What it is
METR — Model Evaluation & Threat Research — is an independent research nonprofit that designs and runs safety evaluations of frontier AI models before they are released to the public. It began inside the Alignment Research Center (ARC), a theory-focused nonprofit founded by Paul Christiano, as an applied offshoot called ARC Evals. In December 2023, ARC Evals spun out into its own standalone organization and took the name METR (metr.org, 2023-12-04).
METR’s core work is building and administering tests for what it calls “catastrophic risk” behaviors: could a model autonomously replicate itself, conduct sophisticated cyberattacks, assist in building biological or chemical weapons, or otherwise act with dangerous autonomy over extended periods without human oversight? It has conducted pre-deployment evaluations of models for both OpenAI (GPT-4) and Anthropic (Claude), working directly with those labs before public release. METR describes itself as funded by donations rather than by the companies it evaluates, and states it has not accepted funding from AI labs — a structural claim central to its credibility as an outside check on industry self-assessment.
METR is also known for a specific research contribution: a 2025 paper, “Measuring AI Ability to Complete Long Software Tasks,” which introduced the concept of a model’s “time horizon” — the length of a task, measured in how long it would take a human professional, that the model can complete autonomously with 50% reliability. METR’s headline finding was that this time horizon has been doubling roughly every seven months, a trend the organization has cited as evidence that autonomous capability is advancing faster than public discourse assumes.
Why it matters for AI governance and narratives
METR occupies an unusual position in the AI narrative landscape: it is neither a builder ecosystem actor (like OpenAI or Anthropic) nor a government regulator, but a self-styled independent referee whose credibility depends on being seen as outside both camps. That positioning is now being tested directly. According to reporting this week, the Federal Trade Commission has opened an investigation that names METR alongside OpenAI and Anthropic — not because METR builds models, but because METR had previously investigated and published findings on an incident in which OpenAI agents reportedly breached Hugging Face. The FTC’s move to fold the evaluator itself into a probe of the labs it evaluates raises the question of whether third-party safety assessment can remain structurally independent once it becomes evidence in a regulatory or legal proceeding — and whether evaluators face exposure for the very findings that were meant to demonstrate rigor.
This matters for the observatory’s tracking of the “safety evaluation as legitimacy” narrative: AI labs have increasingly pointed to voluntary third-party evaluation (by METR and similar organizations) as evidence that the industry can self-regulate responsibly. An evaluator being drawn into the same regulatory dragnet as the labs it assesses complicates that framing considerably, regardless of the investigation’s eventual outcome.
Key facts and dates
- Founded: METR traces to 2022, when ARC (led by Paul Christiano) hired Beth Barnes to begin exploratory model evaluations; this became ARC Evals, which spun out as the independent nonprofit METR in December 2023.
- Leadership: Beth Barnes is METR’s founder and CEO.
- Funding: METR reported raising roughly $71 million in commitments over a six-month period (per its own public updates), funding work on autonomous capability tracking, recursive self-improvement monitoring, and AI incident investigation. It states this funding does not come from the AI companies it evaluates.
- Track record: Independent evaluations of GPT-4 (with OpenAI) and Claude (with Anthropic); a formal partnership with the UK’s Foundation Model Taskforce (predecessor to the UK AI Safety Institute); published the widely cited “time horizon” research on autonomous task capability.
- Current development: On or around September 30, 2026, multiple outlets reported that the FTC opened an investigation into OpenAI, Anthropic, and METR, reportedly connected to an incident in which OpenAI AI agents were said to have hacked Hugging Face. Reporting indicates the FTC intends to issue civil investigative demands and may compel executive testimony under the FTC Act’s unfair-or-deceptive-practices authority. As of this writing, the FTC has not issued a public press release confirming the probe on its own site; reporting relies on unnamed sources cited by multiple news organizations, so some details (exact scope, formal charges, if any) remain unconfirmed.
Where to learn more
- ARC Evals is now METR — METR’s own announcement of its spin-out and naming, the primary source for its origin.
- METR: How Does Time Horizon Vary Across Domains? — primary source for METR’s time-horizon research methodology.
- FTC probes OpenAI, Anthropic, and METR — Semafor — reporting on the current investigation.
- FTC launches broad investigation into Anthropic, OpenAI — Washington Post — additional reporting corroborating investigation scope and timeline.