OpenAI Launches Free ChatGPT for Clinicians — GPT-5.4 Scores 59.0 on HealthBench Pro, 35% Above the Physician Baseline
OpenAI launched ChatGPT for Clinicians on April 23, 2026, extending free access to verified medical professionals across the US. The product is built on the company’s Healthcare stack with HIPAA compliance, role-based access controls, data residency options, and clinical search grounded in millions of peer-reviewed studies and clinical guidelines. Responses include citations with title, journal, author, and publication date — a direct response to clinician complaints that general-purpose AI assistants cite nothing.
The product is not new infrastructure. It is a specialised interface on top of GPT-5.4 with healthcare-optimised post-training, audit logs, and enterprise governance tooling. The marketing angle is clinician burnout: the average US physician spends roughly two hours on documentation for every one hour of patient contact. OpenAI is positioning ChatGPT for Clinicians as administrative triage — consult notes, referral letters, coding support — rather than diagnostic AI.
HealthBench Professional
Alongside the product, OpenAI released HealthBench Professional, an open benchmark evaluating LLMs on clinician-specific tasks. The benchmark covers 5,000 multi-turn conversations evaluated against rubrics developed with input from the American Medical Association and the National Library of Medicine.
GPT-5.4 scored 59.0 on HealthBench Professional. The physician baseline sits at 43.7 — a 35% gap in OpenAI’s favour. The model leads across consult documentation, referral letters, and medical research summarisation. The benchmark also surfaces where AI underperforms: context-seeking, global health scenarios, and adaptive communication to non-expert patients remain weak spots across all evaluated models.
HealthBench has tracked AI progress against physician-written responses since GPT-3.5. The trajectory is steep:
| Model | HealthBench Score |
|---|---|
| GPT-3.5 Turbo | 16% |
| GPT-4o | 32% |
| o3 (April 2025) | 60% |
| GPT-5.4 (HealthBench Pro) | 59.0 |
The original HealthBench, published in May 2025, was produced with 262 physicians across 26 specialties. HealthBench Professional adapts rubrics for clinician-facing workflows rather than patient-facing interactions.
A third variant, HealthBench Hard, represents tasks where performance has stalled. The current top score across all models on HealthBench Hard is 32%.
The Business Logic
Free access to clinicians is not altruism. The US has roughly 1 million licensed physicians. Capturing physician workflows creates two moats: institutional adoption data that feeds post-training, and a referral path into hospital IT procurement cycles.
Anthropic’s Claude has been used internally by several US health systems since late 2025. Google’s MedPaLM lineage is embedded in at least six Vertex AI healthcare partnerships. OpenAI’s move into free clinician access is a land-grab in a vertical where enterprise contracts run seven figures annually.
The HIPAA-compliant free tier is unusual. It suggests OpenAI is willing to subsidise acquisition cost in healthcare the same way it subsidised consumer growth with ChatGPT. Enterprise healthcare AI is forecast to be a $45B market by 2029. At that scale, front-loaded acquisition cost makes economic sense — provided the model holds its benchmark lead long enough to convert free users into paid accounts.
HealthBench Professional is open-source. Lab competitors can benchmark against the same rubrics, which creates accountability pressure in both directions.