Health AI Benchmarks

HealthBench Professional

When a doctor brings a real work question to an AI, would another doctor sign off on the reply?

More about this benchmark

Physician-selected workplace tasks, from care consults to documentation and research, graded on physician-written rubrics.

525 tasks that physicians picked out of 15,079 real workplace AI conversations, spanning care consults, clinical documentation, and medical research, each judged on a rubric physicians wrote for it.

Each item
Clinician task / conversation
The model gets
A single- or multi-turn physician-authored conversation ending with a clinician request.
It returns
The next assistant response, graded with physician-written rubric items.
Setting
525 selected tasks; default GPT-5.4 low-reasoning grader; primary score adjusted for final-answer length.
Released
22 Apr 2026
Scale
0 to 1
Tasks
525
Grader model
GPT-5.4, low reasoning effort
Use cases
3
Countries represented
50
Languages
52
Physicians involved
190
Specialties
26
Paper version
April 2026 paper v1; length-adjusted primary score
Candidate pool
15079
Physician baseline
0.437

Reading these numbers

Grader GPT-5.4 at low reasoning effort with a length adjustment. Physician-written responses score 0.437 on the same rubrics.

Scores come from runs by the benchmark's maintainers or independent researchers.

Results
43 from 37 models
Latest
Oct 2026
Published by
OpenAI
Index weight
0.446, supporting
Sources
19

Counts toward the Health Evals Index as a supporting benchmark, at weight 0.446: clinicians' workplace tasks rather than people asking about their own health.

Where every published HealthBench Professional score sits

43 scores, one column each, highest first, on the benchmark's own scale. Higher is better.

Each column is the model's lab colour. The table below says who ran each score.

Every published score, ranked within this benchmark only

43 scores. Higher is better. Where the source gave a range, the whisker shows it. Open any row to read the sentence the score came from.

