The Healthcare API for Developers

What is the best healthcare API for developers?

Who builds on EvidenceMD

EvidenceMD is built for developers creating clinical products and for healthcare organisations evaluating evidence-grounded AI workflows. Its OpenAI-compatible interface keeps integration familiar, while the clinical evaluation is authored by a physician-led team spanning hospital medicine and critical care.^[1][14]

Why a healthcare API should show its reasoning

In most domains, a confident answer is the product. In medicine it is the beginning of a review. A clinician who cannot see how a recommendation was reached has to reconstruct it before acting, and reconstructing it costs more than looking it up would have. That is why fluency alone has never been the bottleneck for clinical AI adoption — verifiability has.

EvidenceMD treats reasoning as a response channel rather than a hidden implementation detail, and pairs it with evidence. These are two separate guarantees and a clinical product needs both: the thinking channel shows how the model arrived somewhere, while the inline citations show what it arrived from. An answer with neither is a fluent guess; an answer with citations but no reasoning still hides the step where a mistake would occur.^[1][5]

data: {"choices":[{"delta":{"content":"<thinking>"}}]}
data: {"choices":[{"delta":{"content":"Four pillars of GDMT apply, but eGFR"}}]}
data: {"choices":[{"delta":{"content":" constrains the ARNI and MRA choices..."}}]}
data: {"choices":[{"delta":{"content":"</thinking>\n\n"}}]}
data: {"choices":[{"delta":{"content":"## GDMT in HFrEF with CKD Stage 4\n"}}]}
data: {"choices":[{"delta":{"content":"SGLT2 inhibitors retain benefit [1](https://..."}}]}
data: {"choices":[{"delta":{},"finish_reason":"stop"}]}
data: [DONE]

Benchmark: 54.6 on HealthBench Hard

The EvidenceMD clinical system scores 54.6 on HealthBench Hard, the 1,000-example subset of OpenAI's HealthBench isolating the clinical conversations frontier models struggle with most, against 46.2 for gpt-5-thinking under the same benchmark implementation. That is above the highest figure OpenAI has reported for its own models on this subset. Most clinical AI vendors publish no accuracy benchmark at all, so here is ours with every condition attached.^[14][15]

HealthBench itself is worth understanding before reading the number: 5,000 examples, each pairing a conversation with physician-authored rubric criteria, 48,562 unique criteria in total, built with 262 physicians practising in 60 countries. A model-based grader scores each criterion independently. It replaced multiple-choice medical exams because those saturated, and the Hard subset remains unsaturated.^[15]

Scoring works per criterion: the grader assesses each rubric criterion independently, awarding full points where it is met and none otherwise, with negative criteria subtracting when met. Per-criterion values run from −10 to 10, and the overall score is the mean across examples clipped to the zero-to-one interval. The reference implementation uses GPT-4.1 as the grader, which on HealthBench's consensus criteria matched physician judgment within the top half of human annotators across all themes. The Hard examples span seven themes — emergency referrals, global health, context seeking, health data tasks, hedging, communication, and complex responses — scored across five behavioural axes: accuracy, completeness, communication quality, context awareness, and instruction following.^[14][15]

The conditions attached to that number

A benchmark figure presented without its conditions is not a measurement. These are ours, stated in full so you can place the figure precisely — and so you can hold any other vendor to the same standard.

Implementation

Every row was produced under the original public implementation of HealthBench released with the benchmark, which is the implementation used for the EvidenceMD evaluation. In April 2026 OpenAI moved to an updated implementation and recomputed earlier models under it, reporting materially different figures — on Hard, the recomputed unadjusted GPT-5 score is 41.6 against the 46.2 originally published. Those recomputed figures are deliberately not used here, because mixing implementations would not produce a valid comparison.

No length adjustment

OpenAI's updated implementation introduced a length-adjusted scale that penalises longer responses. The original implementation used here does not apply it, so no figure above is length-adjusted. Response length is a genuine confound in rubric-graded evaluation, and it cuts against evidence-grounded systems, whose outputs carry retrieved sources and are longer by design.

System versus model

EvidenceMD is evaluated as a complete clinical system including retrieval and structured reasoning, while the OpenAI figures are model evaluations. This is the more decision-relevant framing for a health system choosing what to build on, but it is not a like-for-like architectural comparison and is not presented as one.

Which model was measured

The evaluated system is evidencemd-deep, the reasoning model of the three available on the API. The figure should not be read as applying to evidencemd-fast or evidencemd-pro, which were not evaluated on this benchmark. If you are choosing a model on the strength of this number, choose evidencemd-deep.

Harness sensitivity

The OpenAI figures are as that company published them rather than re-run under our evaluation setup. Rubric-graded scores carry some sensitivity to grader configuration, so the margins above are best read as indicative rather than exact.

Absent vendors

Claude and Gemini do not appear because we are not aware of a HealthBench Hard figure published by Anthropic or Google for their own models. Anthropic reports HealthBench Professional, which is a separate evaluation that cannot be compared with Hard. Third-party estimates exist, but rubric-graded scores are sensitive to grader model and prompt, so a figure from an unspecified harness cannot sit in the same table as a publisher-reported one.

