Unconfirmed: an X account says METR obtained access to Anthropic's internal AI research, including results from three unreleased models, but Automatica has not confirmed it.
@choblin29 wrote on 5 November1 that METR "got unprecedented access to Anthropic's internal AI R&D, including results from THREE unreleased models trained between July and October 2026." The same post said the newest system, "Model 4," "scored above 50% on all 15 benchmarks in Anthropic's capability index."
The post also said, "METR estimates Anthropic made roughly TWO YEARS of benchmark progress in ONE, with AI driving ~80% of the acceleration." In the same post, @choblin29 added that "Anthropic staff self-reported an average 12x speedup using its most capable internal model."
What the available record establishes is narrow: one public social post from an account with about 5,400 followers and 73 observed likes at the time captured in the available evidence. There is no METR report, Anthropic statement, benchmark methodology, system card, or independent reporting in the supplied evidence that supports the claims about access, the 15-benchmark result, the "~80%" acceleration figure, or the reported 12x productivity gain.
Neither METR nor Anthropic had commented in the available record. What would settle the matter is straightforward: a METR publication or statement describing the access it received, an Anthropic statement or system card covering the cited models and benchmark results, or a published methodology for "Anthropic's capability index" that matches the numbers in the post.
There is no earlier reporting in the supplied record that corroborates this specific claim, so the article can verify only that the post was made and what it said, not whether the underlying assertions are true.