GPT-5.5 Instant improves health answers to physician level

OpenAI's latest model release shows progress in a high-stakes domain: health advice. The company measured improvement using a combination of automated benchmarks like HealthBench and direct comparison against physician-written responses, where GPT-5.5 Instant came out ahead. The scale is notable—over 230 million people use ChatGPT for health questions weekly, and production traffic shows a 71% drop in flagged factuality issues in two months.

The progress builds on a structured feedback loop. A global network of over 260 physicians reviews responses, providing rubrics that guide model behavior. This physician-led approach likely helped the model improve on specific failure modes, like missing red flags or failing to ask for context.

The stakes are high. Better health responses could make a real difference in people's lives, but the model's performance on the most complex clinical situations remains to be fully tested. As OpenAI expands its health tools, including ChatGPT for Clinicians, the real-world impact will depend on whether these improvements hold up in practice, beyond the evaluations and production monitors.