FlashNEWS BRIEFING

METR Releases Time Horizons 2.0: AI Agent Task Durations Double Every 3.6 Months

1 min read

88Z AI, 21 November — METR has updated its AI agent capability assessment, revealing that the doubling time for task completion durations has shortened to 3.6 months. Recent tests show comparable performance between GPT-6 Astra and Opus 5.5, though extremely wide confidence intervals highlight significant uncertainty in current evaluations.

At a glance

  • AI agent task duration doubling time shortens to 3.6 months, aligning with OpenAI's forecast
  • GPT-6 Astra and Opus 5.5 show similar performance, but with extremely wide confidence intervals (11.9 hours to 110.2 days)
METR Time Horizons 2.0 data chart
METR Time Horizons 2.0 data chart

Updated Evaluation Metrics

METR has released the Time Horizons 2.0 report, updating its predictive model for the time spans AI agents can autonomously handle. The data indicates that the duration of tasks AI can manage is doubling every 3.6 months. This rate closely aligns with OpenAI’s previous estimate of a 3.8-month doubling period, suggesting a convergence among leading industry players regarding the pace of AI capability evolution.

Model Performance and Uncertainty

In specific model tests, GPT-6 Astra demonstrated a task handling duration of 40.3 hours, while Opus 5.5 recorded 39.0 hours, showing nearly identical performance. However, the report highlights that these figures come with extremely wide 95% confidence intervals, ranging from approximately 11.9 hours to 110.2 days. This substantial volatility suggests that while average estimates indicate rapid progress, there remains significant variability in the stability of current AI agents on complex, long-horizon tasks, warranting caution when relying on single mean values for deployment decisions.

Sources & editorial note

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.

  1. @scaling01Source published

Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.

Related reading

All stories
88Z.ai