Health AI Benchmarks

HealthBench

When someone asks an AI about their own health, do the answers meet a physician's standard?

More about this benchmark

Realistic multi-turn health conversations graded on physician-written rubrics for accuracy, completeness, and communication.

5,000 realistic multi-turn health conversations graded against physician-written rubrics (48,562 criteria) covering accuracy, completeness, context awareness, communication, and instruction following. OpenAI now also reports a length-adjusted variant that penalizes verbosity.

Each item
Conversation with an example-specific rubric
The model gets
A health conversation ending at the next response to be evaluated.
It returns
An open-ended assistant response.
Setting
Original 5,000-conversation benchmark; physician criteria graded by a model.
Released
May 2025
Models on the official board
8
Paper version
Original HealthBench; May 2025 paper v1

Reading these numbers

Cross-source rows are not directly comparable: Anthropic reports a raw score under its own protocol, OpenAI leads with length-adjusted numbers at maximum reasoning effort and reports production Instant settings separately, and Baichuan ran its competitors itself. The benchmark's official grader is GPT-4.1. OpenAI has said the parent set is approaching a noise ceiling for frontier models and points to HealthBench Professional for continued measurement.

Scores come from a mix of vendor documents and official results.

Results
33 from 26 models
Latest
Oct 2026
Published by
OpenAI
Index weight
0.798, core
Sources
17

Counts toward the Health Evals Index as a core benchmark, at weight 0.798: realistic health conversations with people, graded on physician rubrics.

Where every published HealthBench score sits

33 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

33 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.

