The year AI stops being a tool
you open — and becomes one
that closes the loop.
METR times the tasks frontier agents finish by themselves — at 50% reliability. Not what a demo can do once. What a model can do often, from a single instruction, measured in hours of human work.
“Seventeen-hour horizon” sounds like a workday. It is not. It means the agent finishes a tough job half the time — and the failures aren’t evenly spread. Chain a hundred steps at 99% reliability each and the whole run collapses to barely one in three.
In 2026, an agent that can hold a task for most of a working day stops being a chatbot with a big context window. It reads the repo, books the meeting, files the form, runs the tests, opens the pull request. The interface dissolves into a queue of outcomes.
The same answer that cost twenty dollars in 2022 costs a few cents today. Cost curves like this don’t produce a demo. They produce a default.
By 2027 the marginal cost of a competent answer is close enough to zero that “should we use AI?” stops being a budget question — and becomes a design decision.
Same models. Different institutions. The gap between them is measured in trust, not parameters.
The level isn’t a number.
It’s how much you trust it with
before you check.
A two-minute forecast of what artificial intelligence can actually do by the end of next year — measured in hours of human work, not in benchmark points.
Every number is sourced or labelled as an estimate.
This is a forecast, not a product announcement.
Download MP4 · 1080p 720p 30 fps · 02:00 · with score
What artificial intelligence can actually do by the end of next year — measured in hours of human work, not benchmark points.