Anthropic METR Access: Inside the Three Evaluator Tiers
Anthropic put six METR staff inside the company with employee-like access. The three-tier structure behind it is the real story, and it has limits.
Anthropic has begun giving METR, a nonprofit that evaluates frontier AI models, employee-like access to its internal systems. The 15 October 2026 announcement says six METR employees have been onboarded with company hardware, office space, and access to internal tools roughly on par with Anthropic's own risk assessment teams. The Anthropic METR access deal matters less for the headcount than for the structure around it: three tiers that decide who controls access and who controls publication.
What changed since the September announcement
In September, Anthropic announced evaluator partnerships that we covered in Anthropic's embedded evaluators, including the commercial structure and the independence objection to paid evaluators. This update adds the first employee-like arrangement and spells out the access model. According to Anthropic, last month it committed to stronger independent assessment of its models and safety practices, and signed the White House Accord on Super Intelligence, which includes pledges to work with independent external auditors or evaluators.
Anthropic METR access and the three evaluator tiers
Anthropic describes three tiers. The differences are about control.
Tier | Access | Who approves publication |
|---|---|---|
Tier 1 | Narrowly scoped investigations, with access decided by Anthropic | Anthropic must approve |
Tier 2 | Broader information access | The evaluator publishes without Anthropic's approval, though specific categories can still be redacted |
Tier 3 | Employee-like access, the METR arrangement | Publication rights similar to Tier 2 |
Read the right-hand column first. A tier that needs the company's approval to publish is a different kind of arrangement from one that does not.
What Anthropic can and cannot remove
For METR, Anthropic says it cannot veto publication. METR will publish its analyses of risk levels, incidents or practices. Anthropic can require redactions in three narrow, predefined categories: customer data, attorney-client-privileged information, and model architectures. METR may say publicly whether it thinks the redactions were justified, without revealing what was redacted.
Anthropic adds that architecture information carries a high leak risk and typically low relevance, so sharing it requires explicit internal approval. That is a sensible security reason, and it is also the category an outside reader would most want to inspect if they were asking whether a model is built the way its maker says.
Who else is involved
SecureBio is described as a first-tier evaluator, reviewing the chemical and biological weapons chapter of Anthropic's August 2026 Risk Report after an earlier pilot review.
Accenture and Faculty are working toward a broader partnership covering model evaluation and red-teaming, alignment assessments, safeguard testing and capability building.
Anthropic says it is in early discussions with other organisations, including Hugging Face.
Anthropic also says it is embedding evaluators voluntarily and encourages every frontier AI company to do the same. Whether others will is not something the announcement can answer.
What builders can take from it
This is my reading, not Anthropic's. If your product depends on a model provider, the practical question is which of these layers exists for the vendor you use, and what is published. Our overview of who tests an AI model before release maps the layers of pre-release testing and what each one leaves uncovered, and that uncovered column is a reasonable start for your own test plan. For the broader landscape of what can go wrong with AI products, the AI risks guide is the pillar for this topic.
Two things to watch for next: whether METR actually publishes, and whether any redaction is flagged as unjustified. Those are the points where the structure gets tested.
For related Anthropic safety work, see Anthropic's automated alignment researcher.
Frequently asked questions
What is METR?
Anthropic describes METR as a nonprofit that evaluates frontier AI models. It says METR has reviewed several of Anthropic's Risk Reports over the past year.
Can Anthropic block METR from publishing?
Anthropic says it cannot veto publication. It can require redactions in three predefined categories: customer data, attorney-client-privileged information and model architectures.
How many METR staff have access?
Anthropic says six METR employees have been onboarded.
Does this mean Claude models are independently verified as safe?
No. The announcement describes access and publication rights, not conclusions. The value depends on what METR publishes.
How did this land?
About the author
Staff Engineer, Platform
Carlo works on the platform that turns prompts into running apps. He writes the engineering deep dives and the changelog notes worth reading.
