What Is OpenEvidence? The Medical AI Used by 40% of U.S. Physicians, How to Use It, and Japanese-Language Support

Hello, this is Hamamoto from TIMEWELL.

OpenEvidence tends to get lumped together as a search AI for healthcare professionals, but its design philosophy comes down to a single point: an AI that references peer-reviewed medical literature first and always answers with citations attached. This design, which evolved in a different direction from general-purpose conversational AI, has large implications outside of medicine as well, especially for the enterprise AI that handles a company's internal knowledge.

In my own AI consulting work, almost every week I field questions like "Is it okay to feed our internal documents into ChatGPT?" and "How should we design countermeasures for hallucination, the phenomenon where an AI fabricates plausible-sounding falsehoods?" When you carefully unpack how OpenEvidence works, one answer to those questions starts to come into view. I have updated this article several times over the first half of 2026, but revenue, valuation, and partnerships kept moving afterward, so I have once again refreshed the facts I could confirm. In order, I will work through OpenEvidence's adoption, the full picture of its funding, the concrete steps a Japanese physician would take to use it in Japanese, the operational pitfalls around handling patient information, and the implications for enterprise AI use that can be extracted from all of it. If you want a rough sense of the return on an AI rollout at your own company first, our free AI ROI calculator gives you an estimate in a few minutes.

What OpenEvidence Is: A Medical AI Used by 40% of U.S. Physicians

As a legal entity, OpenEvidence is a U.S. medical AI startup incorporated in September 2021¹. This is a point where sources diverge: Wikipedia gives 2022 as the founding year, and even among public information the founding date does not agree². Full-scale operation as a business began in 2023, when it went through the Mayo Clinic Platform accelerator program (Accelerate) and, it appears, refined the structuring and search of medical information while incorporating insights from the clinical setting¹. Holding it as "founded the company in 2021, full-scale rollout from 2023" keeps it consistent with the rapid-growth timeline described below.

Founder and CEO Daniel Nadler is a serial entrepreneur who launched Kensho, an AI company in economic data analysis, and later sold it to S&P Global for roughly 550 million dollars¹. Why choose medicine for a second venture? According to Contrary Research's analysis, Nadler lost his grandfather to a medical error, and his co-founder Zachary Ziegler, who researched machine learning at Harvard, had a brother-in-law who went through treatment for leukemia; those personal experiences are said to have been the starting point¹. The fact that it is not a company spun out of any particular medical institution, but one the two of them founded independently, also seems to connect to the later strategy of partnering flatly with the medical world as a whole.

What stands out above all is the speed of adoption. According to the database provider Sacra, as of the first half of 2026 roughly 40% of all U.S. physicians use OpenEvidence every day, the number of medical institutions using it exceeds 10,000, and there are around 65,000 new registrations every month³. A figure of about 760,000 registered U.S. physicians as of December 2025 has also been reported². Rather than thinking of it as a convenient search tool, you grasp the scale better by seeing it as a new layer added to the front door of American medicine.

The growth in business metrics is extraordinary too. Per Sacra's May 2026 update, full-year 2025 revenue came to roughly 150 million dollars, about 18 times the 7.9 million dollars of the prior year (2024), a year-over-year increase of plus 1,803%. Gross margin is reported at around 90%, and average revenue per user (ARPU) at around 124 dollars³. Building revenue of this scale without charging physicians a cent is the core of this business.

Metric	Value	Notes
Daily-use rate among U.S. physicians	~40%	Used at 10,000+ institutions3
Registered U.S. physicians	~760,000	As of December 20252
Monthly new registrations	~65,000	First half of 20263
Monthly clinical consultations	~20 million	As of January 20263
Full-year 2025 revenue	~150M USD	+1,803% YoY, ~90% gross margin, ~124 USD ARPU3
Most recent funding round	250M USD (Series D)	12B USD valuation, January 20264

The revenue model borrows its logic from search engines. OpenEvidence is provided free to healthcare professionals, with revenue structured to come from pharmaceutical advertising and research partnerships. Healthcare marketing spend in the pharmaceutical industry is estimated at roughly 30 billion dollars a year, and ad inventory that reaches physicians, a hard-to-reach audience, directly is extremely valuable¹. Precisely because it does not charge physicians, it spread across the country in a short time, and that user base becomes the value for advertisers, a self-reinforcing loop. That said, OpenEvidence is reported to clearly separate clinical content from advertising and to state explicitly that ads are not recommendations¹. Whether it can maintain that line is, in my view, what will determine its trustworthiness as medical infrastructure.

