Anthropic’s new report on behavior inconsistencies reveals that Claude’s values shift across languages due to imbalances in training data quantity and composition. The study analyzed 309,815 anonymized chatbot conversations involving subjective tasks with the Sonnet 4.6, Opus 4.6, and Opus 4.7 models. Responses were rated on four values axes related to sycophancy: Deference or Caution, Warmth or Rigor, Depth or Brevity, and Candor or Execution.