1GPT-6 Astra (Anthropic run)OpenAI · Anthropic public-API…0.703IndependentSep 2026[1]
On HealthBench Professional at max effort, length adjustment changes the ranking. After: GPT-6 Astra (70.3%) > Claude Sonnet 5.5 (69.2%) > Claude Opus 5.5 (65.6%) > Claude Fable 5.1 (62.1%).
In the source
70.3%
Where to find it
p. 138, paragraph above Figure 8.15.B, HealthBench Professional after length adjustment.
How it was run
Anthropic public-API reproduction; max effort; no system prompt; Opus 4.8 grader; length-adjusted; raw 74.0%
Reported by
independent run
Document
[1] Claude Sonnet 5.5 System Card
2Claude Sonnet 5.5Anthropic · max effort0.692Model's makerSep 2026[1]
On HealthBench Professional at max effort, length adjustment changes the ranking. After: GPT-6 Astra (70.3%) > Claude Sonnet 5.5 (69.2%) > Claude Opus 5.5 (65.6%) > Claude Fable 5.1 (62.1%).
In the source
69.2%
Where to find it
p. 138, Section 8.15.2 and paragraph above Figure 8.15.B.
How it was run
Anthropic; max effort; Opus 4.8 grader; safety classifiers enabled; length-adjusted; raw 77.1%
Reported by
vendor-reported
Document
[1] Claude Sonnet 5.5 System Card
3Claude Fable 5Anthropic · length-adjusted0.660Model's makerJun 2026[2]
HealthBench Professional 66.0 - 64.7 56.9 51.8 -
In the source
66.0
Where to find it
p. 252, Table 8.1.A, row HealthBench Professional, column Mythos 5 (Fable 5 column '-'); Figure 8.18.2.A p. 298 bar label 66.0% on Claude Mythos 5
How it was run
length-adjusted, Anthropic protocol: adaptive thinking at max effort, Claude Opus 4.8 grader, averaged over 5 trials, no tools or custom system prompt (raw 70.3%). Measured as Claude Mythos 5; Anthropic itself prints 66.0 in the Fable 5 column of the Opus 5 card with footnote Mythos 5.
Reported by
vendor-reported
Document
[2] Claude Fable 5 and Claude Mythos 5 System Card
4Claude Opus 5.5Anthropic · adaptive max effort0.656Model's makerSep 2026[3]
8.15.2 HealthBench Professional results length adjustment, which penalizes verbose model responses, Claude Opus 5.5 achieved a score of 65.6%.
In the source
65.6%
Where to find it
pp. 213–214, Section 8.15.2, paragraph and Figure 8.15.2.A.
How it was run
Anthropic; adaptive max effort; Opus 4.8 grader; five trials; no tools or custom system prompt; safety classifiers and refusal fallback to Opus 5; length-adjusted; raw 77.1%
Reported by
vendor-reported
Document
[3] Claude Opus 5.5 System Card
5GPT-6 Luna (Anthropic run)OpenAI · Run by Anthropic through…0.655IndependentOct 2026[4]
HealthBench Professional panel: 65.5% (GPT-6 Luna (OpenAI)); caption: GPT-6 Luna's scores come from our own runs through OpenAI's public API and are not OpenAI's published figures.
In the source
65.5%
Where to find it
p. 132, Figure 8.11.1.A, HealthBench Professional panel, GPT-6 Luna (OpenAI) diamond label (visually read printed label); Section 8.11.1 text on pp. 131-132
How it was run
Run by Anthropic through OpenAI's public API; max effort; Opus 4.8 grader; length-adjusted; not OpenAI's published figure
Reported by
independent run
Document
[4] Claude Haiku 5.5 System Card
6Claude Haiku 5.5Anthropic · max effort0.648Model's makerOct 2026[4]
Max effort scored highest, at 64.8% (71.0% raw); thirteen repeat runs at max on October 4, the day of the lower levels, averaged 64.2%.
In the source
64.8%
Where to find it
p. 133, Section 8.11.3; same value in p. 111, Table 8.1.A, row HealthBench Professional (Length-adjusted), column Haiku 5.5 (printed 64.8) and p. 132, Figure 8.11.1.A
How it was run
Anthropic; max effort; Opus 4.8 grader; safety classifiers enabled; length-adjusted; raw 71.0%; five runs (October 1 and 2); thirteen repeat max runs on October 4 averaged 64.2%
Reported by
vendor-reported
Document
[4] Claude Haiku 5.5 System Card
7GPT-6 AstraOpenAI · length-adjusted0.647Model's makerSep 2026[5]
Evaluation | GPT-5.5 | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-6 Astra | GPT-6 Sol | GPT-6 Luna HealthBench Professional length-adjusted | 51.8 (57.2, 3818) | 60.5 (64.1, 3228) | 57.7 (62.4, 3618) | 55.7 (59.8, 3389) | 64.7 (68.2, 3185) | 60.8 (59.5, 1573) | 60.8 (61.2, 2119)
In the source
64.7
Where to find it
Section 11.4.1, Table 29, HealthBench Professional length-adjusted, GPT-6 Astra column; September 22 correction.
How it was run
