How close are we to AGI? A capability check

What's already here

The measurable state of play is strong. Frontier models perform at roughly college-graduate level on knowledge work (~83% GDPval), handle complex agentic coding (~80% SWE-Bench Pro), and run as agents in real workflows. On the capability curve, the trajectory Aschenbrenner described is On track.

The missing piece

The gap between "very capable assistant" and "AGI" is autonomy. The defining bar — a system that can autonomously conduct AI research end-to-end, the "drop-in remote worker" — has not been demonstrated. That single gap is why the headline AGI-2027 prediction is graded Open, not On track.

Where the forecasters land

The bottom line: measurably closer than a few years ago — the whole forecast distribution has compressed from ~2060 to ~2033 in six years — but the finish line (autonomous research) is unproven. The nearest concrete test, Aschenbrenner's 2027, resolves by January 2028. As of July 2026, our auditable AGI-2027 Thesis Tracker puts the whole thesis at 62.5/100 — the single running measure of how close we actually are, and it moves only when a verdict changes. See the live score →

Frequently asked questions

Close on capability, not on autonomy. As of mid-2026 frontier AI reaches ~83% on GDPval and ~80% on SWE-Bench Pro, but no system autonomously conducts AI research end-to-end — the defining AGI bar.

Autonomy. The gap between a very capable assistant and AGI is a system that can independently do AI research and own tasks end-to-end without human cleanup — undemonstrated as of mid-2026.

No consensus: forecasts range from 2026 (Musk) and 2027 (Aschenbrenner) to 2030 (Hassabis), 2033 (Metaculus), and 2040 (academic survey median). Expert medians have compressed from ~2060 to ~2033 in six years.