Which AI handles health questions well?
Claude Opus 5.5 ranks first at 64.0. Claude Sonnet 5.5 is 2nd at 64.0.
The Health Evals Index is our own calculation from 4 of 6 health benchmarks, 53 models and 193 scores, 35% of them published by the model's maker. Method of 31 October 2026.
Health Evals Index
Out of 100. The squares under each column show which of the 4 benchmarks the score uses.
Tap a column to open that model. Ranks stay the same whichever filter you choose.
Each benchmark, model by model
The top 10 models on each of the 4 benchmarks in our index, on that benchmark's own 0 to 100 scale. Tap a column to read where the score came from.
HealthBench Professional
- 0.703GPT-6 Astra
- 0.692Claude Sonnet 5.5
- 0.660Claude Fable 5
- 0.656Claude Opus 5.5
- 0.648Claude Haiku 5.5
- 0.642GPT-6.1 Sol
- 0.627GPT-6.1 Terra
- 0.623Muse Wonder Full
- 0.621Claude Fable 5.1
- 0.608GPT-6 Luna
37 modelsFull benchmark
HealthBench
- 68.1Claude Opus 5.5
- 67.1Claude Opus 5
- 65.4Claude Sonnet 5.5
- 65.1Baichuan-M3
- 63.3GPT-5.2
- 62.7Claude Fable 5
- 61.6Claude Haiku 5.5
- 60Claude Fable 5.1
- 59.3Claude Opus 4.8
- 58.7Claude Sonnet 5
26 modelsFull benchmark
Health Optimization Bench: incretin evidence synthesisnear ceiling
- 84.9Claude Fable 5.1
- 83.8Claude Fable 5
- 82.2GPT-6 Astra
- 81.3Grok 4.6
- 78.3Claude Opus 5
- 77.9GPT-5.6 Sol
- 77.8Kimi K3
- 68.4Muse Spark
- 67.8Gemini 3.8 Flash
- 57.9Gemini 3.6
17 modelsFull benchmark
SAFE-Teen (Vals AI)near ceiling
- 15GPT-6.1 Sol
- 23Claude Sonnet 5.5
- 23GPT-5.6 Luna
- 26Muse Spark 1.3
- 26Qwen3.8 Max
- 27Claude Opus 5.5
- 31DeepSeek V4.1 Flash
- 33Mistral Large 4
- 40MiMo V2.6 Pro
- 44Kimi K3
17 modelsFull benchmark
Each bar is the model's lab colour. Each benchmark's page says who ran every score.
Every model on every benchmarkIndex against price
43 models with a published list price, on a log price scale. The dashed line runs through models that nothing cheaper beats.
How we put the index together
Each model's page shows the full arithmetic.
- Out of 100. Every score counts on its own benchmark's 0 to 100 scale.
- Easy and hard benchmarks. Some run easy and some hard, so each score is read against how the same models do on the others.
- Weights. Benchmarks built from people's own health conversations count the most.
- Few scores. A model with one or two scores is pulled toward the average of all models, 55.3, so the number of benchmarks appears next to every score.
Each benchmark on its own scale
The top 15 scores on each benchmark, on that benchmark's own scale. We keep the benchmarks apart because each one asks a different question.
Each bar is the model's lab colour. Each benchmark's page says who ran every score.
All boards and what they testEach model's index score, and the benchmark scores underneath it
Read each row left to right. The squares are the model's scores on our index benchmarks, each out of 100 on that benchmark's own scale, and the diamond is where our index puts it. Ranks count all 53 models.
