Six people, a badge, and no veto
Start with what is concrete, because most independent evaluation is not. Six named-role researchers from a nonprofit evaluator now work inside Anthropic, on Anthropic laptops, in Anthropic office space, with access the company describes as roughly on par with its own risk assessment teams. Colleagues are encouraged to treat them as colleagues, at least on alignment risk. When they are ready to publish what they make of Anthropic's risk levels, incidents or practices, the company says it cannot stop them.
That last sentence is the one to hold onto. Two days ago this site reported a Reuters exclusive describing both Anthropic and OpenAI privately seeking significant control over what embedded evaluators could publish, weeks after endorsing editorial independence in public. Anthropic's response to Reuters was that evaluators it had worked with before control their own publications, while it reserved comment and approval rights for new firms. The October 15 post puts a version of that answer in writing, with a name attached. For METR, the established evaluator, there is no approval right. That is a checkable commitment. If a METR report is ever delayed, softened or blocked, the gap will be visible against the company's own text.
The tiers do the quiet work
Read past the top tier and the structure changes character. Anthropic is building three tiers of evaluator access. In the first, the company decides what an evaluator sees and approves what it publishes. In the second, an evaluator sees more and can publish without approval. Only the third adds the inside access METR now has. Anthropic says it hopes, over time, to have several evaluators at the third tier, each publishing its own analysis. Today there is one. SecureBio is explicitly placed in tier 1, reviewing a single chapter of a risk report Anthropic wrote, under the approval rule, following its published pilot review of the previous report.
None of this makes the METR arrangement cosmetic. It does locate it. Access here is not a standing right that qualified outsiders hold. It is a level a lab assigns, on a ladder the lab built, under agreements the lab drafted. Anthropic is explicit that the embedding is voluntary and argues every frontier company should do the same. A voluntary scheme has no answer for the company that declines, and so far no other frontier lab has announced an equivalent embed on the same terms; what any lab may have arranged privately is not public. OpenAI pledged a similar approach last month, after Anthropic's chief executive proposed it. A pledge is not a team with laptops.
What is still missing
Three gaps matter more than the bank examiner analogy Anthropic offers, and the company itself notes the analogy is not exact. A bank examiner works for a regulator with statutory authority. METR works under an agreement with the firm it examines, however well that agreement is written.
First, there is no timetable. The post gives no date for a first METR assessment, no reporting cadence and no statement of which models or incidents the team will examine first. A right to publish is worth less before anything has been published under it. The rights have not yet been exercised in public, and no disputed redaction has yet tested how the safeguard works.
Second, the redaction categories are defined by Anthropic, and one of them, information about model architectures, sits close to the subject evaluators are there to judge. METR can protest a redaction publicly, which is a genuine safeguard. It is still a protest about text readers cannot see.
Third, METR has not published its own account of these specific terms — none was located as of 15 October. What readers have is Anthropic's description of the arrangement, not the agreement itself; its full text is undisclosed. That may change when METR publishes, and its version is the one that will settle whether employee-like access felt like access in practice.
How to read the first report, when it comes
When a METR report appears, three details will tell you more than its conclusions: whether Anthropic required redactions and whether METR said they were justified; whether the scope covers incidents and internal practices, not only model scores; and whether a second evaluator has reached tier 3 by then. One embedded team is an experiment. Several, publishing on different subjects without approval, would start to look like oversight.
Sources
- [1] Anthropic, "An update on independent evaluators" (primary source; published 15 October 2026, 17:20 UTC, updated 18:25 UTC — six METR employees onboarded, hardware and office space, access comparable to internal risk teams, no veto, redaction categories and METR's right to characterise them, three tiers, SecureBio in tier 1, Accenture and Faculty partnership, Hugging Face in early discussions, bank supervision analogy, voluntary embedding)Read source
- [2] TechCrunch, Aditya Mehta, "Anthropic is letting METR researchers look under the hood with employee-level access" (published 15 October 2026, 19:14 UTC — corroborating report, three-tier summary, background on recent agent incidents and the September essay behind the proposal)Read source
- [3] ALFW, "The Labs Endorsed Independent Reviews. Then They Negotiated the Publish Button." (13 October 2026 — our earlier coverage of the Reuters reporting on publication-rights negotiations, for the before picture; this piece is its follow-up, not a repeat)Read source