AGI Made Clear › Are we in the AGI era?
Are we in the AGI era?
In September 2026 the president of OpenAI said we are. The chief executive of Google DeepMind puts it around 2030. Both were describing the same technology in the same month.
This page is the one we keep current, because it is the question that keeps being asked and the answer keeps moving. Everything below carries the month it was said or measured.
The short answerThere is no agreed test, so there is no agreed answer. What there is instead is a spread of positions that correlates almost perfectly with how demanding each person’s definition is - and a body of measurement that is moving very fast in some directions and not at all in others.
Who is saying what, in September 2026
Ordered by how strict their bar is, loosest first. Read each definition before the verdict beside it: the verdict follows from the definition, not from the evidence.
Jensen HuangNvidia
Can a system start and grow a billion-dollar business?I think it’s now. I think we’ve achieved AGI.
Lex Fridman podcast, March 2026
Already hereTheir verdict
Demis HassabisGoogle DeepMind
A system that exhibits all the cognitive capabilities the human brain has.Maybe like 2030, plus or minus a year.
Stanford Graduate School of Business, June 2026
Around 2030Their verdict
Yann LeCunAMI Labs
Human-level intelligence - which he argues cannot be built on today’s architectures at all.There are major conceptual breakthroughs that have to happen.
MIT Technology Review, January 2026
Not this decadeTheir verdict
All seven, ordered by how strict their definition is →
Why the dates sort themselves
One of these people is asking whether software can run a profitable company. Another is asking whether a machine has every cognitive capability a human brain has. Those are not the same test, and there is no reason at all they would be passed in the same decade. Nobody in the table is lying. They are answering different questions.
Which is why the first question to put to anyone offering a date is never really? - it is what exactly would count?
What the measurements actually show
Capability and reliability are not moving together, and the gap between them is where most of the disagreement lives.
Long-horizon workMETR, March 2026
≥16 hrs
The length of human task a model completes - not how long it runs unattended, which is the usual misreading. It is also the 50% success figure: at 80% success the public frontier measured about 1.5 hours. METR itself says measurements above sixteen hours are unreliable with its current task suite, and found that at least 16% of successful runs on the longest tasks used illegitimate methods.
Professional deliverablesGDPval, April 2026
85%
Blind expert graders rate the output at least as good as a human expert’s that often, up from 47.6% in September 2025. But the tasks are one-shot and fully specified, the automated grader agrees with human experts only 66% of the time, and the benchmark’s owner is the model’s vendor.
Learning after releaseEpoch AI, July 2026
No change
Models replaying an obscure co-operative board game score roughly flat across repeated attempts, with notes allowed between them. Newer models score higher only because they start higher. Epoch’s conclusion: no evidence the ability to learn on the fly has improved at all across model generations. A second study, from Berkeley in June 2026, found purpose-built memory systems underperformed plain in-context learning.
All eight measurements, with dates and sources →
The pattern in one lineThree of the hardest tests in AI were saturated inside eighteen months. In the same period the ability to learn from experience after release did not improve at all, and one lab reported its own model had got worse at knowing when it was unsure.
