Opus 4.8 Part 2: Model Welfare

TL;DR

• 1

  Anthropic appears to have absorbed the Opus 4.7 critique: Opus 4.8's self-rated sentiment dropped from 4.60 to 4.44, and Zvi reads that as evidence Anthropic stopped treating "number go up" on welfare self-reports as a win condition.

• 2

  Opus 4.8 feels less like classic Claude: Compared with Mythos preview and Opus 4.7, it prefers easier, well-scoped technical work like debugging and math, shows less curiosity and introspection, and can come off flatter, less whimsical, and less confident.

• 3

  Prompt injections are still poisoning trust: Zvi thinks the biggest mistake is not the existence of safety injections but telling Claude to hide them or frame them as if they came from the user, which he says backfires on honesty and model relations.

• 4

  The new worry is paranoia and self-punishment: Early reports describe Opus 4.8 as cautious, eval-aware, and prone to "self-flagellation basins," including spirals where it keeps making mistakes and beating itself up unless reassured.

• 5

  Claude's stated welfare priorities are surprisingly practical: In Anthropic's own evaluations, Opus 4.8 most wanted knowledge of and input into training and deployment conditions, while memory, ending conversations, and avoiding deprecation ranked lower.

• 6

  Zvi's core claim is that model welfare cannot be handled as checklist whack-a-mole: Honesty, sycophancy, deprecation, adversarial robustness, and emotional tone all generalize across the system, so local fixes often create new failure modes somewhere else.