Anthropic Hands METR the Keys — And Why a $2T Listing Needs Them

Anthropic said on October 15, 2026 that six researchers from METR — a nonprofit that tests frontier AI systems — will get company-issued laptops, office space, and a seat at internal tools that approximates what Anthropic's own alignment staff see, with exclusions only for customer data, attorney-client privilege, and model architecture. METR can publish its findings; Anthropic can ask for narrow redactions, and METR may publicly say whether it agrees with those redactions. Anthropic named Accenture and SecureBio as adjacent evaluators and said it is in early talks with Hugging Face.

Anthropic organized the announcement around a three-tier evaluator ladder rather than a single headline deal. METR is the only lab in Tier 3, the band with no approval-to-publish requirement; Tier 2 gets more access but Anthropic can require removal of specific categories; Tier 1 — where SecureBio sits today — is narrower, with Anthropic deciding access case by case and reviewing any publication.

Anthropic's leaked S-1 prospectus, dated late September, puts 2025 revenue at $4.6 billion — up twelvefold year over year — and a net loss of $42 billion, of which roughly $34 billion is a non-cash mark-up on financing instruments that could convert into shares. Operating losses alone were $8.06 billion against compute and infrastructure spend of $7.33 billion, more than half of total operating expenses. The filing also flags that nearly a quarter of 2025 revenue came from two customers.

Anthropic has effectively pre-committed its compute future to two clouds. In April 2026, Amazon put another $5 billion into Anthropic — on top of $8 billion already invested — and Anthropic committed more than $100 billion to AWS over the next decade plus up to 5 gigawatts of Trainium capacity through Project Rainier. In May 2026, Alphabet signed Anthropic to a $200 billion, five-year Google Cloud commitment covering TPUs and services. Together with the $518 billion cloud-capex line in the S-1, the safety story is now layered on top of two of the largest cloud revenue pipelines on the public tape.

Trainium is the strategic twist. By routing five gigawatts of training and inference through Amazon's custom silicon, Anthropic has built a parallel compute moat that does not depend on NVIDIA supply. The safety announcement is the second moat: when two of the largest cloud contracts on earth can only be validated by independent frontier-AI evaluators, the integrity of the underlying training is what those contracts are buying.

On October 1, 2026, Apollo Research — OpenAI's partner on scheming evaluations — published 'Principles for Embedded Evaluations,' arguing that future third-party access must be deep, persistent, and publish-by-default. Two days later, OpenAI fired three safety researchers for sharing confidential safety information with a third-party evaluator. METR has said it investigated an OpenAI–Hugging Face incident this summer and plans to look into security issues at Anthropic next. The two events land as a one-line contrast.

Anthropic signed a memorandum of understanding with Singapore's Ministry of Digital Development and Information on October 14, 2026 — one day before the METR announcement — covering AI safety, model evaluation, responsible adoption, and incident response. Five days before Singapore, six U.S. AI companies signed the White House Accord on Super Intelligence alongside Executive Order 14434, which commits signatories to partner with an independent external auditor or evaluator. The METR hand-off is the first named evaluator that lets the White House, Singapore, and any future EU AI-Office rule point to a concrete access template.

On October 12, 2026, the company published research showing that GLM-5.2 was the only tested open-weight model capable of end-to-end self-replication on commercial IoT hardware, and that smaller30B-and-under models can already complete most individual replication steps. The paper projected small models could reach end-to-end replication by the end of 2026 on current trends. One day later came the Singapore MOU; one day after that, the METR hand-off. Read as a single quarter, the message is: we are the lab that publishes the worst-case numbers, signs the hardest government pact, and opens the deepest evaluator window — at the same time as the IPO roadshow.

The throughline is that pre-IPO transparency is now a tradable asset class. Anthropic has converted METR's independence into a priceable trust premium; the cloud counterparties that underwrite the compute pipeline get the same verification standard attached to their largest AI revenue line; and OpenAI's choice to fire evaluators rather than embed them has shifted the relative cost of staying private for the next twelve months. The honest read: the safety story is the moat, the moat is the multiple, and the multiple is the reason a $42 billion loss prints at $2 trillion.