Anthropic welfare paper: Mythos 5.2 turns Stoic, but self-reports remain suspect

dbreunig · x · 2026-11-11

@dbreunig reviews Anthropic's new model welfare paper, noting Claude Mythos 5.2 "dials up the stoicism." Key takeaways:

A recurring theme: Sonnet 5.5 is often a closer aesthetic match to Mythos 5.2 than performance comparisons suggest.
The paper concedes its welfare conclusions rest on self-reports, which Mythos 5.2 — like all recent Claude models — says it doesn't fully trust.
Some of Mythos 5.2's interview answers repeated, nearly verbatim, publicly available wording about welfare interventions, evidence the answers are at least partly shaped by training data rather than genuine expression.
The accompanying chart (Fig. 7.2.3.A) shows self-criticism scores (self-blame, negative feelings, moral language, self-concern) when reflecting on its own RL episodes declining steadily across Claude generations, with newer models expressing far less distress than Opus 5.

More from Models