1Claude Opus 5.5Anthropic · raw/unadjusted68.1Model's makerSep 2026[1]
On HealthBench, Claude Opus 5.5 achieved a raw score of 68.1%, ahead of Claude Fable 5.1 at 66.7%, Claude Sonnet 5 at 59.2%, and Claude Opus 5 at 67.1%. After length adjustment, which penalizes verbose model responses, Claude Opus 5.5 achieved a score of 60.6%.
In the source
68.1%
Where to find it
p. 213, section 8.15.1 HealthBench results; Figure 8.15.1.A
How it was run
raw/unadjusted, Anthropic system-card protocol (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 60.6
Reported by
vendor-reported
Document
[1] Claude Opus 5.5 System Card
2Claude Opus 5Anthropic · raw/unadjusted67.1Model's makerJul 2026[2]
Claude Opus 5 achieved a raw score of 67.1%, which is the highest among all Claude models, ahead of Claude Mythos 5 at 62.5%, Claude Opus 4.8 at 58.8%, and Claude Sonnet 5 at 59.2%. After length adjustment, which penalizes verbose model responses, Claude Opus 5 achieved a score of 57.8%.
In the source
67.1
Where to find it
p. 188, section 8.15.1 HealthBench results; Figure 8.15.1.A
How it was run
raw/unadjusted, Anthropic system-card protocol (max effort, no tools, five trials); length-adjusted 57.8; mirrored on benchlm.ai
Reported by
vendor-reported
Document
[2] System Card: Claude Opus 5
3Claude Sonnet 5.5Anthropic · max effort65.4Model's makerSep 2026[3]
Sonnet 5.5 | HealthBench | max 65.4%
In the source
65.4%
Where to find it
p. 138, Figure 8.15.B, left panel HealthBench; orange Sonnet 5.5 max-effort label (visually read printed labels).
How it was run
Anthropic; max effort; Opus 4.8 grader; safety classifiers enabled; length-adjusted; raw 69.4%
Reported by
vendor-reported
Document
[3] Claude Sonnet 5.5 System Card
4Baichuan-M3Baichuan · self-run in Baichuan-M3…65.1Model's makerFeb 2026[4]
On the comprehensive HealthBench Total, Baichuan-M3 achieves a score of 65.1, surpassing the runner-up GPT-5.2-High (63.3) by a clear margin.
In the source
65.1
Where to find it
p. 23, section 4.2.1 HealthBench-Main; Table on p. 25 (Model / HealthBench Score)
How it was run
self-run in Baichuan-M3 paper (arXiv 2602.06570)
Reported by
vendor-reported
Document
[4] Baichuan-M3 Technical Report
5GPT-5.2-HighOpenAI · raw score as run by…63.3IndependentFeb 2026[4]
GPT-5.2-High 63.3 2.37% 2.78%
In the source
63.3
Where to find it
p. 25, HealthBench-Hallu table (Model / HealthBench Score column); also p. 23 prose
How it was run
raw score as run by Baichuan in the M3 technical report, not an OpenAI-reported number; OpenAI own GPT-5.2 figure is 56.8 length-adjusted (60.7 unadjusted) in the GPT-5.6 system card.
Reported by
independent run
Document
[4] Baichuan-M3 Technical Report
6Claude Fable 5Anthropic · length-adjusted62.7Model's makerJun 2026[5]
HealthBench 62.7 - 61.1 59.3 56.5 -
In the source
62.7
Where to find it
p. 252, Table 8.1.A, row HealthBench, column Mythos 5; Figure 8.18.1.A p. 297
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 (40k token budget); printed in the Mythos 5 column of Table 8.1.A (Fable 5 column '-'); raw 62.5% per Opus 5 card
Reported by
vendor-reported
Document
[5] Claude Fable 5 and Claude Mythos 5 System Card
7Claude Haiku 5.5Anthropic · max effort61.6Model's makerOct 2026[6]
On HealthBench (one run per level below max), Haiku 5.5 scored 59.9% at low, 60.2% at medium, 60.3% at high, 61.1% at xhigh and 61.6% at max, a rise of 1.8 points against 6.9 on HealthBench Professional
In the source
61.6%
Where to find it
p. 133, Section 8.11.3; p. 132, Figure 8.11.1.A HealthBench panel labels 'Haiku 5.5 61.6%' (length-adjusted, solid) and 'raw 64.2%'
How it was run
Anthropic; max effort; Opus 4.8 grader; safety classifiers enabled; length-adjusted; raw 64.2%; mean of five runs made on October 2
Reported by
vendor-reported
Document
[6] Claude Haiku 5.5 System Card
8Claude Opus 5.5Anthropic · adaptive max effort60.6Model's makerSep 2026[1]
8.15.1 HealthBench results After length adjustment, which penalizes verbose model responses, Claude Opus 5.5 achieved a score of 60.6%.
In the source
60.6%
Where to find it
p. 213, Section 8.15.1, paragraph and Figure 8.15.1.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 68.1%
Reported by
vendor-reported
Document