What the score does not establish

• No specialty view. HealthBench Hard has no specialty partition — no cardiology cut, no oncology cut. Any specialty chart derived from it would be constructed rather than measured, which is why we decline to produce one.
• Averages hide worst-case behaviour. The benchmark's authors report worst-case scores dropping by roughly a third relative to the mean, with emergency and context-seeking behaviours unreliable across models. An aggregate figure does not establish reliability on the individual difficult case.
• Narrow scope. HealthBench measures conversational clinical reasoning. It does not measure documentation accuracy, citation integrity, EHR integration reliability, latency under production load, or safety behaviour in deployment.
• Checkpoint-bound. The production system is periodically retrained on de-identified deployment data, so a HealthBench figure is only meaningful against a named checkpoint. The one above is dated 20 June 2026.

Self-reported benchmark figures, however carefully produced, are self-reported. Independent evaluation designed and run by clinical researchers is the standard this field should be held to, so this result should be read as a benchmark checkpoint rather than a claim of clinical validation.^[14]

Medically fine-tuned, so you write less prompt

EvidenceMD is built on a model fine-tuned on medical data, which means clinical reasoning is trained into the system rather than requested at inference time. A general-purpose model can be brought close to clinical competence through extensive prompt engineering, task-specific fine-tuning, or an attached search tool, and the results can be respectable — but that competence is assembled around the model by whoever deploys it, and it has to be re-assembled and re-validated every time the base model changes underneath you.^[14]

How much the prompt matters to a general model is measurable, not anecdotal. In the ESMO guideline evaluation cited earlier, prompting strategy significantly changed performance: long prompts improved evidence quality and accuracy for GPT-4o and Claude 3.5 Sonnet, while short prompts improved conciseness.^[12] That is a maintenance burden disguised as a configuration detail. In EvidenceMD, clinical behaviour is the default for every query and does not depend on how well the prompt was written.

The practical version of this: a request to EvidenceMD is a clinical question and an optional specialty string. There is no system prompt to maintain, no citation-format instruction to enforce, and no retrieval chain of your own to keep current as guidelines change.^[1]

How this differs from building on OpenAI or Claude

This is the real build-versus-buy decision, and it deserves an honest answer rather than a strawman. GPT and Claude are excellent models. Both vendors now ship healthcare offerings — OpenAI for Healthcare and Claude for Healthcare — and both offer HIPAA BAA paths for API use, so compliance is not the differentiator some vendors imply it is.^[9][10] Both also expose reasoning: Claude has adaptive and extended thinking, though on current models the thinking text you receive is typically a summary rather than the raw chain of thought.^[11] Outside this comparison, DeepSeek returns a raw reasoning trace in a reasoning_content field, which is worth stating plainly: returning a reasoning trace is not by itself unusual in 2026, so we do not claim it as one.^[19]

The difference that matters is where the grounding lives. A general-purpose model reasons over what it memorised during training. EvidenceMD reasons over literature it retrieves at request time, and returns links to it. That distinction is not academic: in a peer-reviewed evaluation of GPT-4o and Claude against ESMO biliary tract cancer guidelines, the models scored well on accuracy, yet nearly 40% of the references they produced were fabricated or misattributed.^[12] A separate analysis of ChatGPT bibliographies found the same failure mode: citations with real-looking authors, journals, and identifiers that do not resolve to real papers.^[13] A citation generated from memory is a plausible string; a citation returned from retrieval is a URL you can open.

Swipe the table horizontally →

To be clear about scope: OpenAI does offer evidence retrieval with citations, and Claude does connect to PubMed — but those ship inside their end-user products, ChatGPT for Healthcare and Claude for Enterprise, not as default behaviour of the API call a developer makes.^[9][10] If you are building a product, the grounding layer is yours unless you buy one. That is the whole argument for a purpose-built clinical API, and it is also why raw token price is the wrong number to compare on.

One endpoint, the whole clinical surface

There is no separate endpoint per capability. You send clinical context to one chat completions route and shape the response with specialty, language, and the model you pick — which means adding a second clinical feature to your product is a prompt change, not an integration.^[1]

Quickstart: reasoning and citations in under ten minutes

Four steps, no sales call. The endpoint follows the OpenAI chat completion schema,^[4] so if you already call OpenAI you are changing a base URL, an auth header, and a model name.

1. 1

   Create a developer account

   Sign up at evidencemd.ai/developers and verify your email. Free starter credits land as soon as the address is confirmed, so the first requests cost nothing and you can judge output quality before spending.^[2]

2. 2

   Provision an API key

   Create a key in the dashboard. An account holds up to five, so give local, CI, staging, and production their own key and rotate them independently rather than sharing one secret across environments.

