In January 2025, something shifted in the public tone of the leaders of the world's largest AI labs. In a matter of weeks, Dario Amodei (Anthropic), Sam Altman (OpenAI), and Demis Hassabis (Google DeepMind) all revised their AGI estimates upward — from "decades" to "years." By 2026, these predictions became more specific, more public, and more consequential.

What is happening, what exactly each of them is saying, and why the divergences among them matter as much as the convergences.

Dario Amodei: 2026–2027 as the Official Horizon

Anthropic's position is the most specific and officially documented. In the company's formal submission to the US Office of Science and Technology Policy in March 2025, Anthropic stated it expects powerful AI systems to emerge in late 2026 or early 2027. This is not a podcast quote — it is a regulatory document.

Amodei went further in public statements: "I am more confident than ever that we are close to powerful capabilities… in the next 2–3 years." And in February 2025: "What I've seen inside Anthropic and outside over the last few months has led me to believe we are on a path to human-level AI systems that outperform humans across all tasks within 2–3 years." The projection refers to systems "broadly better than all humans at nearly everything" — not narrow AGI, but generalist AGI.

Sam Altman: AGI During the Current Presidential Term

Sam Altman adopted a vaguer but equally ambitious formulation: AGI "will probably be developed" during the current US presidential term (which ends in January 2029). In its public roadmap, OpenAI projects AI "research interns" by 2026 — systems that function as autonomous research interns — and fully autonomous AI researchers by 2028.

Altman's expectation for 2026 is specific: tasks that currently take hours will start taking days. "As these systems move from multi-hour tasks to multi-day tasks, which I expect to happen next year..." — a statement from January 2025 that partially materialized: in February 2026, Claude Opus 4.6 crossed the 14.5-hour mark of sustained autonomous work, doubling every 123 days.

Demis Hassabis: More Cautious, But Also Revised

Hassabis occupies the middle ground. His public estimate migrated from "5 to 10 years" in 2024 to "3 to 5 years" in January 2025. He emphasizes what he calls "jagged intelligence" — the observation that current systems display gold-medal performance in Olympic math alongside failures that a twelve-year-old would not commit.

Shane Legg, DeepMind co-founder and Chief AGI Scientist, was more specific in January 2026: he assigned a 50% probability to "minimum AGI" by 2028, defined as an artificial agent that can reliably perform the full range of cognitive tasks an average human can do, without failing in ways that would be surprising if a person had the same task.

The Dissenting Voices: LeCun and Marcus

Not everyone agrees. Yann LeCun argues that current architectures — including all LLMs — cannot achieve AGI in principle, regardless of scale. Gary Marcus holds a similar position: current models have structural limitations that scaling does not solve.

Geoffrey Hinton, who received the 2024 Nobel Prize in Physics, revised his estimates for the short term but maintains profound ambivalence: he estimates a 10–20% chance of AI causing human extinction, and publicly stated that part of him regrets his decades of work in deep learning.

Why Divergences Matter

What is at stake in AGI predictions is not just a calendar date. It is the allocation of hundreds of billions of dollars in infrastructure, the urgency of regulation, and the question of who controls the systems when they arrive.

The most honest rule for interpreting AGI predictions in 2026: any AI lab CEO who gives you an exact date is a salesman. The evidence does not support confident predictions in either direction. It supports monitoring leading indicators, pricing AGI readiness into long-term decisions, and treating any single timeline as the most optimistic outcome of a broad distribution.