METR

METR is a US non-profit research organisation that evaluates the risks and capabilities of advanced AI models and informs the public about them. Its name is short for 'Model Evaluation & Threat Research'. It began as ARC Evals, an evaluation team set up in 2022 when the Alignment Research Center (ARC), an AI safety research group, hired Beth Barnes. The team carried out independent evaluations of OpenAI's GPT-4 and Anthropic's Claude in partnership with each company. In December 2023 it spun out of ARC and became a standalone non-profit (US 501(c)(3)) named METR. As of October 2026 Barnes is its CEO, and its main workplace is Berkeley, California.

METR evaluates how long the tasks are that AI can complete without human help, and whether AI has worrying abilities such as carrying out cyberattacks or making itself hard to shut down. In its March 2025 'time horizon' research, it reported that the length of tasks (measured by how long they take human professionals) that AI agents (AI that carries out tasks on people's behalf) can finish with 50% reliability had doubled about every 7 months over 6 years. In a July 2025 experiment, experienced open-source developers actually took 19% longer when using AI tools, yet they themselves felt that AI had sped them up by 20%. METR says it has never accepted funding from AI companies and is funded mainly by donations.

In January 2026 it released Time Horizon 1.1, which increased its set of test tasks from 170 to 228; looking only at the period since 2024, the doubling time came to about 89 days. In May 2026 it published a report on the risk that AI used inside companies could slip out of control, based on non-public information and internal models from Anthropic, Google, Meta and OpenAI, and said this was the first independent assessment of its kind. On 15 October Anthropic announced that it had given six METR employees access to internal tools and systems roughly on par with that of its own risk assessment teams (company-reported). Anthropic said it cannot stop METR from publishing its findings and can only require predefined kinds of confidential information to be redacted.