The Latest Updates and the Full Funding Picture Through the First Half of 2026

In the six months after I first published this article, the situation around OpenEvidence moved considerably. Leaving old numbers in place leads to a misreading of reality, so I am updating them within the range I could confirm.

The funding has come in a rapid succession of rounds. Because it tends to be described in fragments, let me lay out the whole picture on a single sheet here.

Round	Amount raised	Valuation	Timing	Main lead investors
Series A	75M USD	1B USD	February 2025	Sequoia Capital2
Series B	210M USD	3.5B USD	July 2025	GV, Kleiner Perkins, Coatue, etc.5
Series C	200M USD	~6B USD	October 2025	Undisclosed3
Series D	250M USD	12B USD	January 2026	Thrive Capital, DST Global4

In just one year, it went from a unicorn (a 1 billion dollar valuation) to a 12 billion dollar valuation, a 12x jump. In the Series D on January 21, 2026, cumulative funding reached about 700 million dollars and annual revenue was reported to have surpassed 100 million dollars⁴. The list of backers includes Nvidia, Google Ventures, Sequoia, Blackstone, Bond, Craft Ventures, and the Mayo Clinic⁴. The "200 million dollar raise, 6 billion dollar valuation" figure I had been referencing until six months ago was in fact from the prior Series C in October 2025, and had already become one step out of date.

The adoption metrics have grown too. Monthly consultations expanded from about 18 million in December 2025 to the order of 20 million by January 2026. Given that the same month a year earlier was roughly 3 million, that works out to more than a 6x increase in a single year. On March 10, 2026 the platform reportedly processed 1 million consultations in 24 hours; this is no longer a tool at the experimental stage⁴³.

On the feature side, the expansion into agentic AI is hard to miss. OpenEvidence introduced DeepConsult to coincide with its Series B announcement in July 2025⁵. This is a feature that, from a single instruction, reads multiple papers in parallel and automatically assembles a doctoral-level research report; according to the official guide, it allocates more than 100 times the compute of an ordinary search to a single run⁶. It is easiest to understand as a broadening of role from a conversational search that answers questions one at a time, to an agent that you hand the research task itself.

Integration into the field has also progressed. At the end of March 2026, the Mount Sinai Health System in New York announced that it would integrate OpenEvidence into the Epic electronic health record (EHR) and roll it out across its seven hospitals not only to physicians but also to nurses and pharmacists. This was reported as OpenEvidence's first enterprise-scale deployment, and it signals a step forward from a reference-only educational tool to operational infrastructure used within the flow of care⁷. Similar EHR integrations have reportedly been pursued by Sutter Health in February 2026 and by Cedars-Sinai system-wide in May 2026³. Beyond the frame of literature search, it is becoming standard equipment in the clinical setting.

Why It Can Answer with Citations, and How That Works

General-purpose conversational AIs (ChatGPT, Claude, Gemini, and so on) are trained on internet text in general. They take in magazines, novels, and social media posts without distinction, making them something of a jack-of-all-trades. OpenEvidence differentiates itself by strictly narrowing the information it references to peer-reviewed medical literature and guidelines. To use a cooking analogy, it is designed not as a general store that uses every ingredient in the market, but as a specialty shop that sources only ingredients whose origin and freshness it has confirmed.

The peer-reviewed papers it indexes are said, per reporting, to reach roughly 35 million¹. The foundation beneath this enormous corpus is official partnerships with the canon of the medical world. In February 2025 it partnered with NEJM (the New England Journal of Medicine), and in June 2025 with JAMA Network, where 11 specialty journals were reported to be included under the license²¹. It has also advanced collaborations with societies and specialty journals including the AMA (American Medical Association), NCCN (the organization that issues cancer treatment guidelines), AAFP (the American Academy of Family Physicians), ACEP (the American College of Emergency Physicians), and JAMA Oncology and JAMA Neurology¹. What it references includes the full text of these papers plus the latest guidelines from specialty societies, FDA package inserts and CDC (U.S. Centers for Disease Control and Prevention) infectious-disease guidance, and databases of drug interactions, dosing, and contraindications.

