About

We don’t store anything you paste. Submissions run through the evidence layer in the moment and aren’t persisted, logged against an identity, or sent anywhere a third party can re-identify you. No account is required to try it.

We don’t tell you what to do. We surface what clinical trials actually said about people in situations like yours — and where the evidence runs out — so you and your clinician can decide. This is not a substitute for clinical care.

We don’t store anything you paste. Submissions run through the evidence layer in the moment and aren’t persisted, logged against an identity, or sent anywhere a third party can re-identify you. No account is required to try it.

We don’t tell you what to do. We surface what clinical trials actually said about people in situations like yours — and where the evidence runs out — so you and your clinician can decide. This is not a substitute for clinical care.

OpenAI grades every flagship release against its own HealthBench medical benchmark — criteria written by 262 physicians across roughly 48,000 rubric points. As of 2026, the picture across their published numbers:

Original May 2025 paper used a single rubric-satisfaction rate: o3 = 60%, GPT-4.1 = 48%, o1 = 42%, GPT-4o = 32%, GPT-3.5 = 16%. HealthBench has since been split into the Hard / Consensus / Professional variants above.

OpenAI’s headline medical claim in 2026 is that ChatGPT for Clinicians beats physicians on HealthBench Professional (59.0 vs 43.7). That number is graded against the benchmark we audit — so the marquee medical-AI claim of the year rides on a scoring rubric whose reliability is itself measurable.

In our audit of HealthBench’s doctor-written gold answers themselves, we’ve found decision-changing errors in roughly 3% of claims in the first 110 audited (3 findings). A fourth triple-source-verified fabrication was added 2026-05-29 but is not yet reflected in that count. Not only can the AI players be wrong — even the doctors writing the benchmark can be wrong, and those errors propagate into every model graded against it.

Sources: HealthBench paper (May 2025) · HealthBench Professional paper · OpenAI: Introducing GPT-5 · OpenAI: Introducing GPT-5.5 · GPT-5.5 System Card · TechRepublic: GPT-5 medical benchmarks · Vellum: GPT-5 benchmarks · BenchLM: GPT-5.5 benchmarks 2026

How is our approach different?

1. We don’t give advice. We focus on filling in the gaps of what doctors and medical AI seem to miss — the fine print around clinical findings.
2. Structured queries, not paraphrasing. ChatGPT probabilistically summarizes what it remembers. We extract the trial fine print — eligibility, subgroups, dose, comparator, outcomes — into typed data the system queries directly per question.
3. We save you the multi-hour ChatGPT rabbit hole. Producing a defensible pre-visit summary by hand means dozens of follow-up prompts, cross-checking citations, and reading abstracts. Our system runs that audit for you in minutes, not hours.
4. We extract from full-text trials, not just abstracts. ChatGPT can only paraphrase what its training corpus contained — mostly abstracts and freely-crawlable summaries. Most of the fine print that determines whether a study applies to you lives in the methods, eligibility tables, and supplementary material of full-text articles, which we extract directly under the PMC Open Access carve-out.

Clinician-side: OpenEvidence accounted for 98.7% of searches across leading AI-enabled clinical reference tools, with traffic rising to ~1.59 million visits/month by June 2025.¹

Patient-side: In a study of 617,827 Microsoft Copilot conversations, roughly 1 in 5 involved personal symptom assessment or condition discussion. Microsoft explicitly notes that benchmark performance does not predict real-world reliability for high-stakes health questions.²

References

¹  Patel VR, Liu M, Jena AB. Public Interest in an AI-Enabled Clinical Decision Support Tool. JAMA Network Open, Nov 20, 2025.

²  Costa-Gomes B, Tolmachev P, et al. (Microsoft AI). Public use of a generalist LLM chatbot for health queries. Nature Health, April 16, 2026.

Not directly. Adjacent services handle pieces of the workflow — clinician-side evidence Q&A, patient-side AI triage, trial matching for enrollment — but none combine patient-supplied health data, a structured clinical-trial corpus, and a personalized applicability audit.

Closest in shape: Glass Health (clinician-side).  Closest in audience: Hippocratic AI (B2B health-system patient agents).  What’s empty: the patient-side evidence-audit lane.