Jul, 134 min to read
Frontend
News · OpenAI puts frontier health answers in the free ChatGPT tier with GPT-5.5 Instant
Jul, 94 min to read
FrontendOpenAI puts frontier health answers in the free ChatGPT tier with GPT-5.5 Instant
The model most users actually touch — the fast, free Instant tier — now matches the Thinking models on OpenAI's hardest health evaluations.
What a good health answer looks like in the interface
The sciatica MRI example shows what OpenAI is optimizing the visible response toward. The GPT-5.5 answer enumerates concrete clinical reasons — confirming cause, choosing injection level and side, checking for red flags like infection or cancer — then explicitly notes MRI is not always required and points to studies questioning routine imaging.
Two frontend behaviors stand out. The response carries inline citations to Medscape and PMC articles, and it ends by handing the user a specific question to ask their doctor: "What are you looking for on the MRI, and how would the result change the injection plan?" OpenAI frames progress as recognizing urgency, explaining uncertainty without overstating confidence, and clarifying next steps — all of which are qualities a user reads directly off the screen.
The physician network behind the rubrics
OpenAI grounds the claims in a review pipeline rather than a single benchmark score. It cites a global network of more than 260 physicians across 60 countries, 49 languages, and 26 specialties who have reviewed over 700,000 example responses — described as roughly one review every few minutes — turning that feedback into rubrics and evaluation criteria.
The head-to-head test is notable: physicians wrote responses to representative health conversations with unlimited time and internet access but no AI, and a separate physician panel compared those against model responses across 3,500 reviewed items. OpenAI reports GPT-5.5 Instant was rated higher than both older models and physician-written answers, with fewer failure modes such as missing red flags, failing to tailor to local healthcare context, or not seeking more context from the user.
A production metric, and what it leaves unstated
Beyond benchmarks, OpenAI offers a live-traffic figure: privacy-preserving monitors on billions of health messages a week show the rate of responses flagged for at least one factuality issue has fallen 71% over the last two months. That is a claim about the deployed product, not a lab score, which is the more relevant number for a default free model at this scale.
What the post does not quantify is the residual — the flagged rate that remains, or how the automated factuality monitor itself is validated against physician judgment. For a system now serving frontier-grade health answers to hundreds of millions of free users by default, the specific implication is that the safety margin now depends less on which tier a person can afford and more on how well the physician-authored rubrics and production monitors catch the errors that stay in the interface.
Find this story relevant to you?
Contact us to find a unique solution
Need an AI engineering partner that can actually build?
We help businesses integrate AI, build AI-powered products, automate high-value workflows, and modernize the software systems behind them.
Related reading
More analysis around product delivery, operational AI, and the systems work that makes deployment hold up in reality.