A dark grid of evaluation tiles with green and blue highlights, representing medical AI benchmark results.

Model Evaluation · · 8 min read

How to read medical AI benchmark claims

Acumen Editorial · Updated November 20, 2026

Medical AI vendors now lead with benchmark scores. On November 4, Medscape reported that its Clarus model scored 88.6% on the NOHARM v2 safety benchmark and 85.0% on HealthBench Professional. It put the highest published competitor score on HealthBench Professional at 69.2%. In April, OpenAI launched ChatGPT for Clinicians with a HealthBench Professional score of 59.0, against 43.7 for physicians writing answers with unlimited time and web access.

Both numbers are real. They are also hard to compare, and neither tells a health-system buyer whether a tool is safe for their clinicians. If you lead clinical AI, evaluation or procurement, you need a way to read these claims before they reach a steering committee.

Here is the method we use.

Why are medical AI benchmark claims hard to compare?

Benchmarks change, and so do the conditions around them. A score depends on the benchmark version, the subset of cases, the system prompt, retrieval, the number of attempts and the grader. Two vendors can quote the same benchmark name and run different tests.

Medscape's release, for example, combines automated AI assessments with clinician review under its own ClinVal framework of 16 criteria, and it does not state the number of test questions. That is not unusual. Most press releases do not include enough detail to reproduce the result. Treat the number as a starting point for questions, not as a finding.

Who built the benchmark, and who ran the test?

Start with authorship. A benchmark built by the vendor, run by the vendor on the vendor's own product, is the weakest kind of evidence. It is not worthless. It needs outside confirmation.

The Decoder noted that OpenAI built HealthBench Professional and tested its own models on it, though the benchmark and dataset are openly available. It added that benchmark scores do not necessarily translate to real clinical practice.

Independent studies carry their own ties. A June 2026 preprint led by UCSF researchers compared OpenEvidence with three frontier models on 620 real physician queries. OpenEvidence ranked highest on all five axes. The authors disclose that OpenEvidence co-designed the data collection and ran the survey, while stating they are not affiliated with it. Disclosure like that is good practice. It is also something you should look for every time.

What does the score actually measure?

"Accuracy," "quality" and "safety" are different measurements. Ask which one the headline number is.

NOHARM, from Stanford and Harvard researchers, measures harm. It uses 1,100 cases drawn from real physician-to-specialist consultations, scored by 29 board-certified physicians. Becker's reported that potentially severe harmful errors ranged from 2.9% for the best medical tool to 24.6% for the worst generalist model. Errors of omission, recommending too little, made up more than 80% of severe errors. A hypothetical model that recommended nothing scored worst, at 37%.

That finding matters for buyers. A tool can sound careful and still be unsafe, because the dangerous failure is often what it leaves out. Ask whether the score penalizes omissions, and ask for the severe-error rate, not only the average.

Averages hide the tail. A 2025 literature survey on HealthBench found that strong average scores can mask dangerous weaknesses in high-stakes situations. In healthcare, the worst case is the one that reaches a patient.

Who graded the answers: physicians or an LLM judge?

Grading at scale often uses an LLM judge. That is reasonable for well-specified criteria. It is risky when the judge is related to a system being judged.

The UCSF-led preprint ran three frontier models as judges on the same pairs that 149 physicians rated. The judges generally agreed with physicians on the top system but diverged on lower ranks. One model's judge showed self-preference, with a negative rank correlation against the physicians. The judges also produced more extreme win margins than the physicians did.

So ask who wrote the rubric, who applied it and how the automated grader was calibrated. Ask how many physicians reviewed each case and how often they agreed. HealthBench Professional uses rubrics adjudicated by three or more physicians, which is a useful bar to hold vendors to.

Do the cases look like your work?

A benchmark is a sample. The question is whether it samples your specialties, your patients and your tasks.

Stanford HAI's PhysicianBench shows how much the task shape matters. On 100 multi-step clinical tasks in a simulated EHR, the best agent completed 46% on one attempt and 28% across all three runs. A strong score on single-turn questions says little about a tool that will place orders or draft prior authorizations.

