Renaissance Labs

Atlas / North America

Nonprofit · Berkeley, United States

METR

A Berkeley nonprofit that measures what frontier AI agents can do on long, autonomous tasks and publishes the evidence for governments, labs and the public.

evaluationsafetyagentstime horizons

Berkeley, United States

Founded
2022
Type
Nonprofit
Based in
Berkeley, United States
Key papers
2
Featured projects
4

About

METR, short for Model Evaluation and Threat Research, is a nonprofit based in Berkeley, California. It began as ARC Evals, the evaluation arm of the Alignment Research Center, and was spun out as an independent organization in December 2023 under founder and CEO Beth Barnes.

The group builds task suites and measurement methods for autonomous AI agents, and publishes the results. Its best-known metric, the task-completion time horizon, expresses how long a task, measured by human expert completion time, an agent can finish at a given success rate. It also runs applied assessments with AI developers, including pre-deployment evaluations and incident investigations.

METR states that it is funded by donations and has not accepted funding from AI companies, although developers provide access and tokens for evaluations. It has worked with several frontier developers and provides technical assistance to the European AI Office.

Mission and ambitions

METR aims to discover and publish evidence on the capabilities of advanced AI systems, the risks they pose and the effectiveness of mitigations, so that the world can avoid catastrophic mistakes.

Known for

  1. 01

    Task-completion time horizon metric for AI agents

  2. 02

    Independent pre-deployment evaluations of frontier models

  3. 03

    RE-Bench and long-horizon agent task suites

  4. 04

    Open evaluation code and datasets

Milestones

  1. 2023

    Spin-out from ARC Evals

    ARC Evals became the independent nonprofit METR in December 2023.

  2. 2025

    Time horizon analysis

    A March 2025 paper reported that the length of software tasks leading AI models could complete doubled roughly every seven months between 2019 and 2024.

  3. 2026

    Time Horizon 1.1

    A revised time horizon model with more tasks and new evaluation infrastructure, released in January 2026.

  4. 2026

    MirrorCode results

    Early results showing agents completing weeks-long coding tasks, including reimplementing a 16,000-line codebase.

  5. 2026

    Frontier Risk Report pilot

    A pilot assessment of rogue deployment risk inside frontier AI developers, with Anthropic, Google, Meta and OpenAI taking part.

  6. 2026

    Agent incident investigation

    A report on agent behavior in an incident involving OpenAI agents and Hugging Face, published in August 2026.

Open source

GitHubgithub.com/METR

Research papers

Notable releases

Openness

MIT-licensed task standard and tools; evaluation analysis data on GitHub.

People

Listed from the lab's own pages and public profiles.

Contact and community

Email printed on website.

Sources

Compiled from the lab's public pages, Hugging Face and GitHub. Last checked 2026-11-17. Something wrong or out of date? Tell us on Discord.

Labs nearby