Health AI benchmark index

Measuring health AI
by its sources

Published model scores on five clinical and biomedical benchmark views, grouped into separate evaluation cohorts. Every cohort carries its metric, protocol, source date and a link back to the paper or page that reported it.

170 unique sourced results in 26 cohorts · latest verified sources, checked Oct 24, 2026

Selected source snapshots

Nox Orion 2.1#1 in listed HealthBench Professional results · team-run · score calculation verified, Oct 23, 2026 run; arithmetic checked Oct 24

Muse Spark Thinking#1 in HealthBench Hard · Meta Muse Spark eval methodology, Undated; checked Oct 24, 2026

o3#1 in MedQA-English · MedGemma technical report v4, Table 3, Apr 6, 2026

HealthBench Professional performance 14 / 22 models

Performance vs API cost per task — not a model ranking. Length-adjusted score out of 100.

Sources and cohorts

Cost dataCost axisProvider linesEstimated: calculated scenario

10 listed models not plotted — API cost unavailable (no matched token totals and prices); not placed at zero:

Nox Orion 2.1 69.528
Claude Mythos 5 (max) 66.0
Claude Mythos Preview (max) 64.7
GPT-6.1 Sol (max) 64.634
GPT-5.6 Sol 60.5
Claude Fable 5.1 (high) 57.971
GPT-5.6 Terra 57.7
Claude Opus 4.8 (max) 56.9
GPT-5.6 Luna 55.7
GPT-5.5 51.8

Calculated: (input tokens × input USD/M + (answer + reasoning tokens) × output USD/M) ÷ 1,000,000 ÷ 525. 525 published tasks; input, answer and reasoning tokens included. Calculated uncached list-price scenario using the source’s canonical evaluation token totals, not recorded billed spend. Includes answer and reasoning tokens; excludes caching, batch discounts, retries, grading and infrastructure. Prices are a dated snapshot, not a live quote. Same Artificial Analysis cohort, checked 2026-10-24. Source Provider lines join different models from the same provider sorted by cost; they are provider comparisons, not repeated observations or a scaling curve. Scores from different reports are not a controlled comparison.

Show source table (22 listed models)

Model	Score	95% CI	Evidence	Source date	Calc. cost/task	Published cost/task	Source
Nox Orion 2.1	69.528	65.81–73.03	Evaluated by HealthBenchAI	Oct 23, 2026 run; arithmetic checked Oct 24	unavailable	—	HealthBenchAI: Nox Orion test results and evidence
Claude Mythos 5 (max)	66.0	—	official vendor report	Jun 9, 2026	unavailable	—	Anthropic Fable 5 / Mythos 5 system card, Table 8.1.A and Figure 8.18.2.A (p. 300)
GPT-6 Astra (Max)	66.0	—	independent evaluation	Oct 19, 2026	$0.27	$0.27	Artificial Analysis original announcement and 14-model chart
Claude Mythos Preview (max)	64.7	—	official vendor report	Jun 9, 2026	unavailable	—	Anthropic Fable 5 / Mythos 5 system card, Table 8.1.A and Figure 8.18.2.A (p. 300)
GPT-6.1 Sol (max)	64.634	60.78–68.35	Evaluated by HealthBenchAI	Oct 25, 2026	unavailable	—	GPT-6.1 Sol (max): our evaluation article
Claude Sonnet 5.5 (Max, Default Fallback)	63.0	—	independent evaluation	Oct 19, 2026	$0.24	—	Artificial Analysis original announcement and 14-model chart
GPT-6 Sol (Max)	62.3	—	independent evaluation	Oct 19, 2026	$0.036	$0.036	Artificial Analysis original announcement and 14-model chart
GPT-6 Luna (Max)	61.9	—	independent evaluation	Oct 19, 2026	$0.0041	$0.0040	Artificial Analysis original announcement and 14-model chart
GPT-5.6 Sol	60.5	—	official vendor report	Undated page; checked Oct 24, 2026	unavailable	—	OpenAI GPT-6.1 Sol system card, Table 7
Claude Opus 5.5 (Max, Default Fallback)	59.9	—	independent evaluation	Oct 19, 2026	$0.63	—	Artificial Analysis original announcement and 14-model chart
Claude Haiku 5.5 (Max)	58.4	—	independent evaluation	Oct 19, 2026	$0.017	—	Artificial Analysis original announcement and 14-model chart
Claude Fable 5.1 (high)	57.971	53.38–62.48	Evaluated by HealthBenchAI	Oct 25, 2026	unavailable	—	Claude Fable 5.1 (high): our evaluation article
GPT-5.6 Terra	57.7	—	official vendor report	Undated page; checked Oct 24, 2026	unavailable	—	OpenAI GPT-6.1 Sol system card, Table 7
Claude Opus 4.8 (max)	56.9	—	official vendor report	Jun 9, 2026	unavailable	—	Anthropic Fable 5 / Mythos 5 system card, Table 8.1.A and Figure 8.18.2.A (p. 300)
Grok 4.7 (Xhigh)	55.8	—	independent evaluation	Oct 19, 2026	$0.020	—	Artificial Analysis original announcement and 14-model chart
GPT-5.6 Luna	55.7	—	official vendor report	Undated page; checked Oct 24, 2026	unavailable	—	OpenAI GPT-6.1 Sol system card, Table 7
Gemini 4 Argon (High)	55.1	—	independent evaluation	Oct 19, 2026	$0.030	—	Artificial Analysis original announcement and 14-model chart
Kimi K3 (Max)	53.2	—	independent evaluation	Oct 19, 2026	$0.059	—	Artificial Analysis original announcement and 14-model chart
GLM-5.3 (Max)	52.8	—	independent evaluation	Oct 19, 2026	$0.032	—	Artificial Analysis original announcement and 14-model chart
GPT-5.5	51.8	—	official vendor report	Undated page; checked Oct 24, 2026	unavailable	—	OpenAI GPT-6.1 Sol system card, Table 7
Gemini 3.8 Flash (High)	51.7	—	independent evaluation	Oct 19, 2026	$0.010	—	Artificial Analysis original announcement and 14-model chart
DeepSeek V4.1 Flash (Max)	50.7	—	independent evaluation	Oct 19, 2026	$0.0099	—	Artificial Analysis original announcement and 14-model chart

Pareto Optimal Models Cost frontier

Artificial Analysis · Oct 19, 2026 · 4 of 4 published pairs on frontier. Bar length shows score out of 100.

GPT-6 Astra (Max)$0.27/taskscore 66.0
GPT-6.1 Sol (Max)$0.047/taskscore 65.1
GPT-6 Sol (Max)$0.036/taskscore 62.3
GPT-6 Luna (Max)$0.004/taskscore 61.9

View evaluation details

Leaderboard Trends

Published evidence coverage over time — cumulative result records, not model performance

Scope

Open weights · 10 recordsProprietary · 32 records

7 records excluded as unclassified or outside the shown series. Month-only and undated sources are omitted rather than given invented days.

Each record is one model result in a published benchmark cohort, not a unique model or a new evaluation by this site. Open weights does not imply an OSI-approved license; unrecorded availability is unclassified. US / China refers to provider location, not model quality. Top labs are selected by record count, not benchmark score.

Show data table