Anthropic opens its doors to METR. What can the evaluators publish?

A seat inside, with boundaries

Anthropic’s update describes company hardware, office space and internal tools for the small METR team. Customer information and privileged legal material remain restricted. Model architecture details need special approval when relevant. Other METR staff are separated from this work.

METR can publish risk assessments without Anthropic’s veto. The company retains narrowly defined redaction rights; METR can publicly dispute those redactions without revealing the protected material. Anthropic set out the arrangement in its original announcement:

Look beyond the model’s answers

METR’s proposal, published a day earlier, explains why seeing a model through a chatbot is insufficient. It identifies three questions: might an agent try to evade human control; would monitoring catch it; and could AI-assisted research accelerate development faster than safeguards can keep up?

The group says it is already conducting an alignment assessment with Anthropic: an investigation into whether agents behave as their developers intend. Its proposal also calls for scrutiny of incident records and the methods used to find problems. This is a programme of evidence gathering, not a published verdict on the new arrangement.

Independence has to survive the working relationship

The announcement drew scepticism as well as interest. Technology writer and EZPR chief executive Ed Zitron used a banking analogy in a critical quote-post. His comparison is an opinion about the arrangement, not evidence that an assessment has been suppressed.

METR’s own account of working with AI companies is more revealing than a simple independent-or-not label. It commits to describing access limits, redactions, staff conflicts and informal pressures that could affect its conclusions. It also acknowledges that voluntary cooperation leaves companies with influence over what evidence outsiders receive.

Financial distance is another part of that picture. In its August funding update, METR says it does not take funding from frontier AI companies or donations directed by their staff. It accepts free model access, while saying it could afford to replace those tokens. That is METR’s disclosed policy, rather than an external audit of this engagement.

The next test is the report

Elon Musk replied that SpaceX would probably follow suit. That signals interest from another AI developer; it does not establish that a comparable team or agreement is already in place.

For organisations buying AI systems, the useful next document is the assessment itself: which model and period it covers, what the team could inspect, what was withheld and whether those limits change its conclusions. A badge saying “externally evaluated” cannot answer those questions.

METR explicitly says it does not intend to decide how companies develop or deploy their agents. Its role is to provide evidence for others to judge. That distinction matters to European buyers too: this voluntary access arrangement supplies no blanket approval for a particular deployment.