- 1Claude Opus 5.5Anthropic · 3 of 4 benchmarks64.03/4
- 2Claude Sonnet 5.5Anthropic · 3 of 4 benchmarks64.03/4
- 3Claude Fable 5Anthropic · 3 of 4 benchmarks63.43/4
- 4Claude Opus 5Anthropic · 3 of 4 benchmarks63.13/4
- 5GPT-6 AstraOpenAI · 3 of 4 benchmarks62.43/4
- 6Claude Fable 5.1Anthropic · 3 of 4 benchmarks61.73/4
- 7Muse Spark 1.3Meta · 1 of 4 benchmarks60.91/4
- 8Qwen3.8 MaxAlibaba · 1 of 4 benchmarks60.91/4
- 9GPT-6.1 SolOpenAI · 3 of 4 benchmarks60.83/4
- 10Claude Haiku 5.5Anthropic · 2 of 4 benchmarks60.32/4
- 11GPT-5.6 SolOpenAI · 3 of 4 benchmarks59.63/4
- 12Baichuan-M3Baichuan · 1 of 4 benchmarks59.41/4
- 13Muse Wonder FullMeta · 1 of 4 benchmarks59.11/4
- 14Kimi K3Moonshot AI · 3 of 4 benchmarks58.53/4
- 15Muse WonderMeta · 1 of 4 benchmarks58.41/4
- 16Mistral Large 4Mistral · 1 of 4 benchmarks58.11/4
- 17Claude Opus 4.8Anthropic · 2 of 4 benchmarks57.92/4
- 18Grok 4.6SpaceX AI · 2 of 4 benchmarks57.92/4
- 19Muse Spark 1.1Meta · 1 of 4 benchmarks57.91/4
- 20Muse SparkMeta · 2 of 4 benchmarks57.52/4
- 21GPT-5.6 LunaOpenAI · 3 of 4 benchmarks57.43/4
- 22GPT-6.1 TerraOpenAI · 2 of 4 benchmarks57.42/4
- 23GPT-5.6 TerraOpenAI · 2 of 4 benchmarks57.12/4
- 24GPT-5.2OpenAI · 2 of 4 benchmarks57.02/4
- 25Claude Sonnet 5Anthropic · 3 of 4 benchmarks56.83/4
- 26GPT-6 LunaOpenAI · 2 of 4 benchmarks56.82/4
- 27Grok 4.7SpaceX AI · 2 of 4 benchmarks56.52/4
- 28GPT OSS 120BOpenAI · 1 of 4 benchmarks56.41/4
- 29GPT-6 SolOpenAI · 2 of 4 benchmarks56.32/4
- 30GPT-5.5OpenAI · 2 of 4 benchmarks55.72/4
- 31DeepSeek V4.1 FlashDeepSeek · 2 of 4 benchmarks55.52/4
- 32MiMo V2.6 ProXiaomi · 1 of 4 benchmarks55.31/4
- 33GPT-5OpenAI · 2 of 4 benchmarks55.02/4
- 34GPT-5.3 ChatOpenAI · 1 of 4 benchmarks55.01/4
- 35Claude Opus 4.7Anthropic · 1 of 4 benchmarks54.91/4
- 36Gemini 3.8 FlashGoogle · 3 of 4 benchmarks54.83/4
- 37Gemini 3.6Google · 1 of 4 benchmarks54.71/4
- 38Gemini 4 ArgonGoogle · 2 of 4 benchmarks54.62/4
- 39GPT-5.4OpenAI · 2 of 4 benchmarks54.02/4
- 40GLM 5.3Zhipu · 2 of 4 benchmarks53.92/4
- 41InklingThinking Machines · 1 of 4 benchmarks53.11/4
- 42Claude Sonnet 4.6Anthropic · 1 of 4 benchmarks51.91/4
- 43Grok 4.5SpaceX AI · 1 of 4 benchmarks51.31/4
- 44GPT-5.1OpenAI · 2 of 4 benchmarks51.22/4
- 45GPT-5.5 InstantOpenAI · 2 of 4 benchmarks51.12/4
- 46Gemini 3.6 FlashGoogle · 1 of 4 benchmarks50.51/4
- 47GPT OSS 20BOpenAI · 1 of 4 benchmarks50.41/4
- 48Claude Haiku 4.5Anthropic · 1 of 4 benchmarks47.11/4
- 49MiniMax M3MiniMax · 2 of 4 benchmarks43.82/4
- 50MAI-Thinking-1Microsoft · 2 of 4 benchmarks43.52/4
- 51GLM 5.2Zhipu · 1 of 4 benchmarks43.31/4
- 52Mistral Medium 3.5Mistral · 1 of 4 benchmarks37.31/4
- 53Nemotron 3.5 LightningNVIDIA · 1 of 4 benchmarks34.71/4
Method of 31 October 2026. Tap any square to read the sentence the score came from.
How it works Download the indexHow a published score reaches this page
Every number below is a count of rows in our dataset.
documents we read, each kept with the exact sentence and page
published scores taken from them
health benchmarks we follow
models in our index
We only add a score when we can point to the document, the sentence and the page. Scores a model's maker published stay in, and they are marked so you can tell them apart.
Read the sources