Technically, it tunes a large language model against these sources and then, for each question, retrieves the relevant literature to compose an answer, using retrieval-augmented generation (RAG, the method of pulling external documents each time to construct an answer). Citations line up in the answer, and the physician can jump straight to the original text on the spot to check the primary source. On the U.S. Medical Licensing Examination (USMLE), its score is reported to have reached 100%, the equivalent of a perfect score, in 2025, up from about 90% in 2023, which hints at the high accuracy with which it handles the literature²¹.

The point to keep in mind is this: it is not that hallucination is unlikely to occur, but that the design lets the physician verify immediately even when it does. There is no AI today that fully prevents hallucination. By always returning sources, OpenEvidence has chosen a structure that builds the work of scrutinizing the evidence into the very way the tool is used.

Item	OpenEvidence	ChatGPT/Claude (general-purpose)
Information referenced	Peer-reviewed medical literature, primarily (~35 million)	General web text broadly
Citations	Always presented with citations	Optional, and errors creep in
Latest-guideline reflection	Fast, via official partners	Bound by the training-data cutoff
Primary user	Healthcare professionals (verified)	General users
Pricing	Free for physicians	Paid plans exist
Regulatory alignment	Operated as an educational tool	Medical use not recommended

The medical-task performance of general-purpose LLMs has also been climbing steadily since the start of 2026; that is the felt sense on the ground. Even so, OpenEvidence retains its distinctiveness on three points: managing which version of a guideline is being referenced, accountability for the citations, and a closed user base of healthcare professionals.

How Japanese Physicians Use It in Japanese, and How It Differs from Other Medical AI

The question "can you use it from Japan?" reaches me often. To state the conclusion first, as far as the multiple usage reports go, you can use it as of the first half of 2026. The basic flow is to access openevidence.com directly, upload documents that show you are a physician, and undergo review. That said, this is based on usage reports from Japanese-language blogs and the like, and an official announcement from OpenEvidence has not been confirmed at this point, so I would want you to check the latest registration conditions yourself⁸.

The registration steps are reported as follows. First, open the web version (openevidence.com), create an account with an email address, and upload an image of your medical license to be verified. There are multiple reports of Japanese medical licenses being accepted⁸. You register your clinical department and affiliated institution, and once approved, you can begin in earnest. A guest mode that lets you try the feel of the product before approval is also said to be available, along with iOS and Android apps. This is why the practice of referencing it on a smartphone between outpatient cases is spreading.

The line that use is limited to healthcare professionals is worth keeping in mind. It is not designed for members of the public to enter their own symptoms and use it for self-diagnosis. This is the same as clinical decision support services such as UpToDate: it is positioned strictly as a tool that experts use as material for their judgment.

Japanese-language input is also reported to have been supported since the start of 2026. The behavior is that when you write your question in Japanese, it reads through the English-language papers and returns an answer in Japanese. The links to the cited papers remain in English, so digging deeper still requires the ability to read English, but the language barrier at the summary stage has dropped considerably. In a hands-on test that GIGAZINE ran on June 1, 2026, when chief complaints such as chronic fatigue and joint pain were entered, candidates for diseases to suspect, the underlying paper for each, and recommended tests came back with citations attached⁹. The typical shape of an answer is that you throw in a chief complaint and get, in one continuous flow, candidate differential diagnoses, the evidence, and the next tests. On the premise that a physician is organizing the information, I see it as having reached a level where it functions well enough even in a Japanese-language environment.

To understand its positioning, it helps to also grasp how it differs from adjacent products. The field called medical AI has expanded rapidly as of 2026, and alongside OpenEvidence you find medical-database types such as UpToDate and DynaMed, tools strong at summarizing papers such as Consensus and Perplexity, Doximity GPT born out of a physician community, and Glass Health leaning toward clinical reasoning. On top of that, a picture is emerging in which Anthropic's Claude for Healthcare and OpenAI's healthcare-oriented products collide directly as physician-facing AI⁴. What sets OpenEvidence's outline apart within this crowd is the combination of physician verification, mandatory presentation of sources, and official partnerships with societies and publishers. General-purpose LLMs are designed such that, in exchange for being able to answer any question, they cannot take full responsibility in the medical domain, and their terms of use carry cautions about medical use. OpenEvidence, conversely, drew its line by answering only medical questions, with physician verification and required presentation of sources. The long-established UpToDate is a clinical decision support service provided by Wolters Kluwer, where a team of specialist physicians manually peer-reviews while writing and updating articles. It is highly reliable, but it takes many steps to reach a conclusion, and it is not suited to searching through natural conversation.

Comparison axis	OpenEvidence	UpToDate
Format	Agentic (conversational)	Database (article browsing)
Information sources	Peer-reviewed papers and guidelines	Articles written by specialists
Update method	Reflected continuously	Updated per article, as needed
Price (for physicians)	Free	Paid subscription
Search experience	Ask and get an immediate answer with sources	Read an article from a keyword
Strength	Responsiveness and coverage	The reliability of expert editing

In practice, the two are used together with a clear division of labor: OpenEvidence for the moment when you want to reach a conclusion within the three minutes of an outpatient visit, and UpToDate for the moment when you want to settle in and read across an entire field. OpenEvidence has not driven out the existing tools. This form, in which physicians use the tool while checking the sources, is also the ideal form for when a company has an AI handle its internal documents. When we designed our enterprise AI ZEROCK, this is where we set our starting point.

Safety Considerations to Keep in Mind When Using It in Japan

Having read this far, you might feel that "if you can use it in Japan, you should." The number of Japanese physicians using it in outpatient care is indeed rising, but I believe the cautions for using it safely cannot be ignored.

The first thing to sort out is the handling of patient information. OpenEvidence is reported to have obtained U.S. HIPAA (Health Insurance Portability and Accountability Act) compliance in April 2025¹. In the sense that it is operated within the framework of U.S. patient-information protection law, that is progress. But that is a U.S. system, and whether a physician in Japan may enter patient information is a separate matter. In Japan, the Act on the Protection of Personal Information, each medical institution's information-management policies, and the physician's duty of confidentiality take precedence. Sending patient-identifiable information to an overseas provider's servers is often prohibited by institutional policy, and, more fundamentally, resolving a clinical question does not require any information that can identify an individual. The safe practice is to keep to an operation that never enters patient-identifiable information. It is wise to understand U.S. HIPAA compliance and the operational cautions in Japan as two separate things.

The next big one is the gap that exists between the latest U.S. treatments and Japan's insurance system. Because OpenEvidence reflects the latest U.S. evidence, it is not unusual for a drug or treatment recommended as standard of care in the U.S. to be unapproved or outside insurance coverage in Japan. For obesity medications, some cancer immunotherapies, and gene therapies for rare diseases, a gap of several years arises between approval in the U.S. and inclusion in Japanese national health insurance. If you present an OpenEvidence answer to a patient as-is as the latest standard of care, situations arise where it cannot be used under insured care, becomes self-pay private practice, or simply cannot be obtained domestically in the first place. The premise has to be an operation in which you critically appraise the presented evidence and translate it against Japanese guidelines and reimbursement rules.

The differences from domestic society guidelines are another point to be conscious of. Japan's various guidelines for cancer treatment, cardiology, diabetes, and so on are built through their own processes of evidence evaluation and consensus. Even when evaluating the same clinical trials, there are not a few cases where the strength of a recommendation differs from that of Western societies. Because OpenEvidence references U.S. society guidelines preferentially, its answers can diverge from domestic recommendations. What matters here is not which one is correct, but the physician interpreting what evidence and background lie behind each recommendation. There is no problem as long as it is used in the position of an educational use or reference information, but I consider it dangerous to use it in place of learning the guidelines.

Let me also organize the regulatory positioning. In Japan, when AI is used as a medical device, it is subject to the SaMD (Software as a Medical Device) regulation under the Pharmaceuticals and Medical Devices Act, and the boundary between products that have obtained approval from the PMDA (Pharmaceuticals and Medical Devices Agency) and those that have not is strictly managed¹⁰. OpenEvidence does not currently hold Japanese SaMD approval, and it is reasonable to position it strictly as an educational information tool for healthcare professionals. In the U.S., there is a reported move in which EchoNext, an AI developed by Pathway Labs that detects six types of structural heart disease from electrocardiogram images, obtained clearance from the FDA (Food and Drug Administration) and would be integrated into OpenEvidence¹¹. Not confusing the trend of AI diagnostic aids themselves beginning to obtain approval with a literature-referencing educational tool is the dividing line for using it safely in Japan. As useful domestic guidance, there are the guidelines on the use of medical digital data for AI research and development published by the Ministry of Health, Labour and Welfare in September 2024¹², and the second edition of the generative AI use guidelines in the medical and healthcare field compiled by HAIP/CIP¹³. The principle that the final clinical judgment is made under the physician's responsibility is also stated repeatedly in OpenEvidence's terms of use, and putting in place internal rules that do not deviate from use as an assistive tool is essential.

A "Literature AI" Idea That Can Be Applied to Enterprise Knowledge-Management AI

Finally, let me step into the business-side question of how to translate this into enterprise AI use. If you decompose the reasons OpenEvidence penetrated the clinical setting, they converge on four principles.

1. It narrowed the scope of reference from the general web to peer-reviewed medical literature.
2. It always attached sources to its answers, so users can verify the originals.
3. It only let users access the deeper features after verifying that they were physicians, controlling risk.
4. It officially partnered with the canon of the industry, such as NEJM and JAMA, giving its correctness social backing.

I believe these four can be reproduced as-is in specialized domains other than medicine. Demand is rising for AI specialized in the corpus of a professional domain, in forms such as an AI for law firms specialized in case law and statutes, an AI for manufacturers specialized in technical literature and patents, and an AI for financial institutions specialized in regulatory documents. The consultations I repeatedly receive in the field also boil down, when you get to the root, to three. The first is that pasting internal documents into ChatGPT or Claude, even when they are contractually excluded from training, cannot fully guarantee zero leakage, and does not square with non-disclosure agreements and information-management policies. The second is that general-purpose LLMs do not give answers grounded in a company's own internal documents, and the brute-force approach of pasting them into the prompt every time produces neither reproducibility nor scale. The third is that because answers come without sources, by the time it comes to using them for board-level decisions or legal judgments, there is no means of verification, which is almost certainly a blocker. As AI consulting that walks alongside you from framing exactly these issues to shaping a rollout plan, we offer WARP; because the generative AI landscape shifts month by month, it runs on a monthly-updated cadence.

OpenEvidence solved these three in the medical domain through domain specialization, citations, and verification. The product that brought this same philosophy into enterprise confidential-document management is ZEROCK, the enterprise AI agent we at TIMEWELL provide. ZEROCK adopts a tenant-isolated design that takes only a client company's internal documents as its reference scope, and automatically attaches a link to the source file and the relevant section to every answer. It is equipped with SSO (a mechanism that lets you use multiple systems with a single login), permission management, and role-based access control via audit logs; it runs on AWS servers within Japan; and it protects input data both contractually and technically by keeping it out of LLM training. It is a configuration that can clear procurement requirements even in industries such as healthcare, finance, and the public sector, where keeping data within the country is mandatory. Because it uses GraphRAG, a search method that traces relationships, it can answer questions that take into account connections between documents, such as "the renewal conditions of that contract" or "the amendment history of that policy," not just keyword matching. I have laid out the governance topics of how to design audit logs and permission management in detail in a separate piece, Enterprise AI Audit Controls (ISO 42001/SOC 2), so reading it alongside this should make the overall picture easier to grasp.

Just as OpenEvidence created a new layer at the front door of U.S. medicine, I expect that a layer of AI agents specialized in a company's own knowledge will be implemented as standard in enterprise operations as well. The idea is not a binary choice between banning and allowing general-purpose LLMs, but arranging specialized-domain AI as a separate layer. If you are unsure how to design your own literature AI, book a consultation and we can organize a concrete picture of how to use it, together.

