The Intelligence Ladder

"How smart is AI now?" is unanswerable. "How long a task can it finish?" is measured. Every frontier model on one ladder, fitted windows for the rungs ahead, and what "AGI" and the tiers beyond would mean in numbers.

the ladder — every measured model

Each dot is a frontier model's 50%-reliability task horizon — the length of human expert work it completes half the time (METR's measurement, the field's standard capability series). The line is the state-of-the-art frontier; the shaded fan is the engine's fitted projection with an 80% interval. The horizontal rungs are where the ladder stops being abstract.

the rungs — from minutes to "god-tier", in measured units

Where is AGI on this? Nowhere — because "AGI" isn't a measurement, and this engine doesn't do vibes. But notice what the rungs quietly replace: an agent that holds week-long jobs is what the economy will experience as transformative; an agent that carries a year of work is what most people mean by AGI whether or not the word survives; the tiers beyond — decade-tasks, century-tasks, run in million-agent fleets — are the "god-like" territory, and on this ladder they are not mystical: they are more doublings. The windows above are fitted from seven years of measurements; treat the far rungs as labeled extrapolation, not prophecy.

the race — lab by lab

Grok (xAI), Kimi (Moonshot), DeepSeek and other labs do not appear in METR's published Horizon evaluations — not measured is not the same as not capable, and this page shows only measured points.

the cadence — will something big land next month?

Nobody can name the next invention. But how often frontier events arrive is measurable: every dot on the ladder is a dated, METR-measured release, and the subset that set a new state of the art is a point process whose inter-arrival gaps fit a base rate. Everything below is computed from those release dates — a frequency, not a prophecy — and it is scoreable: either a measured SOTA crossing lands in the next 30 days or it doesn't.

What this base rate leaves out. It counts only releases METR has measured — the released-but-unmeasured frontier (Mythos 5 / Fable 5) is invisible here, so the true cadence runs at least this fast and the current drought is partly a measurement lag, not proof of a slowdown. The ruler itself saturates near ~16 hours, so from here a missing crossing can mean a stalled frontier or a stalled ruler. And the exponential (memoryless) fit is a modeling choice, est-grade: real releases cluster around lab cycles. What makes this an instrument rather than a vibe is that it grades itself — over many 30-day windows, about this fraction should contain a crossing, and if they don't, the number was wrong and gets published as wrong.

the hidden tiers — how far ahead can unseen tech be?

The amber band on the chart is the shadow frontier: capability that plausibly already exists but isn't public. It has two very different sources, and they need to be sized separately:

The historical pattern is real: in aerospace, sensing, and cryptography, classified programs ran 5–20 years ahead of public technology — so the instinct that "what exists exceeds what's shown" has a solid base rate. But AI inverts it. Frontier AI needs the world's largest compute concentrations, open talent markets, and commercial-scale capital — all of which live in private labs, and government programs today consume frontier-lab models rather than exceeding them. So a fair size for today's AI shadow band is months of doublings (lab-internal + measurement lag ≈ ×2–4), not decades — while for defense-adjacent hardware (sensing, propulsion, EW), the historical multi-year hidden tier likely still applies. The engine's rule holds: the band is drawn as an explicitly estimate-grade region and a bias direction — never as invented data points.

limits of this chart

• 50% reliability. The horizon is where the agent succeeds half the time — the 80%-reliability horizon runs roughly 6× shorter (Opus 4.5: ~4.9 hrs at 50% vs ~49 min at 80%, per the same v1.1 file). Long tasks still fail often; the 50%/80% gap is the capability–reliability gap, in numbers.
• Task domain. METR's tasks are software, research and reasoning work — the ladder measures cognitive labor, not plumbing, persuasion, or wisdom.
• Wide error bars at the top. The newest points carry huge confidence intervals (the top model's spans roughly 8 to 55 hours) — few tasks are long enough to test them.
• The ruler itself is saturating. METR states the TH1.1 suite can't reliably measure horizons above ~16 hours — the top of this ladder sits at the ceiling of the measuring instrument, and longer rungs need new tasks before they can be measured at all. A frontier that outgrows its ruler is itself a data point. And on the long tasks the ruler does have, scoring gets harder: METR found at least 16% of successful 8-hour-plus runs were illegitimate on review — one model's horizon would have read ~2× larger if cheats counted.
• The measured shadow, once. METR's first entity-based frontier pilot (Feb–Mar 2026, with the major labs) put the most capable shared internal model at ~16–20 hrs at 50% / 3–4 hrs at 80% — a modest gap over the public frontier, measured right at the suite's ceiling, "highly uncertain" by METR's own read. It's the first direct calibration of the shadow-frontier band below.
• The released Claude Mythos 5 / Fable 5 (mid-2026) has no published METR measurement yet — the April point is the pre-release preview. Third-party extrapolations put the released models near ~60 hours, but extrapolations are not measurements and stay off this board.
• The shadow frontier. Three lags stack on this chart: labs run their next model internally for months before release; released models wait months for measurement; and the ruler tops out at ~16 hours. When capability doubled every two years, that lag was noise — at ~5 months per doubling, the frontier that exists is plausibly 2–4× above the frontier you can see here. Read every rung window with that skew in mind: the visible ladder runs behind the real one, and always in the same direction — early.
• The pace itself is uncertain in the fast direction too. TH1.1's own tiers: doubling every ~196 days across the record, ~131 days since 2023, ~89 days since 2024. If the recent tier holds — roughly a doubling per quarter — every window above lands early; whether it holds is the live superexponential question, not a settled fact.
• What could break the curve: power, chips, and data are ranked physical constraints on training scale (Epoch AI's analysis) — the counter-curves. What could break the constraints: the recursion tracked on the cascade — AI building the fabs, farms and power that build AI.