Most benchmarks answer a narrow question: can a model pass this test? METR asks a more useful one. Give a frontier AI agent a real task, and ask how long that same task takes a skilled human. The longest task an agent can complete on its own, most of the time, is its time horizon.
The appeal is that the unit means something outside the lab. A model that can do a thirty-second job is a clever autocomplete. A model that can do a day's work unsupervised is a colleague.
What the curve shows
- 7 monthsdoubling time of the task horizon, 2019 to 2025 (METR)
- 4 monthsdoubling time since 2024 (METR; AI Digest)
- 89 daysdoubling time on METR's updated 2026 method
From 2019 to 2025 the horizon doubled about every seven months. Since 2024 it has doubled roughly every four. When ChatGPT launched in 2022 it could handle coding tasks that take a person about thirty seconds. Today's agents finish coding tasks that take a human more than fourteen hours, and AI Digest, which tracks the series, reports “no evidence of plateauing.”
A doubling every four months is sixty-four times in two years.
That arithmetic is the whole story. Linear intuition says the machines are getting somewhat better each year. The curve says the length of work they can take on is multiplying.
Why the result is sturdy
The obvious worry is that measuring “how long a human would take” is imprecise. METR addresses it head on: if its absolute measurements were off by a factor of ten, that would shift the arrival of any given capability by only about two years. A curve this steep is not sensitive to exactly where you draw it.
The measurement has also been tightened rather than loosened. METR's Time Horizon 1.1, published in January 2026, grew the task pool from 170 to 228 and more than doubled the number of tasks longer than eight hours, from 14 to 31. It is on that harder set that the doubling time came out at 89 days.
The strongest objection · Toby Ord, Oxford
“Some popular measures like METR time horizons are probably not the right tool for the job, since they can go to infinity without bringing other measures of intelligence along with them.”
He has a point: one curve is not the whole of intelligence. But it does not move alone. Training compute, benchmark scores, cost per task and the first self-improvement results are all moving in the same direction at the same time. Ord's own work names the decisive variable as how quickly the improvement loop turns, and a longer horizon is exactly what lets a machine take on more of that loop itself.
What to watch
The number to follow is not any single score but the doubling time itself. If it keeps shrinking, the case in our manifesto gets stronger every quarter. If it stalls, we will say so here, with the date. That is what an index is for.