METR Secures $71 Million to Scale Independent AI Tests
The nonprofit evaluator will expand research on autonomous capabilities and AI-driven self-improvement while rejecting lab funding.
A major expansion for independent evaluation
Model Evaluation and Threat Research, or METR, said it secured commitments of about $71 million over the past six months, giving the nonprofit substantially more capacity to examine rapidly advancing AI systems. Announced in a series of posts on August 14, the funding will support a larger team and new projects focused on autonomous capabilities, recursive self-improvement and other risks associated with frontier models.
METR develops evaluations that measure whether AI agents can complete increasingly long and complex tasks without human intervention. It has also participated in predeployment assessments involving systems from OpenAI, Anthropic, Google DeepMind, Meta and Amazon. Its work is particularly influential because model developers, policymakers and safety researchers use such evaluations to judge whether new capabilities warrant stronger safeguards.
The organization emphasized that it has not accepted money from frontier AI companies. Labs including OpenAI, Anthropic and xAI have provided model access and inference credits for evaluation work, but METR says those arrangements are separate from financial support. The distinction matters as external evaluators increasingly depend on privileged access to systems built by the same companies they may later criticize.
METR credited institutional and individual supporters, including the Audacious Project, which provided its first institution-scale backing. The announced figure represents funding commitments rather than a conventional venture round, and METR did not publish a detailed allocation schedule.
Why it matters
Independent evaluation has become a bottleneck as AI agents move from short demonstrations toward sustained software engineering, research and computer-use tasks. A $71 million capital base is unusually large for a specialized nonprofit and could let METR build evaluation infrastructure at a scale closer to that of frontier laboratories. Its refusal of lab funding also establishes a consequential governance test: whether an evaluator can retain technical access, recruit scarce researchers and publish uncomfortable findings without becoming financially dependent on the companies under examination.