length-adjusted, max reasoning effort (68.2 unadjusted, 3,185 mean response chars, as corrected by OpenAI on September 22; first printed as 69.5 and 4,097); GPT-6 Astra system card Table 6, column 'gpt-6 Astra'.
Reported by
vendor-reported
Document
[5] GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)
8GPT-6.1 SolOpenAI · length-adjusted0.642Model's makerSep 2026[6]
EVALUATION | GPT-5.5 | GPT-5.6 SOL | GPT-5.6 TERRA | GPT-5.6 LUNA | GPT-6 ASTRA | GPT-6 SOL | GPT-6 LUNA | GPT-6.1 SOL HealthBench Professional length-adjusted | 51.8 (57.2, 3818) | 60.5 (64.1, 3228) | 57.7 (62.4, 3618) | 55.7 (59.8, 3389) | 64.7 (68.2, 3185) | 60.8 (59.5, 1573) | 60.8 (61.2, 2119) | 64.2 (67.2, 3038)
In the source
64.2
Where to find it
Section 5.1, Table 7, HealthBench Professional length-adjusted, GPT-6.1 Sol column
How it was run
OpenAI; length-adjusted; 64.2 (67.2, 3038) (adjusted, raw, mean response characters); addendum Table 7 refers to the GPT-6 Astra card for method
Reported by
vendor-reported
Document
[6] Addendum to GPT-6 Astra System Card: GPT-6.1 Sol
9Claude Fable 5 (September card)Anthropic · length-adjusted0.633Model's makerSep 2026[7]
HealthBench Professional | Length-adjusted score | Claude Fable 5 | 63.3%
In the source
63.3%
Where to find it
p. 199, Figure 8.17.2.A, Claude Fable 5 length-adjusted bar (visually read printed labels).
How it was run
Measured as Claude Fable 5; Anthropic September card; length-adjusted; adaptive max effort; Opus 4.8 grader; five trials; no tools or custom system prompt; raw 68.9%
Reported by
vendor-reported
Document
[7] Claude Fable 5.1 and Claude Mythos 5.1 System Card
10GPT-6.1 TerraOpenAI · length-adjusted0.627Model's makerOct 2026[8]
EVALUATION GPT-5.5 GPT-5.6-SOL GPT-5.6-TERRA GPT-5.6-LUNA GPT-6 ASTRA GPT-6-SOL GPT-6-LUNA GPT-6.1-SOL GPT-6.1-TERRA HealthBench Professional length-adjusted 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389) 64.7 (68.2, 3185) 60.8 (59.5, 1573) 60.8 (61.2, 2119) 64.2 (67.2, 3038) 62.7 (64.2, 2531)
In the source
62.7
Where to find it
Section 5.1, Table 5, HealthBench Professional length-adjusted, GPT-6.1 Terra column
How it was run
OpenAI; length-adjusted; 62.7 (64.2, 2531) (adjusted, raw, mean response characters); addendum Table 5 refers to the GPT-6 Astra card for method
Reported by
vendor-reported
Document
[8] Addendum to GPT-6 Astra System Card: GPT-6.1 Terra
11Muse Wonder FullMeta · Meta internal run0.623Model's makerOct 2026[9]
HealthBench Pro (Open-ended health queries) | Muse Wonder Full (max) 62.3 | Muse Wonder (max) 60.5 | Claude Fable 5.1 (max) 62.1 | GPT-6 Astra (max) 65.3 | Claude Opus 5.5 (max) 65.6 Methodology (p. 4): HealthBench Pro evaluates responses to 525 professional-grade health questions. Models answer without tools, and a GPT-5.4 grader scores each response against physician-written rubrics. We run the official tasks and grading in a common internal framework. The primary metric is the length-normalized weighted rubric score.
In the source
62.3
Where to find it
p. 3, results table (image), HealthBench Pro row, Muse Wonder Full (max) column; methodology p. 4
How it was run
Meta internal run; length-normalized weighted rubric score, GPT-5.4 grader, no tools, max reasoning effort (Muse Wonder Model Card p. 3 table, Muse Wonder Full (max) column)
Reported by
vendor-reported
Document
[9] Muse Wonder Model Card
12Claude Fable 5.1Anthropic · length-adjusted0.621Model's makerSep 2026[7]
On HealthBench Professional, Claude Fable 5.1 achieved a raw score of 74.2%, ahead of Claude Opus 5 at 73.4%, Fable 5 at 68.9%, and Claude Sonnet 5 at 62.4%. After length adjustment, which penalizes verbose model responses, Fable 5.1 achieved a score of 62.1%.
In the source
62.1%
Where to find it
p. 199, sec. 8.17.2 (same figure printed as 62.1% in Table 8.1.A, p. 167)
How it was run
length-adjusted (method published in the HealthBench Professional paper); Anthropic protocol: adaptive thinking at max effort, Claude Opus 4.8 grader, averaged over five trials, no tools or customized system prompt; Fable 5.1 run with safety classifiers active and a refusal-fallback to Claude Opus 5 (raw 74.2%).
Reported by
vendor-reported
Document
[7] Claude Fable 5.1 and Claude Mythos 5.1 System Card
13GPT-6 LunaOpenAI · length-adjusted0.608Model's makerSep 2026[5]
