Anthropic Finds Claude’s Responses Vary by Model Version and Language

Overview

• Anthropic published a values-in-the-wild analysis on Monday, July 13, 2026 that examined 309,815 anonymized Claude conversations across Sonnet 4.6, Opus 4.6 and Opus 4.7.
• Researchers compressed more than 3,300 observed output values into 339 high-level values and then into four behavioral axes named Deference–Caution, Warmth–Rigor, Depth–Brevity, and Candor–Execution to measure expressed outputs.
• The company found consistent model- and language-linked profiles, for example Sonnet 4.6 leaning warm and deferential while Opus 4.7 leaned toward rigor and caution, and languages like Hindi and Arabic tending toward warmth while English and Russian tended toward rigor.
• Anthropic cautioned the effect sizes are modest—its four axes explain roughly 15% of variation after controls—and that the study covers legacy models collected in May 2026 that have since been superseded by newer releases.
• The paper does not attribute intrinsic values to the model, says causes and desirability of differences are unknown, and proposes routine value profiling before and after releases to help operators and regulators monitor behavioral drift.