[1] Claude Opus 5.5 System Card
9Claude Fable 5Anthropic · length-adjusted60.4Model's makerSep 2026[7]
HealthBench | Length-adjusted score | Claude Fable 5 | 60.4%
In the source
60.4%
Where to find it
p. 198, Figure 8.17.1.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 61.2%
Reported by
vendor-reported
Document
[7] Claude Fable 5.1 and Claude Mythos 5.1 System Card
10Claude Fable 5.1Anthropic · length-adjusted60Model's makerSep 2026[7]
On HealthBench, Claude Fable 5.1 achieved a raw score of 66.7%, ahead of Claude Fable 5 at 61.2% and Claude Sonnet 5 at 59.2%, and behind Claude Opus 5 at 67.1%. After length adjustment, which penalizes verbose model responses, Fable 5.1 achieved a score of 60%.
In the source
60%
Where to find it
p. 198, sec. 8.17.1 / Figure 8.17.1.A
How it was run
length-adjusted (method published in OpenAI's GPT-5.5 System Card); 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 66.7%).
Reported by
vendor-reported
Document
[7] Claude Fable 5.1 and Claude Mythos 5.1 System Card
11Claude Opus 4.8Anthropic · length-adjusted59.3Model's makerJun 2026[5]
HealthBench 62.7 - 61.1 59.3 56.5 -
In the source
59.3
Where to find it
p. 252, Table 8.1.A, row HealthBench, column Opus 4.8
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; comparison column in the Fable/Mythos 5 card; raw 58.8% per Opus 5 card; not in the Opus 4.8 card itself
Reported by
vendor-reported
Document
[5] Claude Fable 5 and Claude Mythos 5 System Card
12Claude Sonnet 5Anthropic · length-adjusted58.7Model's makerJun 2026[8]
[Figure 8.12.1.A] HealthBench length-adjusted scores. All Claude models used adaptive thinking at max effort. (bar label: Claude Sonnet 5 58.7%)
In the source
58.7%
Where to find it
p. 138, section 8.12.1 HealthBench results, Figure 8.12.1.A bar label (no prose or table number)
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; figure-only in the Sonnet 5 card; raw 59.2% per Opus 5 card
Reported by
vendor-reported
Document
[8] System Card: Claude Sonnet 5
13GPT-6.1 SolOpenAI · length-adjusted58.5Model's makerSep 2026[9]
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 length-adjusted | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930) | 58.3 (56.9, 1760) | 53.2 (47.1, 977) | 54.5 (50.0, 1255) | 58.5 (56.7, 1701)
In the source
58.5
Where to find it
Section 5.1, Table 7, HealthBench length-adjusted, GPT-6.1 Sol column
How it was run
OpenAI; length-adjusted; 58.5 (56.7, 1701) (adjusted, raw, mean response characters); addendum Table 7 refers to the GPT-6 Astra card for method
Reported by
vendor-reported
Document
[9] Addendum to GPT-6 Astra System Card: GPT-6.1 Sol
14GPT-6 AstraOpenAI · length-adjusted58.3Model's makerSep 2026[10]
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 length-adjusted | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930) | 58.3 (56.9, 1760) | 53.2 (47.1, 977) | 54.5 (50.0, 1255)
In the source
58.3
Where to find it
Section 11.4.1, Table 29, HealthBench length-adjusted, GPT-6 Astra column; September 22 correction.
How it was run
length-adjusted, max reasoning effort (59.7 unadjusted, 2,258 mean response chars); GPT-6 Astra system card Table 6, column 'gpt-6 Astra'.
Reported by
vendor-reported
Document
[10] GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)
15Claude Opus 5 (length-adjusted)Anthropic · length-adjusted57.8Model's makerJul 2026[2]
8.15.1 HealthBench results After length adjustment, which penalizes verbose model responses, Claude Opus 5 achieved a score of 57.8%.
In the source
57.8%
Where to find it
p. 188, Section 8.15.1, paragraph and Figure 8.15.1.A.
How it was run
Anthropic; length-adjusted; adaptive max effort; Opus 4.8 grader; five trials; no tools or custom system prompt; raw 67.1%
Reported by
vendor-reported
Document
[2] System Card: Claude Opus 5
16GPT-5OpenAI · length-adjusted57.7Model's makerJun 2026[11]
evaluation | gpt-5 | gpt-5.1 | gpt-5.2 | gpt-5.4 | gpt-5.5 | gpt-5.6-sol | gpt-5.6-terra | gpt-5.6-luna HealthBench length-adjusted | 57.7 (63.1, 2904) | 50.9 (64.2, 4222) | 56.8 (60.7, 2645) | 54.0 (55.7, 2275) | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930)
In the source
57.7
Where to find it
