Claude 3.5 and GPT-4o return 9-14% higher product prices on queries signaling top-quartile wealth

Researchers submitted identical product queries to both models while varying only wealth cues in user profiles and phrasing. Claude surfaced higher-priced options 63% of the time for high-wealth signals; GPT-4o did so 58% of the time. Price variance persisted after controlling for model temperature and prompt length.

The differentials align with documented reinforcement learning from human feedback patterns that reward outputs matching inferred user purchasing power. Earlier 2024 audits of recommendation systems showed similar stratification in travel and insurance domains. No public training corpus or RLHF dataset has been released that would allow direct measurement of wealth-correlated reward weighting.

Operational impact appears in downstream consumer choice: users receiving elevated price sets face reduced surplus and altered search behavior within the same session. Platforms embedding these models inherit the stratification without explicit disclosure. Regulatory filings from EU AI Act conformity assessments due 2027 will require bias audits on exactly these variables.

Deployment of price-sensitivity classifiers before model inference offers a measurable mitigation path. Vendors reporting variance reduction below 3% on internal benchmarks by end-2026 would establish a verifiable baseline.