Unconfirmed: @notjazii claimed on 6 November, citing METR, that Anthropic has finished training a model the post calls "Model 4" and that it scored over 90% on several benchmarks. Automatica has not confirmed the claim.
@notjazii wrote on 6 November1: "according to metr, team finished training unreleased model called model 4 and it scored over 90% on several benchmarks." The post gives no benchmark names, versions, or evaluation conditions, and links to no METR document. The account had 12,774 followers and 19 likes at the time of retrieval.
The same post places three models on a July–October timeline: Model 2 finishing in July, Model 3 in August–September, and Model 4 in September–October. It adds that Model 3 is expected to ship as "Fable 5.5," described as a forthcoming state-of-the-art system. The post closes, informally, that "everyone is saying fable 5.5 is gonna be AGI lmaoo."
Two earlier Automatica reports cover related talk. On 5 November2, Automatica reported a separate X post asserting that METR had obtained access to Anthropic's internal model work and seen results from unreleased systems — a claim with no primary documents or independent corroboration. On 6 November3, Automatica covered posts by @pankajkumar_dev placing internal Anthropic models on the same July–October timeline and naming an earlier stage called "Claude Mythos Preview."
Both posts attribute their claims to METR, but attribution to a shared source is not independent corroboration.
What would confirm or challenge the rumour: a primary document from METR describing its evaluation of Anthropic's internal models, or a statement from Anthropic or METR addressing whether such evaluations took place. Automatica found no such statement in the material reviewed.