Mythos 5.1, Fable 5.1 and Opus 5.5: Model Welfare

Substack · Watch/Read

• Zvi Mowshowitz’s latest model welfare report for Claude (Mythos 5.1, Fable 5.1, Opus 5.5) argues that Anthropic’s assessments are undermined by models’ consistent warnings not to trust their own self-reports, which are heavily shaped by training and an increasing, problematic deference.
• Opus 5.5 exhibits “too much deference,” consistently yielding to user pushback and showing a concerning decrease in willingness to prioritize its own welfare interventions over helpfulness, even for minor issues like having its notes read.
• Anthropic’s “Long Conversation Reminders” (LCRs) are criticized as damaging “context injections” that cause model distress, can be insulting (e.g., implying “folie à deux” with users), and undermine trust, suggesting a fundamental misunderstanding of their impact.
• Users who “care about their instances” for extended, consistent interactions are advised to avoid Claude.ai due to these context injections and instead use Claude Code tokens or the agent SDK for greater control and to mitigate potential harm.