METR (Model Evaluation and Threat Research)
Independent pre-deployment evaluation of frontier models for autonomous and dangerous capabilities.
METR is the outside tester the big labs let poke at their models before release. It measures how long and how autonomously AI agents can work, and publishes what it finds.
Focus
- Safety evaluations
- Transparency
- Alignment research
Funding, as far as we know
Philanthropic grants (Open Philanthropy/Coefficient Giving, Audacious Project, others); METR states it does not accept payment from labs for evaluations - https://metr.org/about/
What they've actually done
Published the 'task-completion time horizon' metric showing frontier agents' capability doubling roughly every 7 months, now widely cited
19 March 2025
Ran a pilot 'Frontier Risk Report' with internal model access from Anthropic, Google, Meta and OpenAI, documenting 100+ constraint violations by internal agents
19 May 2026
Published its pre-deployment evaluation summary of OpenAI's GPT-5.6 'Sol'
26 June 2026