METR measures what frontier AI agents can do in realistic, extended tasks rather than short benchmark questions. Its widely followed time-horizon evaluations estimate the length of software and research tasks models can complete, giving labs and policymakers a concrete view of rapidly changing autonomous capability.
See something inaccurate or outdated?
The AGI Safety researcher joined METR to investigate incidents of AI misalignment and assess whether current safeguards are sufficient.
September 12, 2026moveJoe Benton leaves Anthropic for METRThe former Scalable Oversight manager has joined METR to work on independent, embedded assessments of AI risks.
September 2, 2026moveThomas Kwa leaves METR for OpenAIThe lead author of METR's time-horizon research has joined OpenAI as a Member of Technical Staff to study recursive self-improvement.
September 1, 2026
