Questions
Short answers about the 14 boards and how to read them.
- Which model is best for clinical work?
- It depends on the job. Pick the board that matches it, such as documentation, coding or EHR agents, and read that board with the configuration line under each row. The index gives one ordering across boards, as this site's own summary, and the compare page sets two to four models side by side.
- How is the index different from a board?
- A board is one benchmark's own results, ranked on its own scale. The index is this site's calculation across boards: each result is read on its board's 0 to 100 scale, adjusted for how easy or hard that board runs for the same models, and averaged over the boards a model has, with boards near their ceiling counting for less. It is a summary; the boards are the evidence. How it is computed.
- Why does no model score near 100 on the index?
- Because no model scores near 100 on most clinical boards. Each result counts at its real value on its board's scale, not stretched so the board's leader reaches 100, and the boards with the most headroom carry the most weight. Right now Bedside Bench, RCM-Bench, MedScribe (Vals AI) and MedXpertQA (MM) are near the ceiling.
- How does the index treat a model with results on one or two boards?
- Every model with a result on an index board is ranked, and it is not marked down for boards it has no result on. A score from few boards rests on less evidence, so it is pulled a little toward the field average of 44.2. The field average's share is one board 60%, two 43%, three 33%, four 27%, five 23%, six 20%. Kimi K3.1 Pro has one board, so its adjusted result of 64.8 counts for 40% and the field average for 60%: its index is 52.4. The pips under each entry show which index boards a score rests on, and the index views can be filtered to models with results on 2, 3 or 5 or more boards.
- What do the bar colours mean?
- The colour is the lab that built the model, and the lab's mark sits beside the name. Every bar is drawn the same way; who ran a result (the benchmark's leaderboard, an independent group or the model's own lab) is written beside it.
- Why does the same model appear twice on one board?
- Because it was run twice under different settings, for example a different reasoning effort, harness or grader. Each row keeps its configuration, and neither is dropped.
- Why is a model's rank here different from the official board?
- This site may carry fewer rows than the official board. Where the full field size is known, the board page says how many of those models appear here.
- Can I use the data?
- Yes. The compilation is CC BY 4.0: credit Clinical Benchmarks and link back. Every result is downloadable as CSV or JSON. The index is at /data/index.csv and /api/v1/index.json, with each model's board results, adjusted results and field share; its version field says which method made it.
- How do I report a wrong number?
- Use the contact link and name the document and page that shows the right value. Corrections go through the same check as any new row.
- Where are the consumer health benchmarks?
- Boards about AI people use for their own health are on the sister site linked at the foot of every page.