Research

How models perform on clinical work

Published results from clinical benchmarks, and the benchmarks we build ourselves.

The Clinical Benchmarks Index

clinicalbenchmarks.ai

Clinical Benchmarks Index

150 models · 25 labs · 10 boards
Data as of 23 Oct 2026

Index out of 100
01Claude Opus 5Anthropic63.2
02GPT-5.6 SolOpenAI61.9
03Claude Fable 5Anthropic61.1
04Gemini 3.1 ProGoogle58.0
05Muse SparkMeta56.1
06Kimi K3Moonshot AI54.5
07GPT-5OpenAI54.1
08GPT-5.2OpenAI53.6
09Claude Sonnet 5.5Anthropic52.2
10GPT-5.4OpenAI51.8
11Claude Opus 5.5Anthropic51.8
12GPT-5.5OpenAI50.5
13Claude Fable 5.1Anthropic50.3
14Claude Opus 4.8Anthropic50.3
15Qwen3.5 397B A17BAlibaba50.3
16Muse Spark 1.1Meta50.3
28Grok 4.7SpaceX AI40.7
35GLM 5.3Zhipu37.3
39DeepSeek V4 ProDeepSeek36.2
51MiniMax M3MiniMax31.8
52MiMo V2.6 ProXiaomi31.3
58Hy4 PreviewTencent30.0
63InklingThinking Machines29.3
82Ling 3.0 FlashAnt Group24.7
83Fleming-R1-32BUbiquant24.7
89Mistral Medium 3.5Mistral23.7
Clinical reasoning and knowledge
Documentation and coding
EHR and workflow agents
Safety
The index is calculated from published results alone: a model’s weighted average over the boards it has results on, times a confidence multiplier that is full from six boards. Each column of dots is one point out of 100. The pips are the ten boards, numbered 01 to 10; point at a board to light its pips.Each row of dots in a column is one point out of 100. The ten boards are numbered 01 to 10; point at a board to see which models have a result on it.
All 150 models Method

The ten boards

Each board is read on its own scale from 0 to 100. These are the weights and notes used in this snapshot.

Clinical reasoning and knowledge

HealthBench Professional
Weight 0.743
MedXpertQA (MM)near its ceiling
Weight 0.463
MAST
Weight 0.995
MedHELM
Weight 0.87

Documentation and coding

MedScribe (Vals AI)near its ceiling
Weight 0.25
MedCode (Vals AI)
Weight 0.911

EHR and workflow agents

PhysicianBench
Weight 0.79
EHR-Complex
Weight 0.875

Safety

MedPICnear its ceiling
Weight 0.482
First, Do NOHARM (v2)
Weight 0.507

How to read the index

Start with the weighted average over the boards a model has results on. Multiply that average by a confidence multiplier. The multiplier is full from six boards. Below six, divide the number of boards by six, then take the square root.

Coverage matters: models can have results on different boards. Read the index alongside that coverage and the conditions of each published result.

Data as of · Revision 7334
Method dated

Read the index method

Health Optimization Bench

Incretin evidence ranking

89 tasks · 16 ranking entries
Snapshot

HOB incretin evidence scores out of 100, 4 September 2026
RankModel / labScore barScore
1Claude Fable 5.1Anthropic84.9
2Claude Fable 5Anthropic83.8
3Grok 4.6xAI81.3
4Claude Opus 5Anthropic78.3
5GPT-5.6 Sol (max)OpenAI77.9
6Kimi K3Moonshot AI77.8
7GPT-5.6 Sol (high)OpenAI77.7
8Muse SparkMeta68.4
9Gemini 3.6Google57.9
10InklingThinking Machines53.9
11Claude Sonnet 5Anthropic50.1
12MiniMax M3MiniMax35.5
13MAI ThinkingMicrosoft AI33
14GLM 5.2Zhipu29.2
15Mistral Medium 3.5Mistral14.4
16Nemotron 3.5 LightningNVIDIA7.8
Rubric-graded scores out of 100 on the 89-task incretin evidence suite. Bars share a 0–100 scale. Model configurations are listed separately.
View the ranking

HOB publishes separate subject-suite and incretin evidence rankings. The snapshot above covers evidence-synthesis questions about incretin therapeutics. The subject-suite ranking uses a separate task set and scores.

The open-source harness and public sample let you inspect the tasks, rubric criteria and reference responses. A sample run is not expected to reproduce a full leaderboard score exactly.

Explore HOBOpen-source harnessHugging Face sample

Clinical Reasoning Bench

In development

What was known at each decision

The benchmark is designed to examine decisions as information becomes available, including predictions recorded before test results are revealed.

The method follows an initial differential, findings and orders, expected test results, then a final diagnosis and plan. Results are designed to remain sealed until expectations are recorded. An early reveal can be recorded with its reason.

In development

Clinical Conversations Bench

Designed to test how a model asks relevant questions and offers a clear next step in a health conversation, with plain language and reasons for its answer.

Evaluation at follow-up

Designed to score a model against the outcome at follow-up, keeping what was predicted during the visit distinct from evidence obtained later.

Multilingual clinical speech

Designed to compare machine transcription and translation with a human reference, checking clinical facts, speaker attribution, omissions and invented facts.

Read the evidence

HealthBench Professional

Published scores on physician-written clinical tasks. We maintain the results site; we do not run the benchmark.

HealthBench Hard

Published results on the difficult subset of HealthBench, with document citations.

Health Evals

Published evaluations of models answering personal health questions, with its own index method.

Clinical Evals

Guides to rubric grading, model graders and reading benchmark results.

Evals.Health

A glossary and question reference for healthcare AI evaluation.

More benchmark method references

Individual reference sites explain benchmark methods and show sourced result panels.

Read more

Explore our workExplore the data