Time horizon is a headline metric for autonomous AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → capability. The expanded task suite and revised estimates change the numbers practitioners cite for model trends.
Key takeaways · AI-distilled
Under TH1.1, METR's post-2023 frontier time-horizon doubling time is 131 days, versus 165 days under TH1, so progress reads about 20% faster; measured since 2024 it falls from 109 to 89 days.
The trend shift comes from older models moving down and newer ones up: two GPT-4 versions fall 35% and 57%, while GPT-5 rises 55% and Claude Opus 4.5 rises 11% to a 320-minute estimate.
More long tasks tightened the top end: Opus 4.5's upper bound was 4.4x its point estimate under TH1 and is 2.3x under TH1.1. METR says intervals remain very wide, and only 5 of its 31 tasks over eight hours have measured human baselines.
Moving from METR's in-house Vivaria agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → to UK AISI's open-source Inspect changed estimates little overall, but GPT-4o and o3 scored significantly higher under Vivaria, which METR reads as sensitivity to scaffold and prompting.
METR re-estimated only 14 of 33 models, skipping ones no longer available, needing major scaffold changes, or far from the frontier, and says even TH1.1 has few tasks the newest models fail, so harder evaluations are its priority.
Terms in this piece · Glossary
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
