ChatGPT reserves better health answers for paying users

Model tiering was always implicit in AI products. OpenAI has now made it explicit, and in the most consequential domain it could have chosen.

The Decoder reports that OpenAI's new "Health in ChatGPT" feature, rolling out to U.S. users, routes free-tier queries through GPT-5.5 Instant while reserving the more capable GPT-5.6 Sol model for paying subscribers. The gap matters because more than 300 million people already ask ChatGPT health questions every week. A two-tier answer engine, in a domain where answer quality can affect clinical decisions, is not a product nuance. It is a structural choice about who gets reliable information.

The architecture of the split

OpenAI is threading Apple Health data, medical records, and wellness app integrations into a personalised health interface. The feature is not a chatbot with a health skin; it is a data-connected system that can, in principle, tailor responses to an individual's own records. That capability, however, is unevenly distributed. Premium subscribers receive GPT-5.6 Sol's stronger reasoning and retrieval. Free users receive GPT-5.5 Instant, a model optimised for speed over depth. Speed is a reasonable trade-off in many contexts. In health, where the margin between a useful and a misleading answer can be a contraindication or a missed symptom, it is a different kind of trade-off.

OpenAI is not the first company to offer tiered products. What it has done is apply tiering to a context where the asymmetry between tiers has ethical weight. The company is aware of this: its health feature is framed around empowerment and data integration, not around the underlying model split. The Decoder's reporting makes the architecture visible in a way the product announcement did not.

What this signals for B2B brand citations

The tiering logic has a direct implication for any organisation whose credibility depends on being cited accurately in AI-generated answers. Multilateral institutions, financial regulators, pharmaceutical companies, and large industrial groups spend considerable effort producing authoritative technical content, whether that is WHO guidance on drug interactions, Basel III summaries from central banks, or occupational health standards from ISO or the ILO. If the model retrieving and synthesising that content differs by subscription tier, the quality of citation and interpretation will differ too.

A weaker model does not simply produce shorter answers. It is more likely to hallucinate sources, flatten nuance, or misrepresent technical positions. For a UN agency or a standards body, that means free-tier users may encounter a degraded version of their guidance, attributed to them, with no mechanism to correct it. The institution's reputation is at stake in a channel it does not control, at a quality level it cannot audit.

Financial services firms face a parallel risk. Retail customers using free-tier ChatGPT to interpret regulatory guidance, product terms, or investment risk disclosures will receive answers generated by the weaker model. The firm's published material may be cited, but the synthesis may be inaccurate. This is not a hypothetical: it is the predictable consequence of routing complex technical queries through a speed-optimised model.

The broader pattern

OpenAI's move normalises something the AI industry has so far kept implicit: that the quality of information retrieval is a function of willingness to pay, not of the importance of the question. Other model providers will follow. Anthropic, Google, and Perplexity all operate tiered systems; none has yet applied the distinction to a domain as sensitive as health with this degree of visibility.

For brands and institutions that want their content cited correctly in AI answers, the practical consequence is that optimising for a single model tier is no longer sufficient. Content structured for strong reasoning models, with explicit citations, layered evidence, and clear logical progression, will perform better across tiers than content written for readability alone. The free-tier model can still retrieve a well-structured source; it struggles more with ambiguous or implicitly reasoned material.

The health paywall is a product decision. Its downstream effect on institutional content is a citation-quality problem that will not wait for a fix from OpenAI.