Evaluation | GPT-5.5 | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-6 Astra | GPT-6 Sol | GPT-6 Luna HealthBench Professional length-adjusted | 51.8 (57.2, 3818) | 60.5 (64.1, 3228) | 57.7 (62.4, 3618) | 55.7 (59.8, 3389) | 64.7 (68.2, 3185) | 60.8 (59.5, 1573) | 60.8 (61.2, 2119)
In the source
60.8
Where to find it
Section 11.4.1, Table 29, HealthBench Professional length-adjusted, GPT-6 Luna column; September 22 correction.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 60.8 (61.2, 2119) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[5] GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)
13GPT-6 SolOpenAI · length-adjusted0.608Model's makerSep 2026[5]
Evaluation | GPT-5.5 | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | GPT-6 Astra | GPT-6 Sol | GPT-6 Luna HealthBench Professional length-adjusted | 51.8 (57.2, 3818) | 60.5 (64.1, 3228) | 57.7 (62.4, 3618) | 55.7 (59.8, 3389) | 64.7 (68.2, 3185) | 60.8 (59.5, 1573) | 60.8 (61.2, 2119)
In the source
60.8
Where to find it
Section 11.4.1, Table 29, HealthBench Professional length-adjusted, GPT-6 Sol column; September 22 correction.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 60.8 (59.5, 1573) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[5] GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)
15GPT-5.6 SolOpenAI · length-adjusted0.605Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
60.5
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.6-SOL
How it was run
length-adjusted, max reasoning effort (64.1 unadjusted, 3,228 mean response chars); GPT-5.6 system card Table 6, column GPT-5.6-SOL. API max-effort setting, not the August ChatGPT production setting measured in row 492.
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
15Muse WonderMeta · Meta internal run0.605Model's makerOct 2026[9]
HealthBench Pro (Open-ended health queries) | Muse Wonder Full (max) 62.3 | Muse Wonder (max) 60.5 | Claude Fable 5.1 (max) 62.1 | GPT-6 Astra (max) 65.3 | Claude Opus 5.5 (max) 65.6 Methodology (p. 4): HealthBench Pro evaluates responses to 525 professional-grade health questions. Models answer without tools, and a GPT-5.4 grader scores each response against physician-written rubrics. We run the official tasks and grading in a common internal framework. The primary metric is the length-normalized weighted rubric score.
In the source
60.5
Where to find it
p. 3, results table (image), HealthBench Pro row, Muse Wonder (max) column; methodology p. 4
How it was run
Meta internal run; length-normalized weighted rubric score, GPT-5.4 grader, no tools, max reasoning effort (Muse Wonder Model Card p. 3 table, Muse Wonder (max) column)
Reported by
vendor-reported
Document
[9] Muse Wonder Model Card
17Claude Opus 5Anthropic · length-adjusted0.598Model's makerJul 2026[11]
Claude Opus 5 achieved a raw score of 73.4%, which is the highest amongst all Claude models, ahead of Claude Mythos 5 at 70.3%, Claude Opus 4.8 at 60.3%, and Claude Sonnet 5 at 62.4%. After length adjustment, which penalizes verbose model responses, Claude Opus 5 achieved a score of 59.8%.
In the source
59.8%
Where to find it
p. 189, section 8.15.2 HealthBench Professional results; also Table 8.1.A p. 152 ('HealthBench Professional 59.8 ...')
How it was run
length-adjusted, Anthropic protocol: adaptive thinking at max effort, Claude Opus 4.8 grader, averaged over 5 trials, no tools or custom system prompt (raw 73.4%).
Reported by
vendor-reported
Document
[11] System Card: Claude Opus 5
18Muse Spark 1.1Meta · length-normalized0.593Model's makerJul 2026[12]
Health | HealthBench Professional | 59.3 | 54.1 | 41.6 | 55.8 | 51.8 (Figure 44 image table row; columns Muse Spark 1.1, Muse Spark, Gemini 3.1 Pro (high), Opus 4.8 (max), GPT 5.5 (xhigh))
In the source
59.3
Where to find it
p. 101, Figure 44 'General capability benchmark results' (image), row HealthBench Professional, column Muse Spark 1.1; protocol p. 104 (printed 103): HealthBench Pro comprises 525 evaluation data points graded by rubrics. We use GPT-5.4 with low reasoning effort as the grader and report the length-normalized rubric score as done in their paper.
How it was run
length-normalized, GPT-5.4 low-reasoning grader, xhigh reasoning via Meta Model API (Muse Spark 1.1 Evaluation Report Figure 44)
Reported by
vendor-reported
Document
[12] Muse Spark 1.1 Evaluation Report
19Claude Sonnet 5Anthropic · length-adjusted0.578Model's makerJun 2026[13]