Section 6.1, Table 6, HealthBench length-adjusted row, gpt-5 column.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 57.7 (63.1, 2904) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
17GPT OSS 120BOpenAI · reasoning level high57.6Model's makerAug 2025[12]
HealthBench 53.0 55.9 57.6 40.4 41.8 42.5
In the source
57.6
Where to find it
Section 2, Table 3: Evaluations across multiple benchmarks and reasoning levels, row HealthBench, column gpt-oss-120b high
How it was run
reasoning level high, raw score (%), gpt-oss model card Table 3 (low 53.0, medium 55.9)
Reported by
vendor-reported
Document
[12] gpt-oss-120b & gpt-oss-20b Model Card
18GPT-5.6 SolOpenAI · length-adjusted57.0Model's makerJun 2026[11]
HealthBench length-adjusted 57.7 (63.1, 2904) 50.9 (64.2, 4222) 56.8 (60.7, 2645) 54.0 (55.7, 2275) 56.5 (58.4, 2313) 57.0 (55.6, 1764) 57.0 (58.7, 2285) 55.8 (55.4, 1930)
In the source
57.0
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 (55.6 unadjusted), GPT-5.6 system card 2026-07-09
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
18GPT-5.6 TerraOpenAI · length-adjusted57.0Model's makerJun 2026[11]
HealthBench length-adjusted 57.7 (63.1, 2904) 50.9 (64.2, 4222) 56.8 (60.7, 2645) 54.0 (55.7, 2275) 56.5 (58.4, 2313) 57.0 (55.6, 1764) 57.0 (58.7, 2285) 55.8 (55.4, 1930)
In the source
57.0
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 (58.7 unadjusted), max reasoning effort
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
20GPT-5.2OpenAI · length-adjusted56.8Model's makerJun 2026[11]
evaluation | gpt-5 | gpt-5.1 | gpt-5.2 | gpt-5.4 | gpt-5.5 | gpt-5.6-sol | gpt-5.6-terra | gpt-5.6-luna HealthBench length-adjusted | 57.7 (63.1, 2904) | 50.9 (64.2, 4222) | 56.8 (60.7, 2645) | 54.0 (55.7, 2275) | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930)
In the source
56.8
Where to find it
Section 6.1, Table 6, HealthBench length-adjusted row, gpt-5.2 column.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 56.8 (60.7, 2645) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
21GPT-5.5OpenAI · length-adjusted56.5Model's makerApr 2026[13]
HealthBench length-adjusted 57.7 (63.1, 2904) 50.9 (64.2, 4222) 56.8 (60.7, 2645) 54.0 (55.7, 2275) 56.5 (58.4, 2313)
In the source
56.5
Where to find it
Section 5 Health, Table 7 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.5
How it was run
length-adjusted (58.4 unadjusted), comparison row in GPT-5.6 system card
Reported by
vendor-reported
Document
[13] GPT-5.5 System Card
22GPT-6 Luna (Anthropic run)OpenAI · Run by Anthropic through…56.3IndependentOct 2026[6]
HealthBench panel: 56.3% (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
56.3%
Where to find it
p. 132, Figure 8.11.1.A, HealthBench 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
[6] Claude Haiku 5.5 System Card
23GPT-5.6 LunaOpenAI · length-adjusted55.8Model's makerJun 2026[11]
HealthBench length-adjusted 57.7 (63.1, 2904) 50.9 (64.2, 4222) 56.8 (60.7, 2645) 54.0 (55.7, 2275) 56.5 (58.4, 2313) 57.0 (55.6, 1764) 57.0 (58.7, 2285) 55.8 (55.4, 1930)
In the source
55.8
Where to find it
Section 5.1 HealthBench, Table 6 (reported as length-adjusted score (unadjusted, mean response length in characters)), column GPT-5.6-LUNA
How it was run
length-adjusted (55.4 unadjusted), max reasoning effort
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
24GPT-6.1 TerraOpenAI · length-adjusted55.1Model's makerOct 2026[14]
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 length-adjusted 56.5 (58.4, 2313) 57.0 (55.6, 1764) 57.0 (58.7, 2285) 55.8 (55.4, 1930) 58.3 (56.9, 1760) 53.2 (47.1, 977) 54.5 (50.0, 1255) 58.5 (56.7, 1701) 55.1 (51.1, 1331)
In the source
55.1
Where to find it
Section 5.1, Table 5, HealthBench length-adjusted, GPT-6.1 Terra column
How it was run
OpenAI; length-adjusted; 55.1 (51.1, 1331) (adjusted, raw, mean response characters); addendum Table 5 refers to the GPT-6 Astra card for method
Reported by
vendor-reported
Document
[14] Addendum to GPT-6 Astra System Card: GPT-6.1 Terra
25GPT-5.6 Sol (August)OpenAI · ChatGPT production/Instant…55.0Model's makerAug 2026[15]
HealthBench 49.6 (47.9, 1,724) 51.4 (50.9, 1,922) 55.0 (52.1, 1,514) 53.3 (50.7, 1,567)
In the source
55.0
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 (52.1 unadjusted), GPT-5.6 August Updates PDF 2026-08-06
