A free clinical search tool that hospitals have wired into their records
OpenEvidence is a Miami company founded in 2022 by Daniel Nadler. It gives US and Canadian clinicians a free tool that answers clinical questions from peer-reviewed journals and clinical guidelines, with citations attached. Access requires a National Provider Identifier, the federal number that identifies US health providers (NBC News).
The FT reports that nearly two-thirds of US doctors use the tool, and that “many of America’s 1 million physicians processed more than 42 million prompts last month”. It quotes Nadler calling the company “the first viral consumer internet app in healthcare” (FT reprint). Crypto Briefing reported on 24 September 2026 that the company logged 42 million “consultations” with US clinicians in August. The company told NBC News in May that about 65% of US doctors used the tool across almost 27 million clinical encounters in April alone, with another 1.2 million users outside the US. The monthly figure comes from the company, and its unit has moved between prompts, consultations and clinical encounters.
Hospitals are adopting it as infrastructure. Mount Sinai announced an enterprise partnership in March 2026 that links the tool from the main electronic health record portal for doctors, nurses and pharmacists, and MaineHealth asks its doctors to keep protected health information out of the tool (NBC News). The FT reports relationships with several hospitals and specialist clinics, including Mount Sinai Health System and Memorial Sloan Kettering Cancer Center in New York, that embed the tool within their electronic health records (FT reprint). Paul Testa, chief health informatics officer at NYU Langone, told the FT the system sought a formal tie-up after finding clinicians already using it as an “independent resource”. Johns Hopkins Medicine announced a collaboration covering medical education and clinical care on 13 October 2026, with the two sides still defining what they will build together (Fierce Healthcare).
Two $250 million rounds, eight months apart
Business Insider reported on 24 September that OpenEvidence had raised $250 million from hospital systems and Andreessen Horowitz at a $15 billion valuation, and that a January round valued the company at $12 billion. The FT names Andreessen Horowitz, Blackstone, Thrive Capital and Sequoia among its backers, and counts nearly $1 billion invested over the previous 18 months. Business Insider’s list also includes DST, GV and Kleiner Perkins, and it counts Nvidia as an investor.
Advertising pays for the product, and the placement rules are not public
The tool is free to clinicians, and Nadler has said the core product “will always be free for users”. Advertising on the app and website brings in the revenue, including ads from pharmaceutical and medical device companies; several clinicians told NBC News the ads were unobtrusive or did not appear at all.
What is not public is the separation between advertising and the clinical answer. OpenEvidence’s terms of service state that the service is “in no way intended to serve as a diagnostic service or platform”, to recommend a particular product or therapy, or to substitute for the clinical judgment of a qualified healthcare professional. There is no published evidence that an advertiser has influenced an answer, and no published description of the controls that keep that from happening. No source explains who the largest advertisers are or how placement is decided.
The revenue figure rests on one source. The FT reports that “a person close to the company” said OpenEvidence booked $100 million in “ad buys” for this year as of the end of 2025 and is “on track to generate between $400 million and $500 million through ads by the end of 2026”. No filing or audited number confirms the range.
Advertising as the price of a free AI product is not particular to medicine. Wolters Kluwer, whose UpToDate Expert AI competes with OpenEvidence, makes freedom from advertising a selling point. “We believe healthcare AI must prioritize patient safety, transparency, and freedom from advertising,” spokesperson Suzanne Moran told NBC News.
The JAMA study: 271 million queries, most about clinical decisions
JAMA published a cross-sectional study of OpenEvidence use on 15 October 2026, covering accounts verified against National Provider Identifiers between 1 January 2024 and 30 June 2026 (JAMA; TechTarget summarised the underlying Harvard release). It counts 271 million queries from 952,008 clinicians, 556,200 of them physicians. Monthly active users rose from 355 in January 2024 to 703,265 in June 2026, and average monthly queries per active user rose from 18.0 to 39.5. Just over half the queries, 51.9%, were classified as clinical decision-making, 33.1% as treatment and management, and 13.9% as diagnosis. Large shares came from outside the clinician’s own specialty.
The study measures use. It contains no accuracy, safety or patient-outcome measure. OpenEvidence did not fund it, but three of the seven authors work for the company, which supplied the de-identified data.
It also shows the limits of the company’s share claim. The 703,265 monthly active users in June 2026 are clinicians rather than physicians, and the physician count across the whole 30 months is 556,200. That does not confirm the company’s share claim, which it gave to NBC News in May as “about 65%” of US doctors and which the FT reported in October as “nearly two-thirds”. JAMA’s average query rate implies roughly 28 million queries in June 2026, against the more than 42 million the company reported for a later month; growth over the following quarter could explain the gap, and no independent count covers it.
An independent study put frontier models ahead of OpenEvidence on three tests
Nature Medicine published an independent comparison on 12 June 2026 from researchers at NYU Langone and the University of Texas at Austin (Nature Medicine). It tested OpenEvidence and UpToDate Expert AI against GPT-5.2, Gemini 3.1 Pro and Claude Opus 4.6 on 500 licensing-exam style questions, 500 HealthBench items and 100 de-identified queries from NYU Langone’s own clinical GPT instance, scored blind by 12 clinicians. General-purpose models led on all three. OpenEvidence scored 89.6% on the exam-style questions, behind Gemini at 97.4%, GPT at 94.2% and Claude at 90.2%. On HealthBench it scored 62.6 against GPT’s 88.0. On the real queries, rated 1 to 4, it scored 3.24 against Gemini’s 3.62, close to the 3.27 of Google’s automatically enabled AI Overview. It was lowest on clarity, at 2.84, which the authors read as a communication weakness rather than a knowledge gap. Rates of flagged harmful content and hallucination did not differ between models.