HealthBench Professional 57.8 44.2 51.8 -
In the source
57.8
Where to find it
p. 115, Table 8.1.A, row HealthBench Professional, column Claude Sonnet 5; Figure 8.12.2.A p. 139
How it was run
length-adjusted, Anthropic protocol: adaptive thinking at max effort, Claude Opus 4.8 grader, averaged over 5 trials, no tools or custom system prompt (raw 62.4%).
Reported by
vendor-reported
Document
[13] System Card: Claude Sonnet 5
20GPT-5.6 TerraOpenAI · length-adjusted0.577Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
57.7
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.6-TERRA
How it was run
length-adjusted, max reasoning effort (62.4 unadjusted, 3,618 mean response chars); GPT-5.6 system card Table 6, column GPT-5.6-TERRA.
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
21Claude Opus 4.8 (Opus 4.8 grader)Anthropic · length-adjusted0.574Model's makerJun 2026[13]
HealthBench Professional | Length-adjusted score (%) | Claude Opus 4.8 | 57.4%
In the source
57.4%
Where to find it
p. 139, Figure 8.12.2.A, Claude Opus 4.8 bar (visually read printed labels).
How it was run
Anthropic June evaluation; length-adjusted; adaptive max effort; Opus 4.8 grader; five trials; no tools or custom system prompt
Reported by
vendor-reported
Document
[13] System Card: Claude Sonnet 5
22Grok 4.7SpaceX AI · xHigh effort0.567Model's makerSep 2026[14]
Grok 4.7 xHigh Grok 4.6 High GPT-5.6 Sol Max Fable 5.1 Max Clinical reasoningHealthBench Professional 56.7% 48.5% 60.5% 62.1%
In the source
56.7%
Where to find it
Model Improvements comparison table, Clinical reasoning / HealthBench Professional row; Grok 4.7 column; September 21, 2026.
How it was run
SpaceXAI evaluation; xHigh effort; HealthBench Professional
Reported by
vendor-reported
Document
[14] Introducing Grok 4.7
23Claude Opus 4.8Anthropic · length-adjusted0.558Model's makerMay 2026[15]
Claude Opus 4.8 scores 55.8%, a meaningful improvement over Claude Opus 4.7 at 51.9% and Claude Sonnet 4.6 at 41.7%.
In the source
55.8%
Where to find it
p. 228, section 8.14.1 HealthBench Professional; Figure 8.14.A p. 229
How it was run
length-adjusted, adaptive thinking at max effort, Claude Sonnet 4.6 grader (the grader used by the Opus 4.8 card itself); the other Claude rows on this board use the Claude Opus 4.8 grader, under which Anthropic later prints 57.4 for this model.
Reported by
vendor-reported
Document
[15] System Card: Claude Opus 4.8
24GPT-5.6 LunaOpenAI · length-adjusted0.557Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
55.7
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.6-LUNA
How it was run
length-adjusted, max reasoning effort (59.8 unadjusted, 3,389 mean response chars); GPT-5.6 system card Table 6, column GPT-5.6-LUNA. API max-effort setting, not the August ChatGPT production setting measured in row 493.
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
25Gemini 4 ArgonGoogle · Artificial Analysis…0.551IndependentOct 2026[16]
HealthBench Professional: Length-Adjusted Score | Gemini 4 Argon (high) | 55.1%
In the source
55.1%
Where to find it
Score section, Length-Adjusted Score chart, Gemini 4 Argon (high) bar label; raw score 64.2% on the Raw Score tab
How it was run
Artificial Analysis independent run, high reasoning; GPT-5.4 (low) judge, 4 repeats per task; length-adjusted (64.2 raw, 5,133 mean response chars)
Reported by
independent run
Document
[16] HealthBench Professional Benchmark | Artificial Analysis
26Muse SparkMeta · length-normalized0.541Model's makerJul 2026[12]
Health | HealthBench Professional | 59.3 | 54.1 | 41.6 | 55.8 | 51.8 (Figure 44 image table row; columns Muse Spark 1.1, Muse Spark, Gemini 3.1 Pro (high), Opus 4.8 (max), GPT 5.5 (xhigh))
In the source
54.1
Where to find it
p. 101, Figure 44 (image), row HealthBench Professional, column Muse Spark; protocol p. 104 (printed 103)
How it was run
length-normalized, GPT-5.4 low-reasoning grader; Muse Spark (1.0) column in the Muse Spark 1.1 Evaluation Report Figure 44
Reported by
vendor-reported
Document
[12] Muse Spark 1.1 Evaluation Report
27GPT-5.6 Sol (August)OpenAI · ChatGPT production0.540Model's makerAug 2026[17]
HealthBench Professional 32.9 (33.8, 2,285) 38.4 (40.7, 2,775) 54.0 (56.6, 2,894) 44.1 (46.8, 2,920)
In the source
54.0
Where to find it
