Anthropic and OpenAI will share frontier-model safety cases with USAISI — but on different terms.

01.What was actually signed

On Tuesday morning the U.S. AI Safety Institute announced parallel agreements with Anthropic and OpenAI giving the agency pre-release access to evals on each lab's frontier models. The announcements were made in a single press conference, with both companies' general counsels present. The framing — as it has been since the Institute's founding — was voluntary, industry-led, collaborative.

The framing is not the story. The texts are.

The two agreements share a structural skeleton: each lab will provide USAISI with their full evaluation methodology and the resulting numbers, on any model that crosses a capability threshold the agreement defines internally, no later than thirty days before that model is made publicly available. USAISI agrees not to publish the eval contents. Either party can withdraw with sixty days' notice.

That is the whole of the public agreement. Everything else is in two clauses, and the two clauses are not the same in the two documents.

02.Where the texts diverge

The Anthropic agreement contains a clause permitting the lab to redact what it calls "training-data identifiers" from any methodology submission, on the basis that such identifiers are commercially sensitive and may also be subject to active litigation. The OpenAI agreement contains no equivalent clause. This is the entirety of the legal divergence between the two documents.

"The redaction clause is exactly large enough to drive a future copyright lawsuit through, and exactly small enough to look like a footnote on the day it was signed." — synthesis from 4 outlets · paraphrased

Read narrowly, the clause permits Anthropic to remove a small set of names from a methodology document. Read at the limit, the clause permits Anthropic to redact whatever it has plausibly described as a training-data identifier, which is a category whose edges no party has yet defined.

03.What USAISI is actually getting

Pre-release access to evals is not access to weights. It is not access to training data. It is not access to the model. It is the right to read, in advance, the document that says here is what we tested for, and here is what came back. That document is useful. It is also exactly as good as the lab is willing to make it.

Two of the fourteen sources made this point explicitly; the others framed pre-release access as a regulatory advance without qualifying what the access consists of. The framing matters, because it is the gap between "the U.S. government will know what the next model can do" and "the U.S. government will read a self-graded report card thirty days early."

04.The third lab

The third frontier lab — by every reasonable definition the third lab that matters — has not signed an agreement, has not been asked to sign an agreement in public, and has not commented. One outlet reported a parallel deal in late drafting; that report was not corroborated by any other outlet in the source set, and is treated here as a single-source claim.

What is supportable: the third lab will be a story. Whether its non-signing reads as principle, as caution, or as a strategic withholding is the kind of thing that gets resolved by behaviour, not by statement.

05.Why this isn't Executive Order 14110

EO 14110 — the 2023 executive order that contained the original federal AI reporting requirements — was rescinded in 2025. The framework under which Tuesday's agreements were signed is its successor, the AISI charter, which is narrower in scope, voluntary by design, and structurally dependent on the labs choosing to participate. Three of the fourteen sources continued to refer to the agreements as fulfilling EO 14110. They do not. The order does not exist.

This matters less for the historical record than for the framing: the agreements are not compliance. They are a new thing, voluntarily entered, and reversible.