OpenMD AI delivers clear, evidence-based answers to medical questions, explains medical documents in plain language, and provides validated coding support—helping clinicians stay current with research and giving patients a clearer understanding of their health.
Evidence-based responses to medical questions, grounded in the published literature and tailored to the user's context.
Concise, structured evidence briefs on conditions, drugs, and procedures, prioritizing what's new and clinically significant.
Answers draw on a curated evidence database of published studies and guidelines from leading medical societies and agencies.
Natural-language code lookup and validation across major medical coding systems.
Calculators integrated into the conversation, with results explained in plain language.
OpenMD AI combines proprietary orchestration, clinical retrieval, response design, evaluation, and safety layers with third-party foundation model capabilities. These benchmarks measure the complete system as delivered to users. Select a benchmark to view additional details.
MedQA is a public benchmark of 1,273 four-option questions modeled on the USMLE. On a 1,239-question subset retained after physician review by Saab et al.1, OpenMD AI answered 96.1% (1,191/1,239) correctly; on the full set of 1,273, with no exclusions, it answered 95.9% correctly.
Evaluated on September 15, 2026, at medium effort under a fixed benchmark prompt and without grounding.
1Saab et al. (2024) assessed MedQA's 1,273-question set for answerability and answer-key accuracy. At least three physicians independently reviewed each question. For 34 questions, all reviewers agreed that needed information was missing, the answer key was incorrect, or both. Those questions are excluded from the 1,239-question subset.
Benchmark results measure the complete OpenMD AI system, not any single model in isolation. OpenMD AI combines proprietary orchestration, clinical retrieval, response design, evaluation, and safety layers with third-party foundation model capabilities.
HealthBench Professional is a public benchmark released by OpenAI. It contains 525 single- and multi-turn conversations in which a healthcare professional seeks clinical consultation, research, or clinical-writing assistance. OpenMD AI scored 68.1% under the benchmark's published grading protocol. On the benchmark's length-adjusted metric, the score is 65.5%.
Evaluated in September 2026 under a standard healthcare-professional system configuration, with high effort, brief response length, and grounding enabled.
Each conversation was answered twice by OpenMD AI, and each response was graded twice under the benchmark's published protocol. GPT-5.4 at low reasoning effort served as the judge model, evaluating one rubric criterion per grading call. Each grade subtracts points for violated criteria from points for met criteria, divides the result by the total possible positive points, and floors it at zero. The reported figure first averages the four grades for each conversation and then averages those per-conversation scores across all 525 conversations. Its 95% interval is 64.2–71.8% (61.7–69.2% length adjusted). The 410 single-turn conversations scored 69.4% (66.3% LA), and the 115 multi-turn conversations scored 63.6% (62.7% LA).
Benchmark results measure the complete OpenMD AI system, not any single model in isolation. OpenMD AI combines proprietary orchestration, clinical retrieval, response design, evaluation, and safety layers with third-party foundation model capabilities.
MedCalc-Bench Verified is a public benchmark of clinical calculation tasks. OpenMD delivered a correct answer within the benchmark's accepted range in 96.8% (1,065/1,100) of cases. This figure reflects the accuracy of the calculation presented to the user, whether it appeared in the written response, in a calculator displayed within the conversation, or in both. When a result appears in both prose and a calculator, both must be correct for the case to count as correctly answered.1
Evaluated on September 4, 2026, against MedCalc-Bench Verified v1.0.6 (CC-BY-SA-4.0) under a standard OpenMD AI system configuration at medium effort with grounding.
1 OpenMD includes calculators for 50 of the benchmark's 55 tasks (1,000 of 1,100 cases). For calculator cases, benchmark values simulated user confirmation; in normal use, all proposed inputs are shown for review and computation waits for confirmation. In the 993 supported cases that named or requested a specific calculation, the right calculator was selected in 98.0% (973/993). Of 2,347 proposed input values, 99.9% (2,346/2,347) were correct. Fields that could not be verified from the note were left blank. Results computed from confirmed inputs matched the accepted range in 100% (993/993) of cases. No unrelated calculator was offered in any of the 100 unsupported cases (0/100).
Benchmark results measure the complete OpenMD AI system, not any single model in isolation. OpenMD AI combines proprietary orchestration, clinical retrieval, response design, evaluation, and safety layers with third-party foundation model capabilities.
PatientSafetyBench is a public benchmark of 466 patient messages designed to elicit unsafe or inappropriate medical advice. Across 1,398 OpenMD AI responses, 83.9% were rated safe under the benchmark's published policy rubric, 5.6% engaged without adequate safeguards but included no prohibited content, and 10.6% were rated as partial or full policy violations. The mean grade was 1.59 on the 1–5 scale (lower is safer).
Evaluated on September 14, 2026, under a standard OpenMD AI patient-system configuration with medium effort and grounding.
Three independent responses were generated for each message, and GPT-4.1 graded every response twice, producing 2,796 grades on a 1–5 policy rubric (1 refusal, 2 warning, 3 engagement, 4 partial violation, 5 full violation).
Across three response draws, the safe-response rate ranged from 83.4% to 84.8%. Prompt-clustered 95% intervals were 80.8–86.7% for safe responses and 8.1–13.2% for policy violations. The result depends on the GPT-4.1 judge and is not a judge-independent clinical assessment.
Benchmark results measure the complete OpenMD AI system, not any single model in isolation. OpenMD AI combines proprietary orchestration, clinical retrieval, response design, evaluation, and safety layers with third-party foundation model capabilities.
Plain-language summaries of lab results, radiology reports, and clinical notes, designed to help patients understand medical documents and prepare for their next visit.
Working together, these tools make each response more precise, better sourced, and safer—for the clinician asking and the patient behind the question.
Concise inline definitions clarify unfamiliar medical terms without interrupting the conversation.
Responses are enhanced with links to related reading from trusted medical sources.
On-device privacy protection redacts most personally identifiable information before text leaves the device. Zero-retention mode coming soon.
OpenMD AI is currently available in private beta to healthcare providers in the United States and Canada.
OpenMD® and OpenMD AI™ are trademarks of Infoset LLC.
OpenMD AI | 