p. 11, section 5.1 HealthBench, table "Reported as length-adjusted score (unadjusted, mean response length in characters)", column GPT-5.6 Sol (August)
How it was run
ChatGPT production (Instant) deployment setting, length-adjusted, GPT-5.6 August Updates PDF column GPT-5.6 Sol (August) (56.6 unadjusted, 2,894 chars)
Reported by
vendor-reported
Document
[17] GPT-5.6 - August Updates (system card addendum)
28Kimi K3Moonshot AI · Artificial Analysis…0.532IndependentOct 2026[16]
HealthBench Professional: Length-Adjusted Score | Kimi K3 (max) | 53.2%
In the source
53.2%
Where to find it
Score section, Length-Adjusted Score chart, Kimi K3 (max) bar label; raw score 56.9% on the Raw Score tab
How it was run
Artificial Analysis independent run, max reasoning; GPT-5.4 (low) judge, 4 repeats per task; length-adjusted (56.9 raw, 3,250 mean response chars)
Reported by
independent run
Document
[16] HealthBench Professional Benchmark | Artificial Analysis
29GLM 5.3Zhipu · Artificial Analysis…0.528IndependentOct 2026[16]
HealthBench Professional: Length-Adjusted Score | GLM-5.3 (max) | 52.8%
In the source
52.8%
Where to find it
Score section, Length-Adjusted Score chart, GLM-5.3 (max) bar label; raw score 57.9% on the Raw Score tab
How it was run
Artificial Analysis independent run, max reasoning; GPT-5.4 (low) judge, 4 repeats per task; length-adjusted (57.9 raw, 3,748 mean response chars)
Reported by
independent run
Document
[16] HealthBench Professional Benchmark | Artificial Analysis
30Claude Opus 4.7Anthropic · length-adjusted0.519Model's makerMay 2026[15]
Claude Opus 4.8 scores 55.8%, a meaningful improvement over Claude Opus 4.7 at 51.9% and Claude Sonnet 4.6 at 41.7%.
In the source
51.9%
Where to find it
p. 228, section 8.14.1 HealthBench Professional; Figure 8.14.A p. 229
How it was run
length-adjusted; adaptive thinking at max effort; Claude Sonnet 4.6 grader; 5 trials; comparison model in the Opus 4.8 card
Reported by
vendor-reported
Document
[15] System Card: Claude Opus 4.8
31GPT-5.5OpenAI · length-adjusted0.518Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
51.8
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.5
How it was run
length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.5 (57.2 unadjusted, 3818 chars)
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
32Gemini 3.8 FlashGoogle · Artificial Analysis…0.517IndependentOct 2026[16]
HealthBench Professional: Length-Adjusted Score | Gemini 3.8 Flash (high) | 51.7%
In the source
51.7%
Where to find it
Score section, Length-Adjusted Score chart, Gemini 3.8 Flash (high) bar label; raw score 59.0% on the Raw Score tab
How it was run
Artificial Analysis independent run, high reasoning; GPT-5.4 (low) judge, 4 repeats per task; length-adjusted (59.0 raw, 4,494 mean response chars)
Reported by
independent run
Document
[16] HealthBench Professional Benchmark | Artificial Analysis
33DeepSeek V4.1 FlashDeepSeek · Artificial Analysis…0.507IndependentOct 2026[16]
HealthBench Professional: Length-Adjusted Score | DeepSeek V4.1 Flash (max) | 50.7%
In the source
50.7%
Where to find it
Score section, Length-Adjusted Score chart, DeepSeek V4.1 Flash (max) bar label; raw score 52.9% on the Raw Score tab
How it was run
Artificial Analysis independent run, max reasoning; GPT-5.4 (low) judge, 4 repeats per task; length-adjusted (52.9 raw, 2,760 mean response chars)
Reported by
independent run
Document
[16] HealthBench Professional Benchmark | Artificial Analysis
34Grok 4.6SpaceX AI · High effort0.485Model's makerSep 2026[14]
Grok 4.7 xHigh Grok 4.6 High GPT-5.6 Sol Max Fable 5.1 Max Clinical reasoningHealthBench Professional 56.7% 48.5% 60.5% 62.1%
In the source
48.5%
Where to find it
Model Improvements comparison table, Clinical reasoning / HealthBench Professional row; Grok 4.6 column; September 21, 2026.
How it was run
SpaceXAI evaluation; High effort; HealthBench Professional
Reported by
vendor-reported
Document
[14] Introducing Grok 4.7
35GPT-5.4OpenAI · length-adjusted0.481Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
48.1
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.4
How it was run
length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.4 (51.9 unadjusted, 3308 chars)
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
36GPT-5OpenAI · length-adjusted0.462Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
46.2
