Share promptsSign inCreate team library
← All benchmarks

Health & medical AI benchmarks

Physician-graded health conversations, medical coding and exam questions. 3 public leaderboards below.

HealthBench (independent run)

5,000 realistic health conversations, each graded against a rubric written by physicians; HealthBench Professional covers clinician use cases.

Run by Artificial Analysis (benchmark by OpenAI). Leaderboard date not stated · retrieved 21 Oct 2026.

  • Anthropic
  • OpenAI
  • Google
  • xAI
  • DeepSeek
  • Z.ai
  • Moonshot
HealthBench (independent run): HealthBench Professional vs HealthBenchScatter plot of 14 models. Leader on HealthBench Professional: GPT-6 Astra (max) at 66.0%. Full values in the table below.45%50%55%60%65%45%50%55%60%65%70%HealthBench →↑ HealthBench ProfessionalGemini 3.8 Flash (high) (Google): HealthBench 48.4%, HealthBench Professional 51.7%GGemini 4 Argon (high) (Google): HealthBench 49.1%, HealthBench Professional 55.1%GGPT-6 Sol (max) (OpenAI): HealthBench 54.2%, HealthBench Professional 62.3%OGLM-5.3 (max) (Z.ai): HealthBench 54.5%, HealthBench Professional 52.8%ZGPT-6 Luna (max) (OpenAI): HealthBench 54.6%, HealthBench Professional 61.9%OKimi K3 (max) (Moonshot): HealthBench 54.7%, HealthBench Professional 53.2%KGrok 4.7 (xhigh) (xAI): HealthBench 56.0%, HealthBench Professional 55.8%XClaude Opus 5.5 (max with fallback) (Anthropic): HealthBench 57.0%, HealthBench Professional 59.9%AClaude Fable 5.1 (max with fallback) (Anthropic): HealthBench 57.2%, HealthBench Professional 57.8%AGPT-6 Astra (max) (OpenAI): HealthBench 58.2%, HealthBench Professional 66.0%OClaude Haiku 5.5 (max) (Anthropic): HealthBench 58.5%, HealthBench Professional 58.4%AGPT-6.1 Sol (max) (OpenAI): HealthBench 58.5%, HealthBench Professional 65.1%ODeepSeek V4.1 Flash (max) (DeepSeek): HealthBench 60.9%, HealthBench Professional 50.7%DClaude Sonnet 5.5 (max with fallback) (Anthropic): HealthBench 62.1%, HealthBench Professional 63.0%AClaude Sonnet 5.5GPT-6 AstraGPT-6.1 SolDeepSeek V4.1 Flash
HealthBench (independent run): Professional (raw) vs HealthBench (raw)Scatter plot of 14 models. Leader on Professional (raw): Claude Opus 5.5 (max with fallback) at 71.4%. Full values in the table below.45%50%55%60%65%70%50%55%60%65%70%75%HealthBench (raw) →↑ Professional (raw)Gemini 3.8 Flash (high) (Google): HealthBench (raw) 57.1%, Professional (raw) 59.0%GGemini 4 Argon (high) (Google): HealthBench (raw) 60.0%, Professional (raw) 64.2%GGPT-6 Sol (max) (OpenAI): HealthBench (raw) 48.2%, Professional (raw) 61.2%OGLM-5.3 (max) (Z.ai): HealthBench (raw) 54.7%, Professional (raw) 57.9%ZGPT-6 Luna (max) (OpenAI): HealthBench (raw) 50.5%, Professional (raw) 62.7%OKimi K3 (max) (Moonshot): HealthBench (raw) 54.2%, Professional (raw) 56.9%KGrok 4.7 (xhigh) (xAI): HealthBench (raw) 54.2%, Professional (raw) 57.4%XClaude Opus 5.5 (max with fallback) (Anthropic): HealthBench (raw) 64.3%, Professional (raw) 71.4%AClaude Fable 5.1 (max with fallback) (Anthropic): HealthBench (raw) 63.9%, Professional (raw) 69.2%AGPT-6 Astra (max) (OpenAI): HealthBench (raw) 56.9%, Professional (raw) 69.7%OClaude Haiku 5.5 (max) (Anthropic): HealthBench (raw) 61.2%, Professional (raw) 64.8%AGPT-6.1 Sol (max) (OpenAI): HealthBench (raw) 56.9%, Professional (raw) 68.3%ODeepSeek V4.1 Flash (max) (DeepSeek): HealthBench (raw) 61.9%, Professional (raw) 52.9%DClaude Sonnet 5.5 (max with fallback) (Anthropic): HealthBench (raw) 65.9%, Professional (raw) 70.8%AClaude Sonnet 5.5Claude Opus 5.5GPT-6 Astra
HealthBench
Share of physician-rubric points earned, adjusted for answer length (long answers lose points), as OpenAI reports it.
HealthBench Professional
Same scoring on clinician-facing tasks (consults, writing, research), length-adjusted.
HealthBench (raw)
Share of rubric points earned with no length adjustment.
Professional (raw)
HealthBench Professional with no length adjustment.
Mean answer length (characters)
Average reply length, excluding reasoning.
HealthBench (independent run) scores
RankModelProviderHealthBenchHealthBench ProfessionalHealthBench (raw)Professional (raw)Mean answer length (characters)
1 Claude Sonnet 5.5 (max with fallback)Anthropic62.1%63.0%65.9%70.8%2,634
2 DeepSeek V4.1 Flash (max)DeepSeek60.9%50.7%61.9%52.9%2,164
3 GPT-6.1 Sol (max)OpenAI58.5%65.1%56.9%68.3%1,732
4 Claude Haiku 5.5 (max)Anthropic58.5%58.4%61.2%64.8%2,457
5 GPT-6 Astra (max)OpenAI58.2%66.0%56.9%69.7%1,780
6 Claude Fable 5.1 (max with fallback)Anthropic57.2%57.8%63.9%69.2%3,113
7 Claude Opus 5.5 (max with fallback)Anthropic57.0%59.9%64.3%71.4%3,217
8 Grok 4.7 (xhigh)xAI56.0%55.8%54.2%57.4%1,691
9 Kimi K3 (max)Moonshot54.7%53.2%54.2%56.9%1,919
10 GPT-6 Luna (max)OpenAI54.6%61.9%50.5%62.7%1,325
Show 4 more rows
RankModelProviderHealthBenchHealthBench ProfessionalHealthBench (raw)Professional (raw)Mean answer length (characters)
11 GLM-5.3 (max)Z.ai54.5%52.8%54.7%57.9%2,036
12 GPT-6 Sol (max)OpenAI54.2%62.3%48.2%61.2%1,008
13 Gemini 4 Argon (high)Google49.1%55.1%60.0%64.2%3,837
14 Gemini 3.8 Flash (high)Google48.4%51.7%57.1%59.0%3,467