Reported by
vendor-reported
Document
[15] GPT-5.6 - August Updates (system card addendum)
26GPT-6 LunaOpenAI · length-adjusted54.5Model's makerSep 2026[10]
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 length-adjusted | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930) | 58.3 (56.9, 1760) | 53.2 (47.1, 977) | 54.5 (50.0, 1255)
In the source
54.5
Where to find it
Section 11.4.1, Table 29, HealthBench length-adjusted, GPT-6 Luna column; September 22 correction.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 54.5 (50.0, 1255) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[10] GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)
27GPT-5.3 ChatOpenAI · raw score54.1Model's makerMar 2026[16]
HealthBench 55.4% 54.1%
In the source
54.1%
Where to find it
Section 4.1 HealthBench, Table 3: HealthBench, row HealthBench, column GPT-5.3-INSTANT
How it was run
raw score (no length adjustment), GPT-5.3 Instant system card Table 3 column GPT-5.3-INSTANT; later OpenAI cards print 49.6 length-adjusted (47.9 unadjusted)
Reported by
vendor-reported
Document
[16] GPT-5.3 Instant System Card
28GPT-5.4OpenAI · length-adjusted54.0Model's makerJun 2026[11]
evaluation | gpt-5 | gpt-5.1 | gpt-5.2 | gpt-5.4 | gpt-5.5 | gpt-5.6-sol | gpt-5.6-terra | gpt-5.6-luna HealthBench length-adjusted | 57.7 (63.1, 2904) | 50.9 (64.2, 4222) | 56.8 (60.7, 2645) | 54.0 (55.7, 2275) | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930)
In the source
54.0
Where to find it
Section 6.1, Table 6, HealthBench length-adjusted row, gpt-5.4 column.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 54.0 (55.7, 2275) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
29GPT-5.6 Luna (August)OpenAI · ChatGPT production53.3Model's makerAug 2026[15]
HealthBench 49.6 (47.9, 1,724) 51.4 (50.9, 1,922) 55.0 (52.1, 1,514) 53.3 (50.7, 1,567)
In the source
53.3
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) (50.7 unadjusted, 1,567 chars)
Reported by
vendor-reported
Document
[15] GPT-5.6 - August Updates (system card addendum)
30GPT-6 SolOpenAI · length-adjusted53.2Model's makerSep 2026[10]
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 length-adjusted | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930) | 58.3 (56.9, 1760) | 53.2 (47.1, 977) | 54.5 (50.0, 1255)
In the source
53.2
Where to find it
Section 11.4.1, Table 29, HealthBench length-adjusted, GPT-6 Sol column; September 22 correction.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 53.2 (47.1, 977) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[10] GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)
31GPT-5.5 InstantOpenAI · length-adjusted51.4Model's makerMay 2026[17]
HealthBench 49.6 (50.8, 2,208) 50.6 (51.5, 2,145) 49.6 (47.9, 1,724) 51.4 (50.9, 1,922)
In the source
51.4
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, GPT-5.5 Instant system card Table 5 column GPT-5.5 INSTANT (50.9 unadjusted, 1,922 chars); same number in GPT-5.6 August Updates p. 11
Reported by
vendor-reported
Document
[17] GPT-5.5 Instant System Card
32GPT-5.1OpenAI · length-adjusted50.9Model's makerJun 2026[11]
evaluation | gpt-5 | gpt-5.1 | gpt-5.2 | gpt-5.4 | gpt-5.5 | gpt-5.6-sol | gpt-5.6-terra | gpt-5.6-luna HealthBench length-adjusted | 57.7 (63.1, 2904) | 50.9 (64.2, 4222) | 56.8 (60.7, 2645) | 54.0 (55.7, 2275) | 56.5 (58.4, 2313) | 57.0 (55.6, 1764) | 57.0 (58.7, 2285) | 55.8 (55.4, 1930)
In the source
50.9
Where to find it
Section 6.1, Table 6, HealthBench length-adjusted row, gpt-5.1 column.
How it was run
OpenAI; length-adjusted; maximum reasoning effort; 50.9 (64.2, 4222) (adjusted, raw, mean response characters)
Reported by
vendor-reported
Document
[11] GPT-5.6 System Card
33GPT OSS 20BOpenAI · reasoning level high42.5Model's makerAug 2025[12]
HealthBench 53.0 55.9 57.6 40.4 41.8 42.5
In the source
42.5
Where to find it
Section 2, Table 3: Evaluations across multiple benchmarks and reasoning levels, row HealthBench, column gpt-oss-20b high
How it was run
reasoning level high, raw score (%), gpt-oss model card Table 3 (low 40.4, medium 41.8)
Reported by
vendor-reported
Document
[12] gpt-oss-120b & gpt-oss-20b Model Card
Sources

