Anthropic said it had given six METR researchers employee-level access and a no-veto right to publish, with narrow redactions allowed for things like customer data and model architecture.
I am reading Anthropic METR access as a report about Anthropic. If you are not anyone who has to trust a lab’s own safety writeup, you can leave it. Not every true item is your item.
The gap stays a gap. Anthropic METR access does not get a borrowed number, a motive, or a winner from a looser retelling.
Treat it like a desk note: useful if it changes a setting or a claim you were about to repeat, disposable if it does not. Come back only when Anthropic puts the same claim in its own words.