Source: Artificial Analysis leaderboard ↗ · About HealthBench (OpenAI) ↗

Top model here: Claude Sonnet 5.5 (max with fallback) (Anthropic). Try prompts for Claude →

MedCode

Assigning the correct medical billing codes from clinical documentation.

Run by Vals AI. Leaderboard as of 13 Oct 2026 · retrieved 21 Oct 2026. Showing top 25 of 107 models listed.

  • Anthropic
  • OpenAI
  • Google
  • xAI
  • Meta
  • Moonshot
MedCode: top 15 by AccuracyRanked bars. Leader: Claude Opus 5 at 63.57%. Full values in the table below.0%20%40%60%80%100%Claude Opus 5 (Anthropic): 63.57%AClaude Opus 563.57%Gemini 3.1 Pro Preview (02/26) (Google): 59.06%GGemini 3.1 Pro Preview (02/26)59.06%Gemini 4 Argon (Google): 58.80%GGemini 4 Argon58.80%Claude Fable 5 (Anthropic): 56.07%AClaude Fable 556.07%Gemini 3 Flash (12/25) (Google): 55.92%GGemini 3 Flash (12/25)55.92%Gemini 3.5 Flash (Google): 55.83%GGemini 3.5 Flash55.83%Claude Opus 4.7 (Anthropic): 54.86%AClaude Opus 4.754.86%Claude Fable 5.1 (Anthropic): 53.51%AClaude Fable 5.153.51%Gemini 3.7 Flash (Google): 53.39%GGemini 3.7 Flash53.39%Claude Opus 4.8 (Anthropic): 53.22%AClaude Opus 4.853.22%Gemini 3.6 Flash (Google): 53.15%GGemini 3.6 Flash53.15%Claude Sonnet 5.5 (Anthropic): 52.92%AClaude Sonnet 5.552.92%GPT 5.1 (OpenAI): 52.73%OGPT 5.152.73%Gemini 3 Pro (11/25) (Google): 52.20%GGemini 3 Pro (11/25)52.20%Claude Haiku 5.5 (Anthropic): 51.54%AClaude Haiku 5.551.54%
Accuracy
Share of coding tasks answered correctly.
MedCode scores
RankModelProviderAccuracyPrice in / out ($/1M tokens)
1 Claude Opus 5Anthropic63.57%$5 / $25
2 Gemini 3.1 Pro Preview (02/26)Google59.06%$2 / $12
3 Gemini 4 ArgonGoogle58.80%$4 / $20
4 Claude Fable 5Anthropic56.07%$10 / $50
5 Gemini 3 Flash (12/25)Google55.92%$0.5 / $3
6 Gemini 3.5 FlashGoogle55.83%$1.5 / $9
7 Claude Opus 4.7Anthropic54.86%$5 / $25
8 Claude Fable 5.1Anthropic53.51%$10 / $50
9 Gemini 3.7 FlashGoogle53.39%$1.5 / $7.5
10 Claude Opus 4.8Anthropic53.22%$5 / $25
Show 15 more rows
RankModelProviderAccuracyPrice in / out ($/1M tokens)
11 Gemini 3.6 FlashGoogle53.15%$1.5 / $7.5
12 Claude Sonnet 5.5Anthropic52.92%$2 / $10
13 GPT 5.1OpenAI52.73%$1.25 / $10
14 Gemini 3 Pro (11/25)Google52.20%$2 / $12
15 Claude Haiku 5.5Anthropic51.54%$0.1 / $0.5
16 Muse SparkMeta51.31%N/A
17 Gemini 2.5 ProGoogle50.59%$1.25 / $10
18 Wonder FullMeta49.93%$4 / $20
19 Claude Opus 5.5Anthropic49.80%$4 / $20
20 GPT 5.2OpenAI49.75%$1.75 / $14
21 GPT 5OpenAI49.63%$1.25 / $10
22 Grok 4.7xAI49.55%$2 / $6
23 Kimi K3Moonshot49.36%$3 / $15
24 Muse Spark 1.2Meta49.35%$1.25 / $4.25
25 Claude Opus 4.5 (Thinking)Anthropic49.16%$5 / $25

Source: Vals AI leaderboard ↗

Top model here: Claude Opus 5 (Anthropic). Try prompts for Claude →

MedQA

US medical licensing exam style multiple-choice questions, including variants that test for demographic bias.

Run by Vals AI. Leaderboard as of 16 Apr 2026 · retrieved 21 Oct 2026. Showing top 25 of 95 models listed.

