong way this year. As of mid-2026 the Metaculus community median for the arrival of a general AI system sits in the early 2030s, with roughly a one-in-four chance before 2029. Weaker definitions, the ones that require passing a battery of tests rather than demonstrating general competence, land in 2026 or 2027.
The direction of travel matters more than the exact figure. Between 2023 and 2025, forecasters repeatedly pulled their AGI dates earlier as each model generation surprised them. Through 2026 they have pushed the strong-AGI date back out. Capability on benchmarks kept climbing; confidence that benchmark capability converts into general competence did not.
The one number that is a measurement rather than an opinion
There is a single line of evidence in this debate that does not depend on anyone's definition of intelligence, and it comes from METR, a nonprofit that evaluates frontier models. Instead of asking whether a model is smart, METR asks how long a task can be before the model stops finishing it. Every task in its suite is timed against a human expert, so the unit is human minutes rather than a score.
The headline metric is the 50 percent time horizon: the task length at which an agent succeeds half the time. On METR's published data, the progression through the last two years is stark. Claude 4 Opus in May 2025 sat at about 100 minutes. GPT-5 in August 2025 reached roughly 3.4 hours. Claude Opus 4.5 in November 2025 hit 4.9 hours. By February 2026, Claude Opus 4.6 measured just under 12 hours, and an April 2026 preview system came in above 17 hours, which is past the point where METR itself warns the task suite stops being reliable.
Fitted across every frontier model released since 2023, that horizon has doubled every 128.7 days, with a confidence interval running from 104 to 158 days. That is the number worth carrying out of this article, because it is the only one here derived from measurement rather than intuition. It is also faster than the roughly seven-month doubling METR found when it included the 2019 to 2022 era, which is why several researchers now describe the trend as accelerating rather than steady.
Extrapolate the fitted line and you get a full eight-hour workday of autonomous software work in 2027, a work week in 2028, and a month in 2029. On the faster recent rate, those dates compress by a year or more.
Four reasons the extrapolation may break
The straight line on a log chart is seductive, and it has held through six model generations. It is still worth stating the failure modes plainly.
Reliability lags capability badly. The 80 percent time horizon, the length at which a model succeeds four times out of five, runs far behind the 50 percent figure across every model METR has tested. A system that finishes a twelve-hour task half the time is not a system you leave alone for twelve hours. Most real deployment thresholds sit closer to the 80 percent line, and that line is where the economic question actually gets decided.
The tasks are software tasks. METR's suite is built from coding and technical reasoning work with clean success criteria. That is the domain where models are strongest and where verification is cheapest. Nothing in the data licenses extrapolating to negotiation, physical work, or any job where "done correctly" is a judgment call.
Compute growth is not guaranteed to hold. The trend was produced during a period of extraordinary capital expenditure on training and inference. Power availability, chip supply, and the willingness of capital markets to keep funding it are all real constraints, and a compute slowdown pushes the curve right.
Continual learning is still missing. The gap Hassabis names has not closed. Frontier models do not accumulate experience across sessions the way an employee accumulates it across a year. Long-horizon autonomy without persistent learning may plateau at the point where the missing memory starts to bind.
What to watch instead of a date
Anyone trying to plan around this is better served by triggers than by forecasts. Four are concrete enough to check.
Watch the 80 percent horizon rather than the 50 percent one, and treat the day it clears a full workday as the moment unsupervised deployment becomes defensible. Watch whether the doubling trend holds on tasks outside software, because that is the test of whether this is general progress or a domain-specific run. Watch for a lab publishing evidence that its models are meaningfully compressing its own research cycles, which is the mechanism every fast timeline depends on. And watch whether continual learning ships as a product feature rather than a paper.
The honest answer to "when is AGI coming" is that the question has no date because it has no agreed definition, and the version of it that does have a date is narrower than the phrase suggests: on current measured trend, AI systems will handle a full working week of software tasks unsupervised around 2028, and that will arrive long before anything worth calling general intelligence does.
The gap between those two events is where the next decade of arguments will happen.