AGI Achieved
A software system matches or exceeds top human performance across nearly all cognitive tasks — not on a narrow benchmark, but broadly across reasoning, planning, creativity, and open-ended problem-solving. This is a purely cognitive milestone; physical presence in the world is not required and is tracked separately under Humanoid Robots at Human-Level Capability.
- Median year
- 2032
- P10 – P90 range
- 2027 – 2044
- Probability ever occurs
- 100%
- Last reviewed
- June 2026
Machine systems match or exceed top human performance across nearly every cognitive task. Every other event in this tree starts moving faster — the pace of research, automation, and geopolitical competition all accelerate simultaneously.
AI systems keep improving but remain narrow — highly capable in specific domains without matching general human reasoning across the board. The category of 'AGI' remains contested and unreached within this window.
Where things stand
As of mid-2026, the debate about AGI timelines is genuinely contested — not a manufactured controversy. Forecasting aggregators place the median at roughly 2031–2033. Lab leaders are considerably more bullish: Dario Amodei has argued that “powerful AI” — systems close to AGI by most definitions — could arrive within two to three years. Sam Altman has made similar statements. Andrej Karpathy and a significant portion of academic AI researchers remain skeptical of near-term AGI, expecting continued progress in narrow capabilities without a general breakthrough before 2040.
The definitional problem is real. “AGI” has no agreed standard, and any announcement will be contested. For this site’s purposes, the definition is deliberately strict: not an impressive benchmark result, but a system that demonstrably replaces or exceeds expert human performance across a broad and diverse set of cognitive tasks over a sustained period. That bar is harder to hit quietly — and harder to fake.
The primary disagreement among informed observers is not about direction but about pace. The argument for faster timelines rests on the scaling hypothesis: that current approaches, with sufficient compute and data, will continue to improve and eventually generalise. The argument for slower timelines rests on the possibility that current architectures hit a ceiling before reaching genuine generality — and that the remaining gap requires conceptual breakthroughs, not just scale.
The p_ever here is set at 1.0: AGI of some form is treated as an eventual certainty, given sufficient time. The uncertainty is entirely about when.
A note on physical embodiment. AGI as defined here is a software milestone — a system could qualify while running entirely in a data centre with no robotic body. The separate question of whether humanoid robots can perform arbitrary physical tasks at human level involves additional engineering challenges in dexterity, locomotion, and real-world perception that are not resolved by cognitive capability alone. That milestone is tracked separately.