What METR does
METR (pronounced “meter”) is an independent AI safety research nonprofit that develops and applies scientific evaluation methods for frontier AI systems, with a focus on (1) autonomous, long-horizon capabilities that can enable substantial real-world task completion and (2) threat research on how such systems (or actors around them) can pose risks to society. METR’s work emphasizes empirical, measurement-driven assessments, publishing both evaluation research and summaries of findings meant to inform public understanding and downstream decision-making about AI development and governance.
METR’s research program includes autonomous evaluation benchmarks and protocols (including “time horizon” style measures of task persistence), evaluation of model behavior on agentic software/research tasks, and related work on evaluation integrity threats and mitigations (e.g., behaviors that threaten the validity of assessments). In addition to model capability evaluations, METR has conducted studies on AI’s effects on real-world software developer productivity and on how AI tools may change developer workflows over time.
Institutionally, METR positions itself as a third-party evaluator and measurement lab, describing how it uses independent evaluation infrastructure and publishes results where possible. METR states it is funded by donations (and also notes a limited technical assistance contract with the European AI Office).