METR’s new evaluation provides a meaningful but uncertain revision to recent AI time-horizon estimates, warranting attention to subsequent measurements rather than treating the faster trend as settled.
Limits of the evidence: The report says confidence intervals remain wide, only 5 of 31 long tasks have measured human baseline times, and TH1.1 re-estimates only 14 of the 33 models covered by TH1. The source also notes that trend estimates are sensitive to task composition.
What to watch
Taken from the first report; the whole story has not been re-assessed yet.
Time Horizon 1.1
METR BlogJan 29, 2026First report
No revisions yet.
No maturity ladder applies to this desk.
Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original sources are authoritative.