AI · Autonomous Agents

AI Autonomy Horizon (METR)

How long a human-expert task can an AI complete alone? Doubling every ~6 months for 7 years.

• 129-day doubling since 2023
• Mythos hit METR measurement ceiling
• Only METR-measured SOTA models plotted
• Suite needs rebuilding to go further

METR measures the autonomy horizon: the task length, in human-expert minutes, that a frontier AI agent can complete with 50% reliability. Every point here is a model METR actually measured and flagged as state-of-the-art in its own results file, plotted at that model’s release date. METR puts the doubling time at 188 days across the whole record and 129 days since 2023. GPT-4 managed about 4 minutes in March 2023, o3 reached 2 hours by April 2025, and Claude Opus 4.6 hit 12 hours in February 2026. In April 2026 Claude Mythos measured 17.4 hours, but with a confidence interval spanning 8.5 to 55 hours: METR excludes any model above 16 hours from its own doubling estimate because the current task suite cannot resolve it. The curve effectively pauses here until the suite is rebuilt for week- and month-long work. Points through 2022 are v1.0 measurements; everything from GPT-4 onward was measured under the v1.1 rebuild.

Source: METR Horizon benchmark results (v1.1 data file)

Download the data: ai_metr.csv · ai_metr.json

All curves · The chart is drawn in your browser and needs JavaScript. Every number it plots is in the table above.