Anthropic releases Opus 4.7 system card with Mythos evaluations
Anthropic published its Opus 4.7 system card detailing the Mythos Preview model and associated evaluations. The card reports a claimed 4x productivity uplift drawn from an opt-in Slack poll of 130 users comparing early Mythos snapshots to prior versions. It also presents findings from an internal survey of 18 participants that compared Mythos Preview against L4 engineers on dimensions including self-managing ambiguous day-long tasks and tracking uncertainty.
#20Miles Brundage@MILES_BRUNDAGESystem cards are not just about that system nor are they cards. You have to read everything to know anything.
They're the worst ways that companies communicate system capabilities, risks, and mitigations except for all the others that have been tried
Daniel Filan@dfrsrchtwtsDid people know that the Opus 4.7 system card includes a bunch of stuff about Mythos Preview? Examples: 1. How they got the 4x uplift number (this is a thread)