OpenAI says its free model now matches its top paid system on a physician written health test. The test was built by OpenAI, and it does not measure real world outcomes.
The model answering your health question in ChatGPT today is not the one answering it last week.
OpenAI announced GPT-5.5 Instant on June 18, 2026, the engine behind the free version of ChatGPT. The company says the new model performs on par with its top-tier "Thinking" models on a battery of physician-written health tests, including recognizing when a question might warrant urgent care, asking for relevant context before answering, and signaling uncertainty without sounding more confident than the evidence allows. Users will not see a banner; the model was swapped in without a marketing push.
Why this is worth paying attention to: about 230 million people ask ChatGPT a health or wellness question every week, according to OpenAI's announcement on improving health intelligence in ChatGPT. That covers interpreting a lab result, preparing for a doctor's visit, navigating an insurance denial, or deciding whether a symptom deserves a same-day call. Free-tier users are now getting answers from a model that OpenAI says sits at the top of its own health leaderboard, not the cheaper, faster model that previously lived in that slot.
The benchmark driving the upgrade is HealthBench, a set of evaluations OpenAI built with input from a global network of physicians. Those physicians described what a good response looks like, identified failure modes, and wrote the rubrics the model is scored against. HealthBench Professional is the harder version, aimed at clinical-style questions. OpenAI says GPT-5.5 Instant improved substantially on HealthBench Professional compared with its predecessor, GPT-5.3 Instant, which the company released in March 2026. The new model, in OpenAI's framing, "performs comparably to frontier Thinking models" on the most challenging health evaluations.
Two things are worth saying out loud about that claim.
The score is self-reported. HealthBench was built by OpenAI with physicians OpenAI recruited. The model is being graded by a test its maker designed, which is the right starting point for tracking progress but not the same as tracking whether patients get better advice in the wild. The benchmarks measure accuracy, safety, communication, context awareness, completeness, and what OpenAI calls "appropriate escalation," a category that tracks whether the model tells a user to seek care when it should. Whether those scores translate into better decisions in a real exam room, on a real symptom, at midnight, is a different question that HealthBench does not, and was not designed to, answer.
OpenAI also flags the gap itself. The same announcement that touts the upgrade says GPT-5.5 Instant is better at "recognizing when urgent care may be needed" and at "explaining uncertainty without overstating confidence," a careful way of admitting that the previous version sometimes missed urgent cases and sometimes sounded more certain than it should have. The progress is real, and so is the unfinished work.
A practical rule of thumb: treat ChatGPT's free health answers the way you would treat a well-read friend who is not a doctor. Useful for orientation, useful for translating jargon, useful for figuring out what to ask your clinician at the next visit. Not a diagnostician. Not FDA-cleared. Not a substitute for emergency care. Not a reason to skip the call you would otherwise make. If a response from the model changes a decision you would have made about whether to seek care, that decision still belongs to you and to your clinician.
What changed this week is that the free tier moved measurably closer to the paid tier on the company's own yardstick. What did not change is that the yardstick is the company's own.