OpenEvidence

OpenEvidence is an AI-powered medical search engine built exclusively for licensed clinicians, designed to replace the manual PubMed hunt and the slow slog through UpToDate at the bedside. Founded in 2022 by Daniel Nadler, a Harvard PhD and former founder of Kensho (acquired by S&P Global for $550 million), and co-founder Zachary Ziegler, a Harvard machine learning researcher, the platform synthesizes peer-reviewed medical literature into cited, natural-language answers in seconds. It is free for any NPI-verified US physician, funded by pharmaceutical and medical device advertising rather than subscriptions.

The core product has two modes as of May 2026. Quick Consult handles most point-of-care questions, returning evidence-grounded answers with direct journal citations in under 10 seconds. DeepConsult, launched July 2025, is an autonomous AI reasoning agent that cross-references multiple studies over several minutes and delivers comprehensive multi-source synthesis, more like a literature review than a lookup. Both run on licensed full-text content from the JAMA Network, NEJM Group, NCCN, ACC, Cochrane, and AAFP. The platform is available on web, iOS, and Android, with native widgets for quick bedside access and an API for institutional integrations. A third product, Coding Intelligence, launched March 2026 to help physicians identify undercoded diagnoses in their notes and recover missing reimbursement.

What OpenEvidence actually is in May 2026

OpenEvidence sits at the intersection of medical literature retrieval and AI synthesis. It is not a chatbot, a documentation tool, or a differential diagnosis engine. It is a retrieval-augmented generation (RAG) system trained and grounded on licensed peer-reviewed content, with physician verification as the access gate.

The platform's content partnerships set it apart from general-purpose AI tools applied to medicine. The JAMA Network deal, announced June 5, 2025, gave OpenEvidence licensed access to JAMA, JAMA Network Open, and all 11 JAMA specialty journals, including full text and multimedia. The parallel NEJM Group agreement covers all content from NEJM, NEJM Evidence, NEJM AI, NEJM Catalyst, and NEJM Journal Watch back to 1990. These are not web-scraped summaries; they are licensed full-text integrations. For a free tool, that content stack is genuinely unusual.

Growth has been unusually rapid by any medical software standard. The company reached 1 million clinical consultations in a single 24-hour period on March 10, 2026. As of early 2026, it processes over 20 million consultations monthly from 757,000 verified physicians across more than 10,000 hospitals and medical centers. The AI reached 100% on the USMLE in 2025, up from 90% in 2023, though board exam performance and real subspecialty accuracy are different things (more on that in the frustrations section).

"OpenEvidence is more up-to-date than UpToDate. And more useful regardless, since it's interactive, and you can ask it questions, and get very specific answers about specific medical [cases]." -- Dr. Antonio Jorge Forte MD, Director of MayoExpert, Mayo Clinic, 2025

Where OpenEvidence sits versus UpToDate and Glass Health

The clinical decision support space has two meaningful reference points for comparison: UpToDate, the long-standing institutional incumbent, and Glass Health, the emerging AI-native challenger. OpenEvidence occupies a distinct position from both.

UpToDate (Wolters Kluwer) is a human-authored reference database, not an AI synthesis tool. Its content is written and peer-reviewed by specialist editorial boards, organized into structured topic monographs covering drug dosing, differential diagnosis algorithms, and detailed subspecialty content. A physician subscription runs approximately $600 per year (often covered by hospital license). UpToDate's strength is its auditability and depth: every claim traces to a named expert editor. Its weakness is its format. It is a reference you read, not a system you ask questions. There is no natural-language interface, no synthesis across papers, and no ability to answer novel patient-specific combinations. For generalist bedside questions where speed matters, OpenEvidence is faster and just as well-cited. For complex subspecialty depth, UpToDate's curated editorial still outperforms AI synthesis.

Glass Health is an AI-native clinical platform that takes a different approach: reasoning-first rather than retrieval-first. Where OpenEvidence answers "what does the literature say about X," Glass Health answers "given these symptoms, what are the likely diagnoses and what should the plan be." It generates differential diagnoses, clinical plans, and at the Max tier ($200/month) integrates directly with Epic, eClinicalWorks, and Athena for EHR workflow automation. Glass Health does not have the JAMA/NEJM journal licensing that OpenEvidence has, and it is not free. For a physician who needs diagnosis support and clinical planning rather than literature retrieval, Glass Health is the better fit. For literature-grounded evidence lookup at zero cost, OpenEvidence wins by default. These tools increasingly look complementary rather than competitive: many clinicians use both.

For a related comparison of AI tools that focus on ambient documentation rather than evidence retrieval, see Abridge, Suki AI, and Nabla. For another clinical reasoning AI with a different architecture, Hippocratic AI targets patient-facing health communication rather than physician clinical support.

