THE FRONTIER SIGNALSHow much work can an AI complete autonomously?
METR reports longer human-equivalent task horizons on its software and research suite. Opus 4.6 is about 12 hours at 50% success.
What was claimed?
Models can complete increasingly long tasks.
What does the evidence support?
Supported with limits
METR reports longer human-equivalent task horizons on its software and research suite. Opus 4.6 is about 12 hours at 50% success.
What are the limits?
TH1.1 snapshot updated May 8. A 50% threshold means frequent failure; above 16 hours the suite is insufficient for reliable measurement.
Reliability in other work domains and unsupervised deployment remain unresolved.
Claim or evidence date: 2026-05-08. Last verified: .
METR · Current TH1.1 ↗ · METR · TH1.1 dataset ↗
Who produced this claim?
METR conducts and publishes the task-horizon evaluation.
An external evaluator has a methodological perspective too. Independence from model production does not eliminate task-selection or measurement limits.
The series covers models from multiple developers under evaluator-defined conditions.
Public estimates and confidence intervals allow scrutiny. They do not constitute replication by a second evaluator.
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →