OpenEvidence vs UpToDate Expert AI vs ClinicalKey AI vs DynaMed: The Free One Has a BAA. The Paid One Says No PHI.

In its CY 2027 fee schedule final rule, CMS called OpenEvidence "likely" the most widely used AI tool in primary care, used by more than 40% of U.S. physicians according to the outside reports CMS cited. (We covered the rule here.) The established reference publishers have since added their own AI answer engines. Wolters Kluwer has UpToDate Expert AI, Elsevier has ClinicalKey AI, and EBSCO has Dyna AI inside DynaMed.

If you run a small practice, you probably already use at least one of these. Three questions matter more than any marketing comparison. What does the answer come from? What does it cost you? And are you allowed to type patient details into it? The answers are not what you'd guess. The free tool is the one whose terms include a business associate agreement (BAA). The best-known paid tool says it is not designed for protected health information (PHI) at all.

We read each vendor's own terms, product pages and press releases, and the independent research that names these tools, in November 2026. This post covers what we found.

Quick verdict

• OpenEvidence is the default for a reason. It's free for verified U.S. clinicians, answers in seconds from a wide licensed literature base, and its Terms of Use include a BAA. The costs are advertising, a $100 liability cap, and a business model that depends on your attention.
• UpToDate Expert AI is the most conservative tool. It answers only from UpToDate's own edited content and declines more often than the others. Its disclaimer says it is not intended for PHI. It makes sense if you already pay for UpToDate or your hospital provides it.
• ClinicalKey AI is Elsevier's literature-grounded tool, now with full text from journals such as NEJM and The Lancet. Individual monthly and annual plans exist. It publishes the most detailed public model card of the four.
• DynaMed with Dyna AI is the cheapest paid route to curated, point-of-care content with AI search, starting at $149 a year for students. It has the least independent evaluation of the four.
• None of them has been shown to be safe to use without checking the source. In the only independent head-to-head study, both specialized tools scored below general-purpose chatbots on every benchmark. Neither the study nor the vendors' rebuttals settles the question.

The four tools at a glance

• OpenEvidence (OpenEvidence Inc.)
  • Price: free for verified U.S. healthcare professionals, paid for by advertising
  • Answers from: broad medical literature, plus content agreements the site lists with NEJM, JAMA, Nature, NCCN and Cochrane
  • PHI: allowed for U.S. covered entities under a BAA built into the Terms of Use
  • Extras: Visits, a free ambient scribe; a phone dialer; CME
• UpToDate Expert AI (Wolters Kluwer)
  • Price: included in UpToDate subscriptions. Independent price trackers list U.S. individual plans at $579 a year (standard), $699 (Pro Plus) and $219 (trainee). Confirm at checkout.
  • Answers from: UpToDate content only
  • PHI: not intended for personal data, and "not designed to protect such information"
  • Extras: CME credit earned inside the workflow (added March 2026)
• ClinicalKey AI (Elsevier)
  • Price: individual monthly or annual subscription in the U.S. since October 2024. The pages we checked don't list a price. Medical students in the U.S. and Canada get one year free.
  • Answers from: Elsevier's clinical overviews, drug monographs and journal content, plus full text from more than 130 premium journals and partner societies
  • PHI: "built to be HIPAA-compliant." It has no access to patient records.
  • Extras: SMART on FHIR EHR launch (Epic and Oracle Health), and a public CHAI model card
• DynaMed / DynaMedex with Dyna AI (EBSCO)
  • Price: individual DynaMedex plans are priced by clinical role, from $149 a year for students. Free trial available.
  • Answers from: DynaMed's curated content, plus Micromedex drug content in DynaMedex
  • PHI: we found no public statement for individual subscribers
  • Extras: Dyna AI Mode interface (February 2026), and CME/MOC tracking

What the independent evidence actually says

The Nature Medicine study

The only peer-reviewed head-to-head study of these tools is Vishwanath et al., "General-purpose large language models outperform specialized clinical AI tools on medical benchmarks," from the NYU Langone group led by Eric Oermann. It was published online in Nature Medicine on June 12, 2026 (doi 10.1038/s41591-026-04431-5). It tested OpenEvidence and UpToDate Expert AI against GPT-5.2, Gemini 3.1 Pro and Claude Opus 4.6. ClinicalKey AI and DynaMed were not tested.

• MedQA, 500 exam-style questions: Gemini 97.4%, GPT-5.2 94.2%, Claude 90.2%, OpenEvidence 89.6%, UpToDate Expert AI 88.4%.
• HealthBench, 500 items scored for alignment with physician judgment: GPT-5.2 88.0, Gemini 79.3, Claude 77.0, OpenEvidence 62.6, UpToDate Expert AI 61.3.
• 100 real clinical questions that physicians had asked a chatbot in practice, rated blind by 12 U.S. clinicians on a 1–4 scale: Gemini 3.62, GPT-5.2 3.54, Claude 3.52, Google's AI Overview 3.27, OpenEvidence 3.24, UpToDate Expert AI 3.17. The two clinical tools scored about the same as a Google search summary.
• Refusals: UpToDate Expert AI declined 19% of the real clinical questions. Google's AI Overview declined 6%, and the other systems 1–3%. OpenEvidence had the lowest clarity score, and the authors attributed that to how it communicates rather than what it knows.

The authors list real limitations. The clinical tools have no public API, so they were queried by hand in a browser. OpenEvidence was accessed in September 2025 and February 2026, and UpToDate in November 2025 and February 2026. Public benchmarks such as MedQA may have leaked into the general models' training data, and OpenAI built HealthBench. Agreement between reviewers on individual items was low. And the study did not assess citation quality or speed, which are probably the two things clinicians value most in these tools. The authors call it "a snapshot of a rapidly evolving landscape."

The vendors' responses

Both vendors objected. OpenEvidence asked the journal to retract the paper, arguing that its methods were flawed and the benchmarks contaminated. Nature Medicine referred the company to its standard Matters Arising process for published rebuttals. We found no report of a retraction as of November 2026.

Wolters Kluwer's chief medical officer, Peter Bonis, also disputed the results. The company says its own testing covered more than 1,600 clinical queries scored against more than 15,000 criteria, and that UpToDate Expert AI gave clinically aligned information on 99.9% of them. It also argues that declining to answer is sometimes the right response, and that the paper never checked whether the refusals were appropriate. That's a fair point. A tool that says "UpToDate doesn't cover this" is behaving differently from one that invents an answer. But the 99.9% figure is the vendor's own internal rubric, and no one has independently replicated it.

The counter-study

On June 27, 2026, a team led by Jean Feng (UCSF) posted a preprint to arXiv (2606.28960; not yet peer reviewed). In it, 149 physicians in 36 states compared answers blind and side by side on 620 real questions, plus 187 HealthBench items. OpenEvidence beat Claude Opus 4.8, Gemini 3.1 Pro and GPT-5.5 on all five dimensions: accuracy, clinical utility, source quality, verifiability and completeness. Its win margins ranged from 25 to 39 percentage points.

Read its conflict-of-interest section before you quote it. The questions were de-identified and rewritten from OpenEvidence's own user queries. OpenEvidence co-developed the data collection plan, ran the data collection, and paid the survey respondents. None of the authors is affiliated with OpenEvidence, the analysis plan was pre-specified, and the data is public. But it measures relative preference between answers to questions asked on OpenEvidence's own platform. It also did not include UpToDate, ClinicalKey AI or DynaMed.

What to take from the evidence

There is no independent evidence that any of these four tools beats the others on the questions you actually ask. The numbers that do exist point in opposite directions, depending on who designed the test. That's not a reason to stop using them. It is a reason to keep checking the cited source, the same way you'd check any AI-generated note.

If your conclusion is "just use ChatGPT," PHI is the catch. Consumer chatbots generally don't sign BAAs, and enterprise versions sign them only under a negotiated contract. See what HIPAA compliance actually means for AI tools.

OpenEvidence: free, fast, and paid for by advertising

OpenEvidence is a search-and-answer engine for verified U.S. healthcare professionals. It is not intended for patients, and its Terms say so. It cites its sources inline, and its site lists content agreements with NEJM, JAMA, Nature, NCCN and Cochrane.

The legal documents are better than many assume. We read the Terms of Use and Privacy Policy, both dated September 17, 2026, and the BAA, dated February 5, 2025.

• A BAA comes with the Terms. The BAA is "incorporated by reference into and made a part of" the Terms of Use. It applies when the user is a HIPAA covered entity and OpenEvidence handles PHI on its behalf. You don't sign anything separately. If your health system has its own access agreement with OpenEvidence, that agreement controls instead.
• No PHI for AI training. The Terms say: "we do not use Protected Health Information (as defined by HIPAA) for any AI training purposes." The BAA does let OpenEvidence de-identify PHI under HIPAA's standards and use the de-identified data "to provide or improve its products or services." That's a common clause, but it is broader than "never used."
• Advertising. OpenEvidence shows ads from "closely vetted advertisers." It says the ad system and the answer system are "fully unconnected" and that advertisers "cannot and will never be able to influence the content." The advertising policy is dated August 12, 2024. The Privacy Policy allows interest-based advertising through cookies and pixels, with an opt-out. It also says OpenEvidence doesn't do cross-context behavioral advertising "based on user queries, uploads or chats, or PHI."
• Liability and disputes. Total liability is capped at $100. Disputes go to individual arbitration. You can opt out by writing to the company within 30 days of first accepting the arbitration terms.
• Your inputs are your responsibility. If you enter patient information, the Terms make you confirm you have the necessary consents and authorizations, including any your employer requires. If you are employed, your organization's AI policy controls. The BAA covers OpenEvidence's obligations to a covered entity. It does not give you permission to use the tool under your employer's policy.

Visits is a free scribe. In August 2025 OpenEvidence launched Visits. It transcribes encounters, drafts notes with evidence added to the assessment and plan, and lets you query uploaded patient documents. It is "free for all verified US Healthcare Professionals." The Terms say Visits records your conversation, and that you must tell the other people in the room where the law or your organization requires it. Our consent and recording-law guide covers which states require that. A free scribe built into an evidence tool is a real alternative to the free scribes we've compared. Judge it on the same terms, including where the audio goes.

Best for: fast questions answered from current literature, especially by clinicians who don't have institutional UpToDate access. Watch for: ads next to clinical answers, an evidence base that is mostly vendor-run, and the $100 cap.

UpToDate Expert AI: the curated answer, with no PHI

Wolters Kluwer announced UpToDate Expert AI on September 24, 2025. By August 2026 it said about 2,500 U.S. hospitals and health systems, over 90% of its U.S. Enterprise customers, had chosen it. That's why many employed clinicians already have it without asking for it.

It works differently from OpenEvidence. Answers are summaries of UpToDate's own topic reviews, the content that UpToDate's physician editors write and update, rather than of the open literature. That's why it refuses more often. If UpToDate doesn't cover a question, Expert AI is built to say so rather than reach further.

The PHI rule is explicit. UpToDate's disclaimer, as quoted in Vanderbilt University Medical Center's January 2026 rollout notice, says Expert AI "is not intended for use with personal data" and "is not designed to protect such information." It adds that outputs are generated by AI, "not reviewed by a human, and may contain errors," and that the tool is for clinicians only. Don't paste notes, names, dates of birth or MRNs into it. Ask the question in general clinical terms.

For individuals: Wolters Kluwer's site blocked our automated checks, so we couldn't confirm U.S. individual prices on its own pages. Independent price trackers that checked in early November 2026 list $579 a year for standard, $699 for Pro Plus and $219 for trainees, and say Expert AI is now included in all three U.S. plans. Wolters Kluwer's own product pages also say U.S. trainee subscriptions now include it. A tracker writing in mid-2026 said it required Pro Plus, so the plans have changed recently. Check what's included at checkout before you pay for Pro Plus to get the AI. Since March 2026, clinicians can also earn CME credit for questions answered inside Expert AI.

Best for: clinicians who want answers limited to edited, graded recommendations, and anyone whose employer already licenses UpToDate. Watch for: the higher refusal rate, the no-PHI rule, and the price if you're paying yourself.

ClinicalKey AI: the literature tool with a public model card

Elsevier opened ClinicalKey AI to individual U.S. residents and physicians on monthly or annual subscriptions on October 8, 2024. On February 18, 2026 it added full text from premium journals and societies, including The Lancet, NEJM, the European Society for Medical Oncology and the American College of Cardiology. Elsevier says content is updated every 24 hours.

Two details stand out:

• It has a public model card. Elsevier lists ClinicalKey AI in the Coalition for Health AI (CHAI) registry. The entry gives a general-availability date of March 1, 2026, and describes how the system works: retrieval-augmented generation with citation validation and guardrails. It also spells out what the tool is not for. That list includes standalone diagnostic or treatment decisions, time-critical triage, and "queries requiring access to patient-specific EHR data," because the system has no access to patient records. Federal certification is dropping its own AI transparency requirements (see HTI-5), so a voluntary card like this is more useful than it was. Our model-card reading guide shows what to look for.
• It is "built to be HIPAA-compliant," with a footnote that this applies "where recognized by country regulators." Elsevier says user data is encrypted and that AI partners don't get access to it for training. We didn't find a public BAA for individual subscribers. Ask before you enter any patient details.

Best for: clinicians who want literature-grounded answers with full-text citations, and organizations that want EHR launch through SMART on FHIR (Epic and Oracle Health). Watch for: no published individual price on the pages we checked, and no independent comparison yet.

DynaMed with Dyna AI: the low-cost curated option

EBSCO launched Dyna AI in July 2024 and added it to mobile in October 2024. On April 1, 2025 it opened Dyna AI to individual DynaMedex subscribers who don't have an institutional subscription. In February 2026 it added Dyna AI Mode, an AI interface built into DynaMed, DynaMedex and Dynamic Health. EBSCO says the interface shows the question, the reasoning and the underlying evidence.

Like UpToDate, it answers from curated, editor-graded content. DynaMedex also includes Micromedex drug content and dosing tools, which is useful in primary care. Individual pricing depends on your clinical role. EBSCO publishes only the student price ($149 a year), and a free trial is available.

The weakness is evidence. We found no independent head-to-head study that includes Dyna AI, and EBSCO's pages that we reviewed don't say whether individual subscribers may enter PHI. Treat it as a no-PHI tool unless you have a BAA in writing.

Best for: price-conscious clinicians who want curated recommendations and drug dosing in one subscription. Watch for: the lack of outside testing.

Two more you'll run into

• Doximity Ask (formerly Doximity GPT, then DoxGPT) is free for verified U.S. physicians, NPs, PAs and medical students. It was built partly on Pathway, the clinical AI company Doximity acquired. Doximity markets it as HIPAA-compliant. The public pages we reviewed don't describe its BAA terms, so confirm them in the app before entering patient details. Doximity's free scribe is covered in our pricing guide.
• Glass Health combines differential-diagnosis and plan drafting with an ambient scribe. It is the only one of these tools in our directory, and its opt-in BAA and ad-supported free tier are worth reading first. See our Glass Health review.

How to choose

• You're a solo clinician with no institutional access and a tight budget: start with OpenEvidence. Read the ads as ads, and click through to the cited source on anything that will change management.
• Your hospital or system already licenses UpToDate: use Expert AI for questions where you want graded, edited recommendations. Use OpenEvidence for newer literature that UpToDate hasn't covered yet. Follow your employer's PHI rules in both.
• You want one paid subscription that covers curated content and drug dosing: price DynaMedex for your role against UpToDate. Note that UpToDate has the only independent outside testing of the two.
• You want literature-grounded answers and an EHR launch for a group: ask Elsevier for ClinicalKey AI pricing and its BAA, and read its CHAI model card before the demo.
• You want a scribe and an evidence tool in one place: compare OpenEvidence Visits and Glass on the same checklist you'd use for any AI scribe: BAA, audio retention, consent, and EHR transfer.

Using any of them safely

1. Ask without identifiers by default. "68-year-old with CKD stage 4 and new AFib: anticoagulant choice?" works as well as pasting the chart, and it keeps you inside every tool's rules. Use PHI only where a BAA covers you and your employer allows it.
2. Open the citation before acting on anything that changes management. None of the published studies checked whether these tools' citations actually support their claims.
3. Treat a refusal as information. If UpToDate Expert AI declines and OpenEvidence answers confidently, find out why.
4. Keep the clinical judgment yours, and on the record. These tools present themselves as reference tools for clinicians, which matters under FDA's clinical decision support guidance. Your note should show your reasoning, not "per AI."
5. If you'll attest to the new MIPS improvement activity for clinician use of AI (IA_AHW_4, starting with the 2027 performance year), check its documentation requirements first. See our fee schedule breakdown.

To compare evidence tools with the rest of the category, browse our clinical decision support directory, or see side-by-side picks for family medicine, nurse practitioners and small clinics. If you use these tools before visits, our pre-charting workflow shows where they fit.

---

This comparison is informational only and is not legal, medical, or procurement advice. We are not affiliated with OpenEvidence, Wolters Kluwer, Elsevier, EBSCO, Doximity, or Glass Health. Terms, prices, and features were checked on vendor websites and in published research in November 2026, and they change often. UpToDate individual prices come from independent trackers because the vendor's own pages blocked our checks. Confirm current terms, BAA coverage, and pricing with each vendor, and your organization's AI policy with your compliance team, before you use any of these tools with patient information.