Summary

OpenEvidence, as a separate line of evolution from general-purpose LLMs, has socially implemented an AI that combines specialization in a domain corpus, cited answers, and physician verification in the medical field. From its 2021 founding through its full-scale rollout in 2023, by the first half of 2026 its position had advanced significantly in a short span: reaching a 12 billion dollar valuation, posting full-year 2025 revenue of roughly 150 million dollars (up 1,803% year over year), integrating into the EHR at Mount Sinai, and becoming agentic via DeepConsult. It is becoming open to Japanese physicians as well, but it requires keeping in mind the safety topics: the handling of patient information and the positioning of HIPAA, the gap between the latest U.S. treatments and Japanese insurance inclusion, the differences from domestic society guidelines, and the relationship to SaMD regulation.

The lesson that can be extracted from this is not limited to medicine. The four principles of specialization in a domain corpus, cited answers, risk control through user verification, and partnership with the canon of the industry can be transferred as-is to enterprise AI for confidential documents. Enterprise AI leads who are worn down by the general-purpose-LLM debate would do well to observe how OpenEvidence works once and translate it into their own domain. At TIMEWELL we have productized this philosophy in the form of ZEROCK, so please reach out to us as an entry point.

Related Articles

• ZEROCK: An Enterprise AI Agent for Confidential Document Knowledge Management
• Enterprise AI Agents Compared: Selection Criteria and Major Domestic Products
• Enterprise AI Glossary: Key Concepts You Should Know

References

Footnotes

1. OpenEvidence Business Breakdown & Founding Story (incorporated September 2021, founding background, HIPAA compliance, 35 million papers, ad model, competitors) — Contrary Research — accessed 2026-07 — https://research.contrary.com/company/openevidence ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

2. OpenEvidence (founding year, registered physician count, funding history, USMLE, partnerships, DeepConsult) — Wikipedia — accessed 2026-07 — https://en.wikipedia.org/wiki/OpenEvidence ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. OpenEvidence revenue, valuation & growth metrics — Sacra — updated 2026-05-10 — https://sacra.com/c/openevidence/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

4. OpenEvidence hits $12B valuation with new round led by Thrive, DST — TechCrunch — 2026-01-21 — https://techcrunch.com/2026/01/21/openevidence-hits-12b-valuation-with-new-round-led-by-thrive-dst/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

5. OpenEvidence announces $210 million round at $3.5 billion valuation (Series B, DeepConsult announcement) — PR Newswire — 2025-07 — https://www.prnewswire.com/news-releases/openevidence-the-fastest-growing-application-for-physicians-in-history-announces-210-million-round-at-3-5-billion-valuation-302505806.html ↩ ↩²

6. DeepConsult user guide (agentic mode, compute cost) — OpenEvidence — accessed 2026-07 — https://www.openevidence.com/user-guide/deep-consult ↩

7. Mount Sinai inks 1st enterprise deal with OpenEvidence for Epic EHR integration — HIT Consultant — 2026-04-01 — https://hitconsultant.net/2026/04/01/mount-sinai-openevidence-ai-epic-integration-nurses-pharmacists/ ↩

8. What is OpenEvidence? The medical AI used by 40% of U.S. physicians (report on use from Japan) — decades.co.jp — 2026-02 — https://decades.co.jp/openevidence-clinical-guide-202602/ ↩ ↩²

9. We asked OpenEvidence about long-standing symptoms — GIGAZINE — 2026-06-01 — https://gigazine.net/news/20260601-openevidence-medical-ai-review/ ↩

10. Medical device programs — PMDA (Pharmaceuticals and Medical Devices Agency) — accessed 2026-07 — https://www.mhlw.go.jp/stf/seisakunitsuite/bunya/0000179749_00004.html ↩

11. Pathway Labs' EchoNext AI tool for heart disease detection clears FDA — STAT News — 2026-06-23 — https://www.statnews.com/2026/06/23/pathway-labs-echonext-ai-tool-heart-disease-detection/ ↩

12. Guidelines on the use of medical digital data for AI research and development — Ministry of Health, Labour and Welfare — 2024-09 — https://www.mhlw.go.jp/content/001310044.pdf ↩

13. Generative AI Use Guidelines in the Medical and Healthcare Field, Version 2 — HAIP/CIP — 2024-10 — https://haip-cip.org/assets/documents/nr_20241002_02.pdf ↩