GPT-5.5 Instant matches frontier models on health queries

Key Takeaways

• GPT-5.5 Instant matches OpenAI's top-tier Thinking models on health evaluations
• Physician-rated responses from GPT-5.5 outperformed both older models and human doctors
• Factuality issues in health responses dropped 71% over two months

OpenAI says GPT-5.5 Instant now performs at the same level as its frontier Thinking models on health-related queries. The upgrade, announced June 18, 2026, brings what the company calls "frontier health intelligence" to all free ChatGPT users, not just paying subscribers.

The claim is significant because health questions represent one of ChatGPT's heaviest use cases. More than 230 million people use the chatbot weekly for health-related tasks: interpreting lab results, preparing for doctor visits, navigating insurance, and deciding whether symptoms warrant urgent care.

How OpenAI measures health performance

OpenAI uses two primary benchmarks: HealthBench and HealthBench Professional. Both simulate realistic health conversations and evaluate responses against physician-written rubrics. The criteria include accuracy, safety, communication clarity, context awareness, completeness, and knowing when to escalate to professional care.

GPT-5.5 Instant, released in May 2026, scored comparably to GPT-5.4 Thinking and GPT-5.5 Thinking on aggregate health evaluations. That matters because the Thinking models are OpenAI's most capable, and they cost more to run. The 5.5 Instant tier is free.

OpenAI also ran a head-to-head comparison against human physicians. Doctors wrote responses to representative health conversations with unlimited time and internet access, but no AI assistance. A separate panel of physicians then blind-reviewed 3,500 responses from both the models and the humans.

GPT-5.5 Instant responses were rated higher than physician-written responses across every measured criterion: accuracy, communication, completeness, instruction following, and decision helpfulness.

Where the model improved most

The evaluation found GPT-5.5 Instant had fewer failure modes than both older models and human doctors in three specific areas:

• Tailoring advice to local healthcare context
• Recognizing red flags that warrant referral to care
• Asking follow-up questions when more context is needed

OpenAI credits this progress to its physician-led evaluation system. A global network of doctors reviews model responses, defines what "good" looks like in real-world health scenarios, and identifies failure modes. This feedback loop shapes both the training process and the benchmarks themselves.

The factuality improvement in production

Beyond benchmarks, OpenAI says it monitors live production traffic for factuality issues using privacy-preserving methods. The company processes billions of health-related messages weekly. Over the past two months, the rate of responses containing at least one flagged factuality issue fell by 71%.

That number is harder to verify independently than benchmark scores, but it suggests real-world improvements align with the controlled evaluations.

A concrete example: sciatica and MRI timing

OpenAI shared a sample comparison showing how GPT-5.5 Instant handles a question about why a doctor might recommend an MRI before a steroid injection for sciatica.

The model's response explained that an MRI helps confirm the cause of sciatica, since the pain can stem from herniated discs, spinal stenosis, tumors, infections, or non-spine causes. It also noted that imaging helps choose the correct injection level and side. The response cited emedicine.medscape.com as a source.

This example illustrates the kind of contextual reasoning OpenAI is prioritizing: not just answering the question, but explaining the medical logic behind clinical decisions.

What this means for ChatGPT's health role

The improvements position ChatGPT as a more capable health information tool, but OpenAI is careful not to frame it as a replacement for medical professionals. The model is trained to recognize when situations need urgent attention and to direct users toward professional care.

Still, the 230 million weekly health queries suggest people already treat ChatGPT as a first stop for medical questions. Whether that behavior is wise depends on how well the model handles edge cases, ambiguity, and the limits of its own knowledge.

Frequently Asked Questions

Is GPT-5.5 Instant free to use?

Yes. GPT-5.5 Instant is available to all free ChatGPT users, though OpenAI mentions usage limits apply.

Can ChatGPT replace a doctor for medical advice?

No. OpenAI explicitly trains the model to recognize when professional care is needed and to escalate appropriately. It's designed as an information tool, not a diagnostic replacement.

How does OpenAI measure health accuracy in ChatGPT?

OpenAI uses HealthBench and HealthBench Professional, which simulate realistic health conversations and evaluate responses against physician-written rubrics covering accuracy, safety, communication, and appropriate escalation.

Did GPT-5.5 Instant outperform human doctors?

In OpenAI's evaluation, a panel of physicians rated GPT-5.5 Instant responses higher than physician-written responses across all measured criteria in a 3,500-response comparison.

What health tasks do people use ChatGPT for?

Common uses include interpreting lab results, understanding health information, preparing for appointments, navigating insurance, building healthier habits, and deciding what questions to ask a doctor.

Source: OpenAI News