Tracking AGI progress since 2026

How close
are we to
human-level
intelligence?

The AGI Guide tracks the world's leading experts, benchmark scores, and milestone events — so you know exactly where we stand. Not hype. Not speculation. Data.

Median expert estimate
2029
50% probability — Metaculus community
Anthropic's formal submission
2026–2027
To US Office of Science and Technology Policy
Most aggressive lab CEO
2026
Elon Musk, xAI — end of 2026

The window is narrow.

AGI will arrive as a threshold, not a moment. Labs are already on internal timelines of 2027–2028. When the first system crosses the threshold — whether it's late 2026 or 2032 — the world changes overnight.

The AGI Guide exists so you can watch the approach in real time, understand the debate, and track the markers that actually matter.

Current probability window
2026
2033
10% by 2027
50% by 2029
90% by 2041

The experts are not aligned.

That's the point. When Sam Altman and Demis Hassabis give different answers, the honest answer is: nobody knows. The AGI Guide tracks all of them.

Sam Altman
CEO, OpenAI
2026–2028
"We know what to do to build AGI."
Elon Musk
CEO, xAI
2026
End of 2026 — smarter than any human.
Dario Amodei
CEO, Anthropic
2026–2027
Formal White House filing, 2025.
Demis Hassabis
CEO, Google DeepMind
2030–2035
Cites jagged intelligence gaps in current systems.
Geoffrey Hinton
Turing Award Winner
2028–2053
90% confidence interval, revised from decades.
Daniel Kokotajlo
AI Futures Project
2029 (median)
Originally 2027 — updated after slower 2025 progress.
Ranges span from 2026 to 2053 — the AGI Guide tracks all of them so you don't have to pick a side.

What we're measuring

Benchmarks aren't perfect. But when multiple independent tests move simultaneously, that's a signal — not hype.

ARC-AGI
Measures ability to learn novel tasks, not just pattern match. Created by François Chollet.
o3: 87.5% — humans avg ~60%
FrontierMath
Research-grade mathematical proofs at the Olympiad level. Models scored near zero until 2024.
Rapid improvement — current models at ~45%
METR Task-Length
How long a task a human expert can do — and whether AI can match it. Updated monthly.
Claude Opus 4.5: 4h49m at 50% — human expert baseline 5h
Human-Last Exam (HLE)
Humanities equivalent of GPQA — tests expert-level multidisciplinary reasoning.
Early stage — major gap remaining

How forecasts have moved

Forecasters have updated their timelines 8+ times in the last 24 months. The direction is the story.

2023
~2040 median
Post-ChatGPT surge — most forecasters pulled timelines forward
2024
2032–2035 median
xAI Grok, Meta Llama, Gemini releases underwhelm — some push back
2025
AI 2027 scenario published, Kokotajlo median → 2029
Wider AGI debate enters mainstream — US government, VP Vance reference it
2026
Every updater pulls earlier — Anthropic advances the frontier
Updated in progress — first half of 2026 shows consistent compression

The most consequential event in human history is not a prediction. It's a planning problem.

Whether AGI arrives in 2026 or 2035, the organizations and individuals who understand the landscape will be better prepared. The AGI Guide exists to give you that understanding — without the noise.