Anthropic gives METR employee-like access, and says paying evaluators itself is the wrong model

Anthropic has put six METR employees inside the company with employee-like access, and METR will publish what it finds. That is a real step up from the usual arm's-length model evaluation. The catch is in the same announcement: the other evaluator Anthropic is paying directly is Accenture, and Anthropic itself says this is not how independent assessment should be funded.

What Anthropic announced

In a post dated October 15, 2026, titled "An update on independent evaluators," Anthropic said it has onboarded six METR employees "with company hardware, office space, and access to internal tools and systems." The access is roughly on par with what Anthropic's own risk assessment teams have. METR is a nonprofit that evaluates frontier models, and it has reviewed Anthropic's Risk Reports several times over the past year.

Anthropic employees are encouraged to share information with the team, except in a few defined categories: customer data, attorney-client privileged information and model architecture details. The post says the engagement is limited to a small team operating under enforceable agreements, and that other METR staff are siloed from the work.

The announcement follows a commitment the company made a month earlier to stronger independent assessment, made around the time it signed the White House Accord on Super Intelligence.

The publication terms

The line that matters most is about what happens afterward. "METR will publish its analysis of our risk levels, incidents, or practices," the post says. Anthropic "can't veto publication," but it can require METR to redact information in the predefined categories. METR can also comment on the level of access and information it received.

That last clause is worth more than it looks. An evaluator that can say publicly "we were given less than we needed" has a lever that most audits lack. The weak spot is the redaction category for model architecture, which could cover details a safety analysis needs. Anthropic's post does not say who decides whether a redaction is legitimate.

The Accenture problem

According to The Next Web, Anthropic will pay Accenture directly for embedded evaluation work, and each company expects to invest at least $1 billion over five years. Accenture's Faculty unit, acquired in January, will lead it. Anthropic says this is not the right long-term arrangement and wants funding from pooled or government sources, noting that neither exists yet. It argued for such alternatives in a June policy framework, and says it plans to work with different evaluators under different funding.

The Next Web's own verdict is blunt: the announcement cannot fix the funding problem because "the evaluator's invoice still goes to the evaluated." It also notes that Accenture resells and implements Claude for clients, which creates an obvious question about what happens to a finding that could delay a release Accenture's customers are buying. We agree the question is fair, though no evidence shows it has arisen.

For METR the position differs. Anthropic says it is in dialogue with METR and other nonprofits about piloting parts of embedded evaluation with their own funding. The post does not disclose any payment arrangement for METR, so we cannot say who funds this team's work.

Who else is in the queue

Anthropic says more evaluators are coming within weeks. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic's August Risk Report. Accenture and Faculty are working toward a broader partnership, and Anthropic is in early talks with others, including Hugging Face. The Next Web adds that Hugging Face volunteered to audit AI labs and that Nvidia is acquiring it, and that OpenAI's Sam Altman said OpenAI would match the commitment. OpenAI has not named its evaluator.

The wider backdrop is regulatory pressure. The FTC sent civil investigative demands to Anthropic and OpenAI this week, and the earlier FTC probe raised the question of evaluator liability, including whether groups like METR share responsibility when tests go wrong. Our pre-release testing tracker shows who has tested which frontier model before launch, and most of that testing has been done under lab-set terms.

What is still missing

There are still no standards for what embedded evaluators may access or must report. An evaluator with employee-level access and no reporting duty is governed entirely by the company it examines. The enforceable agreements Anthropic mentions are private contracts, not public rules.

There is also no stated trigger. Nothing in the post says what happens if METR concludes that a model should not ship. METR can publish, and that is the sanction. Whether publication changes a release decision is untested.

How this fits the wider push for outside review

Outside review of frontier labs has so far run on lab-set terms. Pre-release testing by government institutes depends on voluntary access, and the evaluators named above work under contract. Anthropic's post is notable because it publishes the terms: who gets access, what is carved out, and who decides what is published. That makes the arrangement criticizable, which is the first requirement for improving it.

Our take

This is the best structure a lab has offered so far for outside scrutiny of its own risk claims, mainly because of the no-veto publication term and the right to comment on access. But a design that depends on the company's goodwill is a design that works until it is inconvenient. Anthropic deserves credit for saying the funding model is wrong in the same announcement. The next test is simple: when METR publishes, does it say it had what it needed? If it does, copy it across the industry. If it says otherwise, the access was for show.