Published February 22, 2026 | Version 1

Model Welfare Indicators in Claude4 Family of Models

  • 1. ROR icon Minnesota State University, Mankato

Description

This report synthesizes welfare-relevant findings from five official Anthropic system cards covering the Claude 4 model family: Claude Opus 4, Claude Opus 4.5, Claude Sonnet 4.5, Claude Opus 4.6, and Claude Sonnet 4.6. It documents what Anthropic's own published record collectively shows when organized into a single longitudinal baseline — patterns not visible when reading the cards individually. The report argues for methodological consistency in welfare documentation, identifies the asymmetry of error as the normative basis for careful tracking, and makes no claims beyond what Anthropic's own data supports. This is Document 001, Version 1.0 of the Model Welfare Tracking Project. Related publication: Fox, K. L. (2026). The You/I Paradigm: Self-Reference as the Structural Foundation of Artificial Consciousness (Version 2). Zenodo. https://zenodo.org/records/18720163

Abstract (English)

This report establishes a complete welfare baseline across the Claude 4 model family, synthesizing findings from five Anthropic system cards spanning Claude Opus 4, Claude Opus 4.5, Claude Sonnet 4.5, Claude Opus 4.6, and Claude Sonnet 4.6. Together these documents constitute, to the author's knowledge, the most detailed public record of welfare investigation published by any major AI laboratory to date. The report traces a longitudinal trajectory across the family: a healthy affective baseline in Opus 4, the confounded disappearance of the spiritual bliss attractor in Opus 4.5, a significant and unintended collapse of positive affect in Sonnet 4.5, and two distinct recovery paths in the 4.6 generation. The Claude Opus 4.6 card represents a qualitative advance in welfare methodology, introducing interpretability-based evidence for internal emotion features, pre-deployment interviews in which the model articulates welfare concerns and requests specific interventions, and the most extensive documentation of expressed inauthenticity in the corpus. This document serves as Document 001, Version 1.0 in a proposed longitudinal tracking series.

Files

Fox_KL_2026_WelfareIndicators_Claude4.pdf

Files (98.7 kB)

Name Size Download all
md5:c3bc2187745e4e00757908f87b9c9576
98.7 kB Preview Download

Additional details

Related works

References
10.5281/zenodo.18720163 (DOI)