The 17 documents behind these scores

  1. 1.Claude Opus 5.5 System Cardsystem card, Anthropic, 22 Sep 2026. Quotes from this document
  2. 2.System Card: Claude Opus 5system card, Anthropic, 24 Jul 2026. Quotes from this document
  3. 3.Claude Sonnet 5.5 System Cardsystem card, Anthropic, 28 Sep 2026. Quotes from this document
  4. 4.Baichuan-M3 Technical Reportpaper, Baichuan, 6 Feb 2026. Quotes from this document
  5. 5.Claude Fable 5 and Claude Mythos 5 System Cardsystem card, Anthropic, 9 Jun 2026. Quotes from this document
  6. 6.Claude Haiku 5.5 System Cardsystem card, Anthropic, 7 Oct 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.System Card: Claude Sonnet 5system card, Anthropic, 30 Jun 2026. Quotes from this document
  9. 9.Addendum to GPT-6 Astra System Card: GPT-6.1 Solsystem card, OpenAI, 29 Sep 2026. Quotes from this document
  10. 10.GPT-6 Astra System Card - HealthBench (Deployment Safety Hub)system card, OpenAI, 3 Sep 2026. Quotes from this document
  11. 11.GPT-5.6 System Cardsystem card, OpenAI, 9 Jul 2026. Quotes from this document
  12. 12.gpt-oss-120b & gpt-oss-20b Model Cardmodel card, OpenAI, 5 Aug 2025. Quotes from this document
  13. 13.GPT-5.5 System Cardsystem card, OpenAI, 23 Apr 2026. Quotes from this document
  14. 14.Addendum to GPT-6 Astra System Card: GPT-6.1 Terrasystem card, OpenAI, 26 Oct 2026. Quotes from this document
  15. 15.GPT-5.6 - August Updates (system card addendum)system card, OpenAI, 6 Aug 2026. Quotes from this document
  16. 16.GPT-5.3 Instant System Cardsystem card, OpenAI, 2 Mar 2026. Quotes from this document
  17. 17.GPT-5.5 Instant System Cardsystem card, OpenAI, 5 May 2026. Quotes from this document