  • Anthropic
  • OpenAI
  • Google
  • xAI
  • DeepSeek
  • Alibaba
  • Z.ai
  • Moonshot
MedQA: top 15 by AccuracyRanked bars. Leader: o1 at 96.52%. Full values in the table below.0%20%40%60%80%100%o1 (OpenAI): 96.52%Oo196.52%GPT 5.1 (OpenAI): 96.38%OGPT 5.196.38%Gemini 3.1 Pro Preview (02/26) (Google): 96.37%GGemini 3.1 Pro Preview (02/26)96.37%GPT 5 (OpenAI): 96.32%OGPT 596.32%GPT 5.4 (xhigh) (OpenAI): 96.09%OGPT 5.4 (xhigh)96.09%o3 (OpenAI): 96.06%Oo396.06%GPT 5 Mini (OpenAI): 96.06%OGPT 5 Mini96.06%Gemini 3 Pro (11/25) (Google): 96.03%GGemini 3 Pro (11/25)96.03%o4 Mini (OpenAI): 96.02%Oo4 Mini96.02%Claude Opus 4.5 (Thinking) (Anthropic): 95.88%AClaude Opus 4.5 (Thinking)95.88%Gemini 3 Flash (12/25) (Google): 95.81%GGemini 3 Flash (12/25)95.81%Claude Opus 4.6 (Thinking) (Anthropic): 95.41%AClaude Opus 4.6 (Thinking)95.41%Qwen 3.5 Plus (Alibaba): 95.21%QQwen 3.5 Plus95.21%o3 Mini (OpenAI): 94.83%Oo3 Mini94.83%Claude Sonnet 4.5 (Thinking) (Anthropic): 94.71%AClaude Sonnet 4.5 (Thinking)94.71%
Accuracy
Share of exam questions answered correctly.
MedQA scores
RankModelProviderAccuracyPrice in / out ($/1M tokens)
1 o1OpenAI96.52%$15 / $60
2 GPT 5.1OpenAI96.38%$1.25 / $10
3 Gemini 3.1 Pro Preview (02/26)Google96.37%$2 / $12
4 GPT 5OpenAI96.32%$1.25 / $10
5 GPT 5.4 (xhigh)OpenAI96.09%$2.5 / $15
6 o3OpenAI96.06%$2 / $8
7 GPT 5 MiniOpenAI96.06%$0.25 / $2
8 Gemini 3 Pro (11/25)Google96.03%$2 / $12
9 o4 MiniOpenAI96.02%$1.1 / $4.4
10 Claude Opus 4.5 (Thinking)Anthropic95.88%$5 / $25
Show 15 more rows
RankModelProviderAccuracyPrice in / out ($/1M tokens)
11 Gemini 3 Flash (12/25)Google95.81%$0.5 / $3
12 Claude Opus 4.6 (Thinking)Anthropic95.41%$5 / $25
13 Qwen 3.5 PlusAlibaba95.21%$0.4 / $2.4
14 o3 MiniOpenAI94.83%$1.1 / $4.4
15 Claude Sonnet 4.5 (Thinking)Anthropic94.71%$3 / $15
16 Grok 4.20 (Reasoning)xAI94.55%$2 / $6
17 Kimi K2.5Moonshot94.37%$0.6 / $3
18 GLM 5Z.ai94.27%$1 / $3.2
19 GPT 5.2OpenAI94.13%$1.75 / $14
20 DeepSeek V3.2 (Thinking)DeepSeek93.92%$0.56 / $1.68
21 GLM 4.7Z.ai93.74%$0.6 / $2.2
22 Claude Opus 4.1 (Thinking)Anthropic93.59%$15 / $75
23 GPT 5 NanoOpenAI93.26%$0.05 / $0.4
24 Claude Opus 4.5 (Nonthinking)Anthropic93.16%$5 / $25
25 Gemini 2.5 Pro ExpGoogle93.14%$1.25 / $10

Source: Vals AI leaderboard ↗

Top model here: o1 (OpenAI). Try prompts for ChatGPT →

ChatJudge does not run these tests. Scores are copied as published by each source, with a link and the date we read them, and only change when we refresh from the source. Different leaderboards use different settings, so compare models within one chart, not across charts. Spotted a mistake? Tell us.