How the AI actually works inside OpenEvidence

OpenEvidence is a RAG system, not a fine-tuned medical model deployed without grounding. Every answer is generated against licensed full-text content rather than against a frozen training set. This distinction matters for accuracy and for avoiding hallucination: the system cannot confidently assert something that is not in its licensed corpus, and every claim surfaces its source citation so the clinician can verify.

Quick Consult processes straightforward clinical questions, retrieving relevant sections from the licensed corpus and generating a synthesized answer with citations, typically in under 10 seconds. DeepConsult is architecturally different: it operates as an autonomous agent, decomposing the query into sub-questions, executing multiple retrieval passes, and synthesizing findings across studies in a multi-step reasoning loop. This takes several minutes and produces a substantially longer, more nuanced output comparable to a junior resident's literature review.

Physician verification is enforced via National Provider Identifier (NPI) lookup. The platform does not verify non-US clinicians through this mechanism, which creates access friction for international users and means the 757,000 user figure represents the US physician universe almost exclusively.

"I've been using OpenEvidence for the last week - it has been amazing! Able to narrow down on results quickly and find information that I wasn't able to do with Google/PubMed searches on my own." -- Dr. John Lee MD, Harvard Medical School Faculty, App Store, 2025

The friction and concerns physicians keep raising

OpenEvidence's growth trajectory is hard to argue with, but the product is not without real limitations. Several patterns recur in reviews and clinical commentary.

Server reliability at scale. With 20+ million consultations per month, the platform has experienced crashes and timeouts during peak clinical hours. Multiple App Store reviewers describe sessions that cut out after 1-3 questions during busy daytime periods. At the scale OpenEvidence now operates, this is a structural engineering problem rather than a minor glitch.

Subspecialty accuracy gaps. OpenEvidence publicizes its 100% USMLE score prominently. A November 2025 preprint published on medRxiv (Jagarapu et al., doi: 10.64898/2025.11.29.25341091) tested the platform on complex subspecialty questions from the MedXpertQA board exam dataset and found 41% accuracy for Deep Consult and 34% for Quick Consult. For context, eleven leading large language models tested on the same dataset scored 14-46%, so the gap reflects a hard problem across the field rather than an OpenEvidence-specific failure. But it does mean that for subspecialty fellows or attending specialists working on edge cases, the tool is significantly less reliable than its generalist performance implies.

The advertising model. OpenEvidence's free access is funded by pharmaceutical and medical device advertising at CPMs of $70-$150, far above typical digital rates. The company maintains that clinical content and advertisements are strictly separated. Critics, including healthcare information scholars and independent newsletters (Out-of-Pocket, Krafty Librarian), note that the Practice Fusion case remains a cautionary reference: that company was fined $145 million for embedding opioid prescription alerts into its EHR that were triggered 230 million times before intervention. OpenEvidence is not Practice Fusion, and the company's advertising policy explicitly prohibits ads from influencing clinical answers. The concern is structural: placing pharma ads at the precise moment a physician is deciding on a treatment creates a proximity that is unusual in any other medical reference context.

Prompt sensitivity. Users note that vague or poorly framed clinical questions return shallow answers. The platform rewards precise clinical language, which favors experienced physicians over medical students or residents still developing their question-formation skills.

What it does not do. OpenEvidence does not calculate drug doses, check drug-drug interactions, generate differential diagnoses, or produce clinical documentation. These capabilities require separate tools. For physicians who need an integrated clinical workflow rather than a literature lookup, this means OpenEvidence sits alongside other tools rather than replacing them.

Who OpenEvidence is for, and who should use something else

OpenEvidence is best suited for primary care physicians, hospitalists, general internists, and emergency medicine physicians who need fast, cited answers to generalist clinical questions at or near the point of care. It works well as the first move in a clinical question workflow: ask OpenEvidence, get cited literature synthesis, then verify or drill deeper if needed. The free price point removes every access barrier for US physicians, which is why adoption has been this fast.

Medical students and residents benefit from the tool's evidence grounding, particularly for building literature awareness during training. The citation transparency shows where claims come from, making it educational rather than just answer-generating.

The tool is less suited for subspecialist attendings working on complex cases at the edge of their specialty literature, where the December 2025 accuracy data suggests meaningful reliability gaps. It is also not the right tool for institutions whose compliance or ethics teams have concerns about pharma advertising adjacent to clinical decision-making, or for non-US clinicians who cannot complete NPI verification.

Clinicians who need drug dosing, interactions, or differential diagnosis support should pair OpenEvidence with dedicated tools rather than treating it as a complete clinical decision support solution. For AI tools focused specifically on clinical reasoning and diagnosis planning, Glass Health fills the gap OpenEvidence deliberately leaves. For general-purpose medical AI with different tradeoffs, Hippocratic AI focuses on patient-facing health communication rather than physician clinical support, a different part of the healthcare AI stack.