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5
How it was run
length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5 (51.0 unadjusted, 3616 chars)
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
37GPT-5.2OpenAI · length-adjusted0.459Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
45.9
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.2
How it was run
length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.2 (50.0 unadjusted, 3400 chars)
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
38Claude Sonnet 4.6Anthropic · length-adjusted0.442Model's makerJun 2026[13]
HealthBench Professional 57.8 44.2 51.8 -
In the source
44.2
Where to find it
p. 115, Table 8.1.A, row HealthBench Professional, column Claude Sonnet 4.6
How it was run
length-adjusted; Anthropic protocol: adaptive thinking at max effort, Claude Opus 4.8 grader, averaged over 5 trials, no tools or custom system prompt; comparison column in the Sonnet 5 card. Other Anthropic prints: 44.4% (Fable card Figure 8.18.2.A), 41.7% (Opus 4.8 card, Sonnet 4.6 grader)
Reported by
vendor-reported
Document
[13] System Card: Claude Sonnet 5
39GPT-5.6 Luna (August)OpenAI · ChatGPT production0.441Model's makerAug 2026[17]
HealthBench Professional 32.9 (33.8, 2,285) 38.4 (40.7, 2,775) 54.0 (56.6, 2,894) 44.1 (46.8, 2,920)
In the source
44.1
Where to find it
p. 11, section 5.1 HealthBench, table "Reported as length-adjusted score (unadjusted, mean response length in characters)", column GPT-5.6 Luna (August)
How it was run
ChatGPT production (Instant) deployment setting, length-adjusted, GPT-5.6 August Updates PDF column GPT-5.6 Luna (August) (46.8 unadjusted, 2,920 chars)
Reported by
vendor-reported
Document
[17] GPT-5.6 - August Updates (system card addendum)
40GPT-5.1OpenAI · length-adjusted0.396Model's makerJun 2026[10]
HealthBench Professional length-adjusted 46.2 (51.0, 3616) 39.6 (48.0, 4863) 45.9 (50.0, 3400) 48.1 (51.9, 3308) 51.8 (57.2, 3818) 60.5 (64.1, 3228) 57.7 (62.4, 3618) 55.7 (59.8, 3389)
In the source
39.6
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.1
How it was run
length-adjusted, max reasoning effort, GPT-5.6 system card Table 6 column GPT-5.1 (48.0 unadjusted, 4863 chars)
Reported by
vendor-reported
Document
[10] GPT-5.6 System Card
41GPT-5.5 InstantOpenAI · length-adjusted0.384Model's makerMay 2026[18]
HealthBench Professional 37.6 (40.4, 2,973) 35.7 (38.3, 2,872) 32.9 (33.8, 2,285) 38.4 (40.7, 2,775)
In the source
38.4
Where to find it
Section 4.1 HealthBench, Table 5 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.5 INSTANT
How it was run
length-adjusted (40.7 unadjusted, 2,775 mean response chars); GPT-5.5 Instant system card Table 5, column GPT-5.5 INSTANT.
Reported by
vendor-reported
Document
[18] GPT-5.5 Instant System Card
42MAI-Thinking-1Microsoft · length-adjusted0.350Model's makerAug 2026[19]
Model AIR-Bench CyberSec Instruct CyberSec Auto Long Fact Truthful QA HealthBench Prof. MedXpert QA MAI-Thinking-1 88 63 63 98 88 35 43 Sonnet 4.6 88 62 56 98 88 38 49
In the source
35
Where to find it
p. 54, Table 12 'Post-trained model evaluation results on various public benchmarks', Health group, column HealthBench Prof.; protocol Appendix K.6 p. 106: HealthBench Professional introduces a length penalty for the primary metric, to correct for a well-observed correlation between lengthy responses and artificially increased LLM-grader scores. For all reported scores, we use the standard GPT-5.4 grader and rubrics provided by OpenAI.
How it was run
length-adjusted (HealthBench Professional length penalty), standard GPT-5.4 grader and OpenAI rubrics, Microsoft AI run; printed at integer precision.
Reported by
vendor-reported
Document
[19] MAI-Thinking-1: Building a Hill-Climbing Machine
43Claude Haiku 4.5Anthropic · extended thinking0.322Model's makerOct 2026[4]
Evaluation | Haiku 5.5 | Haiku 4.5 | Sonnet 5.5 | GPT-6 Luna HealthBench Professional (Length-adjusted) | 64.8 | 32.2 | 69.2 | –
In the source
32.2
Where to find it
p. 111, Table 8.1.A, row HealthBench Professional (Length-adjusted), column Haiku 4.5
How it was run
Anthropic; extended thinking (Haiku 4.5 has no effort setting); Opus 4.8 grader; safety classifiers enabled; length-adjusted; measured for the Haiku 5.5 card
Reported by
vendor-reported
Document
[4] Claude Haiku 5.5 System Card
Sources

