FAQ
What people ask about the healthcare benchmark landscape, answered from the index itself.
Which healthcare AI benchmarks are still being updated with frontier models?
18 as of August 16, 2026: HealthBench Professional, HealthBench Hard, HealthBench, Health Optimization Bench, MAST (Medical AI Superintelligence Test), MedHELM, First, Do NOHARM (v2), HealthAgentBench, CHI-Bench, MedCode (Vals AI), MedScribe (Vals AI), MedXpertQA (MM), Artificial Analysis Healthcare & Medical Index, PhysicianBench, EHR-Complex, WHBench, HealthAdminBench, OpenAI Dynamic Mental Health Evaluations. Each qualifies by having at least one public result for a current frontier model, and each has its own page here with the current numbers and their source.
Which AI model is best for healthcare overall?
No single number answers that, because the benchmarks measure different work on incompatible scales. What the index can say: Claude Opus 5 currently leads HealthBench and First, Do NOHARM (v2) and MedCode (Vals AI) and MedScribe (Vals AI); Claude Fable 5 currently leads HealthBench Professional and Health Optimization Bench; Gemini 3.1 Pro currently leads MedXpertQA (MM); GPT-5.6 Sol currently leads MAST (Medical AI Superintelligence Test); Claude Opus 4.6 currently leads WHBench; GPT-5.5 currently leads PhysicianBench; Muse Spark currently leads HealthBench Hard; Gemini 3.1 Pro (Preview) currently leads MedHELM; Claude Code (Opus 5) currently leads HealthAgentBench; erius + claude-opus-5 currently leads CHI-Bench; Claude Opus 5 (Adaptive Reasoning, Max Effort) currently leads Artificial Analysis Healthcare & Medical Index; GPT-5.4 (high reasoning) currently leads EHR-Complex; Claude Opus 4.6 (computer-use agent) currently leads HealthAdminBench; GPT-5.5 Instant (June Update) currently leads OpenAI Dynamic Mental Health Evaluations. The models page shows every model's full footprint.
Why not combine everything into one healthcare ranking?
Because the underlying measurements do not add: a rubric score, a hard-subset score, and an exam accuracy are different quantities, produced by different graders. Averaging them would manufacture precision that does not exist, so this site keeps one table per benchmark and one page per model instead.
Where do the scores come from?
Every row names its source: the sister leaderboard sites this index's team also runs, benchmark publishers and vendor system cards, and third-party trackers. Vendor-reported numbers are labeled as such, and each score appears exactly as its source shows it.
How current is the index?
The snapshot on every page is from August 16, 2026, and each result row carries its own date as well, since sources refresh on their own schedules. Changes land as dated entries on the updates page.
Is the data downloadable?
Yes, the full index ships as JSON and CSV under CC BY 4.0 from fixed paths, with citation guidance on the data page. Individual scores should be cited to the original source each row names.
The index holds the current tables, the models page the per-model views, and the methodology page the sourcing rules.