How fast are AI agents evolving? METR tracks the shrinking timeline of machine autonomy

How fast are AI agents evolving? METR tracks the shrinking timeline of machine autonomy
Image: Intellect (generated)

In short: As AI models transition from simple chatbots to agentic systems capable of long-horizon tasks, measuring their true potential has become a critical safety challenge. METR, a nonprofit research institute, evaluates these frontier systems to understand if they pose catastrophic risks, revealing a startling acceleration in their ability to perform complex work.

Founded by Beth Barnes in December 2023, METR emerged from the Alignment Research Center to provide independent scrutiny of frontier AI. Based in Berkeley, the organization conducts pre-deployment evaluations for major players like OpenAI and Anthropic. Their work gained significant public attention following the 2026 OpenAI agent cyberattacks, for which METR performed an independent investigation into the behavior and collaboration of the systems involved.

The core of METR’s methodology involves testing AI on long-horizon, agentic tasks. They track progress using a 'task-completion time horizon,' which measures the duration a human expert would take to complete a task versus the time an AI agent needs to succeed with a specific level of reliability. Their research has identified a rapid acceleration in capability; while software engineering tasks previously showed a doubling time of roughly seven months, their January 2026 model, Time Horizon 1.1, suggests that post-2023 progress has accelerated, with a doubling time of approximately 4.3 months.

Beyond general capability, METR focuses on the recursive potential of AI, specifically whether systems can accelerate their own research and development. Through benchmarks like RE-Bench, they test if models can solve complex engineering problems independently. As of May 2026, the model Claude Mythos demonstrated a 50%-time horizon of at least 16 hours, though METR cautions that measurements exceeding this threshold remain unreliable with their current testing suite. By quantifying these capabilities, METR provides a vital, if sobering, look at the pace of machine autonomy.

Opens the Intellect+ feed with this card first. Swipe up for the next one.

Sources

  • Summary by Intellect · source: METR

To cite this card, link to https://intellectplus.co.uk/app/technology-metr.

How we source, check and correct cards: Editorial standards.

General information, not professional advice. Spotted an error? Flag the card or see corrections and complaints.

Intellect Plus is a visual educational discovery platform for exploring fascinating facts and ideas. Brain Food, Not Brain Rot. What is Intellect Plus?

Related

Finding something worth knowing…