Check the specialty mix, the share of hard cases and whether cases came from real practice. OpenAI says about a third of HealthBench Professional examples come from red teaming and that the hardest conversations are overrepresented 3.5 times. That design choice changes what the score means.

How to read a benchmark claim: a quick reference

Claim you will seeWhat to askRed flag
"Top score on HealthBench"Which version, which subset, how many cases, how many runs?No version or case count
"Beats physicians"Under what conditions did physicians answer, and who graded both?Physicians graded on a different setup
"Safest model"What is the severe-error rate, and are omissions counted?Only an average safety score
"Clinician-validated"How many clinicians, which specialties, what agreement rate?"Clinicians" with no number
"Independent study"Who funded it, designed the data and ran it?No conflict disclosure
"Graded by AI"Which judge model, calibrated against whom?Judge from the same vendor

Checklist: what to ask a vendor about its benchmark scores

  • Which benchmark, which version and which subset produced this number?
  • Who built the benchmark, and who ran this test?
  • How many cases, and how many runs per case?
  • What prompt, retrieval and tools were active during the test?
  • Does the score count harm and omissions, or only accuracy?
  • What is the severe-error rate, and can we see the worst failures?
  • Who graded the answers, and how was any LLM judge calibrated?
  • How many physicians reviewed each case, and what was their agreement?
  • Do the cases cover our specialties and our tasks?
  • Will you run our private, held-out cases and share the results?

The last question is the one that settles it. A public score tells you how a tool compares. Your own cases tell you whether it works for you.

How Acumen helps

We build the private evaluation sets that turn a vendor claim into a decision. Practicing physicians, matched by specialty, write cases from real practice and the rubrics that grade them, including criteria for omissions and severity. Specialists grade the outputs, a third reviewer adjudicates disagreements, and we report agreement rates. Then we run every shortlisted tool on the same frozen set, so you compare like with like.

We work in healthcare, life sciences and pharma only. The physicians grading your shortlist know the specialty, and the documentation holds up in procurement and regulatory review.

See how it works on our Medical Model Evaluation page. For the full evaluation method, read How to evaluate a clinical LLM.

Request a sample dataset Schedule a conversation

FAQ

How do you read a medical AI benchmark claim?

Ask four questions. Who built and ran the test? What does the score measure: accuracy, quality or harm? Who graded the answers, physicians or an LLM judge? Do the cases match your specialties and tasks? Then ask for the benchmark version, the number of cases and runs, and the worst failures. Without those details, a headline score is a claim, not evidence.

What is HealthBench Professional?

HealthBench Professional is an OpenAI benchmark that evaluates models on real clinician tasks: care consults, writing and documentation, and medical research. Rubrics are adjudicated by three or more physicians. OpenAI built it and tests its own models on it, though the benchmark and dataset are openly available. Scores depend on the version and test setup, so compare them carefully.

What is the NOHARM benchmark?

NOHARM is a medical safety benchmark from Stanford and Harvard researchers. It uses 1,100 cases from real physician-to-specialist consultations, scored by 29 board-certified physicians, and measures potentially severe harmful errors. Becker's reported severe-error rates from 2.9% to 24.6% across 24 systems, with errors of omission making up more than 80% of severe errors.

Can an LLM judge grade medical AI fairly?

Sometimes, for well-specified criteria, once it is calibrated against physician grades. A June 2026 UCSF-led preprint found LLM judges agreed with physicians on the top system but diverged on lower ranks, and one judge favored its own model. Use physicians to write rubrics and review high-risk cases, and ask vendors how any automated grader was calibrated.

Do vendor benchmark scores predict real clinical performance?

Not on their own. Benchmarks sample cases that may not match your specialties, patients or workflows. Stanford's PhysicianBench found the best agent completed 46% of multi-step EHR tasks on one attempt and 28% across three runs. Treat public scores as a shortlist filter, then test each tool on your own private, held-out cases.

What should a health system ask a medical AI vendor about benchmarks?

Ask which benchmark version and subset produced the number, who built and ran it, how many cases and runs were used, and what prompts and retrieval were active. Ask for the severe-error rate, the worst failures, the grading method and physician agreement. Finally, ask the vendor to run your own held-out cases and share the results.

From the Acumen archive. Explore expert data and Healthcare & Pharma.

← Back to Insights