MEASURED AI · REALITY, WITH RECEIPTS
Is AI becoming
more useful?
Capability is one part of the answer. Follow measured results through time, then check the cost and the evidence from real work.
Can it do the task?
At what cost?
Does it work in practice?
AI horizons: months & years ↗ Explore measured history and conditional longer-term scenarios.
A DIFFERENT CAPABILITY LENS · METR TH 1.1
AI Task Horizons
From measured tasks to possible month- and year-scale workloads. Explore the evidence, adjust future assumptions and see what sustained autonomy would still need to demonstrate.
Human task time, not unattended AI time
A two-hour horizon at 50% means a fitted model predicts success half the time on tasks that take a human expert about two hours. It does not mean two hours of continuous autonomous operation. These tests mainly cover software, machine learning and cybersecurity.
Loading historical task-horizon measurements…
What this can — and cannot — tell us
Editorial interpretation: handling longer, well-defined tasks could reduce how often people need to break work into small steps. Real benefit still depends on error rates, supervision, cost and whether the tasks resemble actual work. This chart measures none of those benefits directly.
Human baselines include timed attempts and estimates. Different scaffolds, task suites and elicitation choices affect the results. Wide intervals are not a precise ranking. Only entries explicitly labelled TH 1.1 are plotted. Three TH 1.0 entries bundled in the same source download are retained but excluded. No cross-version growth rate is inferred.
Newer or absent models remain unmeasured here. The snapshot is separate from EBIGRAPH’s current benchmark edition and does not alter its cohort or methodology.
Loading released measurements…