HealthBench Professional
Updated November 2, 2026

HealthBench Professional Leaderboard

GPT-6 Astra (Anthropic run) leads at 0.703, ahead of Claude Sonnet 5.5 (0.692) and Claude Fable 5 (0.660). OpenAI's benchmark scores models on 525 tasks physicians selected from real clinician conversations, graded against physician-written rubrics on a 0 to 1 scale.
43
models ranked
9
labs represented
525
physician-selected tasks
0.437
physician-written responses

Length-adjusted score

The score the board ranks on: rubric points earned, adjusted for answer length as OpenAI reports it. 43 models, 525-task set, compiled from published documents.

  • OpenAI
  • Anthropic
  • Meta
  • SpaceX AI
  • Google
  • Moonshot AI
  • Zhipu
  • DeepSeek
  • Microsoft
Sources

Score before the length adjustment

Rubric points earned with no adjustment for answer length, for the 30 of 43 models whose source prints it. Longer answers earn more here, so the order differs from the board.

  • OpenAI
  • Anthropic
  • Google
  • Moonshot AI
  • Zhipu
  • DeepSeek
Sources

Full ranking

Click a column to sort. Each model links to its full result; the line under the name says which document the score was read from. List API prices per 1M tokens as of November 2, 2026. GPT-5.6 models charge higher rates above 272K input tokens. MAI-Thinking-1 is in public preview on Microsoft Foundry without final list pricing.

License
1
0.703
0.7401.05M$10.00$50.002026-09-03proprietary
2
0.692
0.7711M$2.00$10.002026-09-28proprietary
3
0.660
0.7031.0M$10.00$50.002026-06-09proprietary
4
0.656
0.7711M$4.00$20.002026-09-22proprietary
5
0.655
–1.05M$0.10$0.502026-09-22proprietary
6
0.648
0.7101M$0.10$0.502026-10-07proprietary
7
0.647
0.6821.05M$10.00$50.002026-09-03proprietary
8
0.642
0.6721.05M$2.00$10.002026-09-29proprietary
9
0.633
0.6891.0M$10.00$50.002026-06-09proprietary
10
0.627
0.6421.05M$1.00$5.002026-10-26proprietary
11
0.623
–1M––2026-10-12proprietary
12
0.621
0.7421.0M$10.00$50.002026-09-01proprietary
13
0.608
0.6121.05M$0.10$0.502026-09-22proprietary
14
0.608
0.5951.05M$2.00$10.002026-09-22proprietary
15
0.605
0.6411.05M$4.00$20.002026-07-09proprietary
16
0.605
–1M––2026-10-12proprietary
17
0.598
0.7341.0M$5.00$25.002026-07-24proprietary
18
0.593
–1.0M$1.25$4.252026-07-09proprietary
19
0.578
0.6241.0M$2.00$10.002026-06-30proprietary
20
0.577
0.6241.05M$2.00$12.002026-07-09proprietary
21
0.574
–1.0M$5.00$25.002026-05-28proprietary
22
0.567
–500K$2.00$6.002026-09-21proprietary
23
0.558
–1.0M$5.00$25.002026-05-28proprietary
24
0.557
0.5981.05M$0.20$1.202026-07-09proprietary
25
0.551
0.642–$2.00$10.002026-09-30proprietary
26
0.541
–1.0M––2026-04-08proprietary
27
0.540
0.5661.05M$4.00$20.002026-07-09proprietary
28
0.532
0.5691M$3.00$15.002026-07-16Kimi K3 License (open weights)
29
0.528
0.5791M$1.40$4.402026-08-18open
30
0.519
–1M$5.00$25.002026-04-16proprietary
31
0.518
0.5721.05M$5.00$30.002026-04-23proprietary
32
0.517
0.5901M$0.75$3.752026-09-02proprietary
33
0.507
0.5291M$0.30$1.202026-09-10open
34
0.485
–500K$2.00$6.002026-08-12proprietary
35
0.481
0.5191.05M$2.50$15.002026-03-05proprietary
36
0.462
0.510400K$1.25$10.002025-08-07proprietary
37
0.459
0.500400K$1.75$14.002025-12-11proprietary
38
0.442
–1M$3.00$15.002026-02-17proprietary
39
0.441
0.4681.05M$0.20$1.202026-07-09proprietary
40
0.396
0.480400K$1.25$10.002025-11-12proprietary
41
0.384
0.407400K$5.00$30.002026-05-05proprietary
42
0.350
–256K$2.00$8.002026-08-12proprietary
43
0.322
–200K$1.00$5.002025-10-15proprietary

What the benchmark measures

HealthBench Professional evaluates language models on the work clinicians bring to a model: care consults such as differential diagnosis and management questions, clinical writing and documentation, and medical research. The 525 tasks were selected by physicians from 15,079 real clinician conversations, difficult cases were deliberately overweighted, and roughly a third of the set is adversarial.

190 physicians across 50 countries and 26 specialties wrote and adjudicated the rubrics. The full construction is on the benchmark page.

How to read the scores

A score is earned rubric points divided by possible points, clipped to 0 to 1. The length adjustment subtracts credit from answers longer than 2,000 characters and adds it to shorter ones, so verbosity cannot buy credit. Absolute numbers run low by design: the best result is 0.703, and physician-written responses score 0.437.

Scores depend on the grader version, the reasoning effort and the deployment setting a document reports, so 2 numbers from different documents can differ for reasons other than model quality. How each one is read is on the methodology page.

Data

The full results are published as JSON and CSV at stable URLs, licensed CC BY 4.0. Every score links to the document it was copied from on the sources page; citation format and version history are on the data page.