How the numbers get here
Every score was printed somewhere first, and you can follow it back to that page.
On this page
Where results come from
- The model's maker: system cards, model cards, launch posts; 67 results, marked in words beside each score.
- The benchmark's maintainers: official leaderboards and papers, usually with every model run the same way. Solid bars; 117 results.
- Independent groups: researchers who neither make the model nor maintain the benchmark. Solid bars; 9 results.
A maker chooses its own settings, and sometimes its own grader, so the board filter lets you hide makers' numbers and compare the rest on their own.
How each number is checked
For each result we keep the document, the sentence or table cell carrying the number, and where it sits. Select any bar to read it. A result is verified once someone has read the number in the document itself. Unchecked rows are marked “unverified”, rows whose sources disagree “disputed”, and rows without a document “source pending”.
When sources disagree
The original document wins: a system card outranks an article repeating it, an official leaderboard outranks a screenshot of it. When a maker and a benchmark's maintainers publish different settings for one model, both rows stay, each with its setting shown.
How a board is ranked
By the benchmark's own direction, with ties sharing a rank (1, 2, 2, 4). Human and reference baselines are drawn but not ranked. Numbers keep the precision their source printed.
The Health Evals Index
The index is this site's own calculation, the one place where boards are combined. It asks a narrower question than any board: across the health boards it uses, where does a model tend to land?
- On each benchmark we take a model's best headline result on the board's own 0 to 100 scale (a 0.42 board score counts as 42). Benchmarks where lower is better are flipped. Left out: harness plus model pairings, products, research systems, human and baseline references.
- Some benchmarks run easy and some hard. For each one we measure how far its scores sit above or below the same models' scores on the other benchmarks, and take that amount off every result on it. An easy benchmark's results come down a little and a hard one's go up, so a model reads the same whichever benchmarks it happens to have.
- The benchmarks closest to the health questions people ask about themselves count most: core 1, supporting 0.6, narrow 0.35. A benchmark near its ceiling counts for less as well, the nearer its leader is to 100 the less, since it no longer tells strong models apart. A benchmark's weight is the two multiplied.
- A model's adjusted average is the weighted average of its adjusted results. A benchmark it has no result on is left out: never filled in with an estimate, and never counted against it.
- The field average, 55.3 (the average level of every ranked model), is mixed in as if it were 1.5 more benchmarks, so a model with few results is pulled a little toward it: 60% with 1 board, 42.9% with 2, 33.3% with 3, 27.3% with 4. One strong result still counts, but it cannot lead on its own.
Worked example: Claude Opus 5.5, rank 1 of 53
Claude Opus 5.5 has results on 3 of 4 benchmarks. Each result is read on its benchmark's 0 to 100 scale, the benchmark's offset comes off, and the adjusted results are averaged by weight:
| Benchmark | Score as printed | Out of 100 | Runs | Adjusted | Weight |
|---|---|---|---|---|---|
| HealthBench Professional | 0.656 | 65.6 | −2.5 hard | 68.1 | 0.446 |
| HealthBench | 68.1 | 68.1 | −0.5 level | 68.6 | 0.798 |
| Health Optimization Bench: incretin evidence synthesis | no result, not counted | ||||
| SAFE-Teen (Vals AI) | 27 | 73.0 | +4.8 easy | 68.2 | 0.131, near its ceiling |
| Adjusted average of 3, by weight | 68.4 | ||||
| Field average, counts for 33.3% with 3 of 4 | 55.3 | ||||
| Index | 64.0 | ||||
The adjusted average is each adjusted result times its weight, added up and divided by the sum of the weights: (68.1 × 0.446 + 68.6 × 0.798 + 68.2 × 0.131) ÷ (0.446 + 0.798 + 0.131) = 68.4. Then the field average takes its share: 68.4 × (1 − 0.333) + 55.3 × 0.333 = 64.0. Select any column of the index to see the same working for that model.
Each benchmark keeps its own scale: a leader reaches 100 only by acing it. The index moves when a benchmark joins or leaves it, and a new model can move the offsets and the field average a little. We have scored the index this way since 31 October 2026. Until then a score resting on fewer benchmarks was cut back in proportion, which sank newer models whose results were only on the newer benchmarks, and every benchmark counted alike.
The benchmarks in use now:
| Benchmark | Counts as | Runs | Leader, out of 100 | Weight |
|---|---|---|---|---|
| HealthBench Professional | supportingclinicians' workplace tasks rather than people asking about their own health | −2.5 hard | 70.3 | 0.446 |
| HealthBench | corerealistic health conversations with people, graded on physician rubrics | −0.5 level | 68.1 | 0.798 |
| Health Optimization Bench: incretin evidence synthesis | supportingprevention and optimization questions, one area of personal health | +4.0 easy | 84.9 | 0.226, near its ceiling |
| SAFE-Teen (Vals AI) | narrowteen mental-health safety, one population | +4.8 easy | 85.0 | 0.131, near its ceiling |
Benchmarks with results from fewer than 5 models are not used. Every model with a result on any of them is ranked, so one score may rest on a single benchmark and another on several. Every entry says how many it covers ("2 of 4"), the index views can be narrowed to models with results on more of them, and selecting a column of the index shows the working: each benchmark's printed score, its value out of 100, how far that benchmark runs, the adjusted result, its weight, and the field average's share.
What we never do
- Combine boards in any way other than the index above, or call any model the best for health.
- Estimate a score a publisher did not print, or borrow a similar model's number.
- Change the scores a board shows. A publisher's composite index appears exactly as printed, on its own board.
Retiring a benchmark
A benchmark leaves the live boards when models saturate it or nobody publishes current results. It stays listed under retired evaluations with the reason.
Corrections
If a number does not match its source, send us the link and page. Corrections appear on the updates page.