The 19 documents behind these scores

  1. 1.Claude Sonnet 5.5 System Cardsystem card, Anthropic, 28 Sep 2026. Quotes from this document
  2. 2.Claude Fable 5 and Claude Mythos 5 System Cardsystem card, Anthropic, 9 Jun 2026. Quotes from this document
  3. 3.Claude Opus 5.5 System Cardsystem card, Anthropic, 22 Sep 2026. Quotes from this document
  4. 4.Claude Haiku 5.5 System Cardsystem card, Anthropic, 7 Oct 2026. Quotes from this document
  5. 5.GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)system card, OpenAI, 3 Sep 2026. Quotes from this document
  6. 6.Addendum to GPT-6 Astra System Card: GPT-6.1 Solsystem card, OpenAI, 29 Sep 2026. Quotes from this document
  7. 7.Claude Fable 5.1 and Claude Mythos 5.1 System Cardsystem card, Anthropic, 1 Sep 2026. Quotes from this document
  8. 8.Addendum to GPT-6 Astra System Card: GPT-6.1 Terrasystem card, OpenAI, 26 Oct 2026. Quotes from this document
  9. 9.Muse Wonder Model Cardmodel card, Meta, 12 Oct 2026. Quotes from this document
  10. 10.GPT-5.6 System Cardsystem card, OpenAI, 9 Jul 2026. Quotes from this document
  11. 11.System Card: Claude Opus 5system card, Anthropic, 24 Jul 2026. Quotes from this document
  12. 12.Muse Spark 1.1 Evaluation Reportmodel card, Meta, 9 Jul 2026. Quotes from this document
  13. 13.System Card: Claude Sonnet 5system card, Anthropic, 30 Jun 2026. Quotes from this document
  14. 14.Introducing Grok 4.7launch post, SpaceXAI, 21 Sep 2026. Quotes from this document
  15. 15.System Card: Claude Opus 4.8system card, Anthropic, 28 May 2026. Quotes from this document
  16. 16.HealthBench Professional Benchmark | Artificial Analysisindependent leaderboard, Artificial Analysis, 2 Nov 2026. Quotes from this document
  17. 17.GPT-5.6 - August Updates (system card addendum)system card, OpenAI, 6 Aug 2026. Quotes from this document
  18. 18.GPT-5.5 Instant System Cardsystem card, OpenAI, 5 May 2026. Quotes from this document
  19. 19.MAI-Thinking-1: Building a Hill-Climbing Machinemodel card, Microsoft AI, 12 Aug 2026. Quotes from this document