The authors write that clinical tools “may carry institutional legitimacy and are likely safe for routine use” but “are not superior to frontier models on knowledge, communication or clinical alignment”. They also state a limitation that applies across the market: proprietary clinical AI cannot be examined from the inside, so they queried the two clinical tools through a browser, and they describe the results as “a snapshot of a rapidly evolving landscape rather than a permanent ordering”.
OpenEvidence responded on LinkedIn that the paper’s benchmarks are “clearly ill-equipped to judge clinical decision support tools”, alleging an “obvious” conflict of interest and inattention to data contamination (TechTarget). Days after the Nature Medicine paper appeared, STAT reported that the company had approached a UCSF researcher to run a comparison of its own, supplying the data, paying the survey respondents and co-developing the collection plan without disclosing its involvement. TechTarget relayed those details on 10 July 2026. The resulting preprint, submitted on 27 June 2026, reports 620 real point-of-care queries and 149 practising physicians rating answers head to head, with the specialised tool ahead on all five dimensions by 25 to 39 percentage points (arXiv). It is not peer reviewed, and two of its authors also worked on the JAMA usage paper.
The two Mayo Clinic studies, and what they measured
The FT cites a Mayo Clinic study, described there as coming from “an investor in the startup”, that found OpenEvidence “provided accurate, evidence-based recommendations” in all five sampled cases while “its impact on treatment decisions was more limited”. The paper behind it is a clinical-nutrition comparison (Nutrition in Clinical Practice): five complex cases from the authors’ own nutrition support service, put to ChatGPT, OpenEvidence, Gemini and Copilot, with blinded scoring by five physician nutrition experts. All four produced clinically appropriate answers, Gemini scored highest on relevance and clarity, and the paper concludes that “no consistent differences in performance were observed across the five cases”. It did not test impact on treatment decisions, and its only disclosed support is a grant for statistical consultation.
A larger and later Mayo study is more favourable and more specific (Mayo Clinic Proceedings: Digital Health). It paired 125 gastroenterology eConsults with OpenEvidence answers, and five blinded internists scored both with the eSQUARE instrument. OpenEvidence scored higher on perceived quality, 32.5 against 26.1, and was judged to give an appropriate amount of information more often, 70.9% against 33.9%. A hepatologist reviewing the answers found errors in 13 of 125 (10.4%), 12 minor and one potentially major. The authors state the limits: eSQUARE measures perceived written quality and “does not independently establish diagnostic accuracy”, the error review was exploratory and done by a single study author, OpenEvidence’s answers were about twice as long as the eConsults, 639 words against 318, and length was associated with higher scores. No patient outcomes were measured.
An evidence-grading feature that nobody has graded
In July 2026 OpenEvidence launched EvidenceGrade, which rates each cited paper from A to D for quality, certainty and relevance as an answer is produced, built on the GRADE framework and a March 2026 partnership with Cochrane (TechTarget). Samuel Finlayson, the company’s senior vice president of medical AI, called it “a complement to traditional evidence synthesis” and “a starting point that will improve as the clinical community helps us refine it”. No published study compares its grades with expert appraisal or a reference standard.
The EU and UK shutdown over the AI Act
On 27 April 2026 OpenEvidence became unavailable across the EU and the UK. The page that replaced it attributed the withdrawal to “regulatory uncertainty” over the EU Artificial Intelligence Act and initially gave users contact details for named Members of the European Parliament, removed by 29 April, according to a peer-reviewed letter in The Lancet Regional Health - Europe (letter). The authors, board members of the Italian Society of Artificial Intelligence in Medicine, note that the company signed neither a 2023 industry letter critical of AI regulation nor the EU’s voluntary AI Pact, and describe the shutdown as a way of channelling professional frustration into “bottom-up” lobbying. Some of the authors disclose pharmaceutical consulting. The FT reports the exit “stymied its expansion into two key markets”.
What no study has measured
Independent peer-reviewed evaluations exist, but they are small and specialty-bound. A review in npj Health Systems in September 2026 examined 4,979 references behind 150 prompts across five medical specialties and found no fabricated references, though three abstracts had attribution errors (npj Health Systems). It assessed the citations themselves; whether those citations support the answers was not part of its test. A triage study in the International Journal of Medical Informatics found the tool missing fewer emergencies than a consumer health assistant while over-triaging non-urgent cases (Int J Med Inform). None of that work measures whether OpenEvidence changes treatment decisions, prescribing or patient outcomes, and none examines whether advertising revenue shapes what a tool used for treatment questions surfaces. The Nature Medicine authors ask for independent, real-world evaluation of AI tools before clinicians adopt them at scale, and the JAMA study’s first author, Vishal Patel, says the next questions are accuracy and effects on patient care. Two things would answer them: outcome studies run by researchers with no stake in the company, and a public account of how advertising is placed relative to clinical answers.