3. 3

   Send a request with reasoning enabled

   POST to /api/v1/chat/completions with an x-api-key header and include_thinking set to true. Set stream explicitly — streaming is strongly recommended for anything user-facing, because a non-streaming call returns nothing until the full answer is ready and long clinical questions can hit the server timeout and fail with HTTP 504.^[1]

   curl https://evidencemd.ai/api/v1/chat/completions \
     -H "Content-Type: application/json" \
     -H "x-api-key: $EVIDENCEMD_API_KEY" \
     -d '{
       "model": "evidencemd-pro",
       "messages": [
         {"role": "user", "content": "GDMT for HFrEF in CKD stage 4?"}
       ],
       "stream": true,
       "include_thinking": true,
       "specialty": "Cardiology"
     }'

4. 4

   Split the reasoning, then parse the citations

   Buffer everything between the thinking tags into a reasoning channel and the rest into the answer. Then extract provenance — and note the one place this API diverges from what most developers expect: EvidenceMD does not return a separate sources array. Citations are inline markdown links inside the answer body, in both streaming and non-streaming mode, and an answer with no links is an answer with no retrieved evidence.

   from openai import OpenAI
   
   # EvidenceMD authenticates with x-api-key, not a bearer token, so the key goes
   # in default_headers and api_key just satisfies the SDK constructor.
   client = OpenAI(
       base_url="https://evidencemd.ai/api/v1",
       api_key="unused",
       default_headers={"x-api-key": os.environ["EVIDENCEMD_API_KEY"]},
   )
   
   stream = client.chat.completions.create(
       model="evidencemd-pro",
       messages=[{"role": "user", "content": "GDMT for HFrEF in CKD stage 4?"}],
       stream=True,
       extra_body={"include_thinking": True, "specialty": "Cardiology"},
   )
   
   reasoning, answer, in_thinking = "", "", False
   for chunk in stream:
       text = chunk.choices[0].delta.content or ""
       if "<thinking>" in text:
           in_thinking = True
           text = text.replace("<thinking>", "")
       if "</thinking>" in text:
           in_thinking = False
           continue
       if in_thinking:
           reasoning += text   # render in a reviewer panel, or log for audit
       else:
           answer += text      # render to the user

   import re
   
   # EvidenceMD embeds provenance as inline markdown links in the answer body.
   # There is no separate sources array in streaming or non-streaming mode.
   CITATION = re.compile(r"\[(\d+)\]\((https?://[^)]+)\)")
   
   citations = CITATION.findall(answer)
   # [('1', 'https://pubmed.ncbi.nlm.nih.gov/36346411/'), ...]
   
   # No inline links means no evidence was retrieved for this answer.
   if not citations:
       flag_for_human_review(answer)

Three models, one endpoint

Switching model is a string change. The useful and slightly counterintuitive detail is that fast and pro cost the same 4 credits, so unless latency is your binding constraint there is no cost reason to choose fast.^[1][2]

30 languages, same evidence, same price

Set language on the request and the clinical answer comes back in that language with its citations intact. There is no per-language surcharge, so a product serving a multilingual patient population pays the same model-specific per-request price as an English-only one — which is the difference between multilingual support being a roadmap item and being a request parameter.^[1][3]

# Same evidence base, same price, patient-facing language.
response = client.chat.completions.create(
    model="evidencemd-pro",
    messages=[{"role": "user", "content": "Explain my new metformin prescription"}],
    stream=False,
    extra_body={"language": "Spanish"},
)

One caveat worth knowing before you design around it: the chain-of-thought channel is English-only and is automatically suppressed when language is set to a non-English value. If you need both visible reasoning and a translated answer, run the reasoning pass in English for your reviewers and the patient-facing answer in the target language.

Pricing, printed here rather than behind a sales call

Credits cost $0.05 each and are charged per request, not per token, so a 200-word answer and a 2,000-word answer cost the same and your forecast is one multiplication rather than a token model. Invalid requests are rejected before billing, so a malformed call costs nothing.^[2]

Credit price

$0.05One flat rate. 20 credits per dollar.

fast / pro

$0.204 credits per request, any answer length.

deep

$0.255 credits per request, 64k reasoning tokens.

Monthly minimum

$0No subscription. A quiet month costs nothing.

What teams build with it

The common thread is that all of these need an answer someone can check. Each is worked through in more depth in the use-case guide.^[7]

• Clinical copilots and point-of-care Q&A inside an existing clinician workflow
• Telehealth triage and patient-facing symptom checkers
• AI scribes and documentation review that reason over the note, not just transcribe it
• EHR-embedded decision support surfaced at the point of ordering
• Multilingual patient education and discharge instructions
• Literature synthesis and research tooling for pharma and life sciences
• Medical education products where the reasoning is the lesson
• Agentic clinical workflows that chain several reasoning calls together

Security and compliance

Honest limits worth knowing before you build

Four constraints that are easier to design around than to discover in week three.

Frequently asked questions about the EvidenceMD API

Sources and related guides

Every bracketed marker in the text above links here. Sources 1–3 are the EvidenceMD API specification, published pricing, and product reference behind every parameter and figure on this page; 4 and 6 are the external standards and indexes referenced; 5, 7, and 8 are deeper reading; 9–11 are the vendor documentation behind every OpenAI and Anthropic claim, cited to the vendor rather than paraphrased; 12 and 13 are the peer-reviewed evidence for the citation-fabrication figures; 14 –16 are the technical white paper, the HealthBench paper, and the OpenAI system card behind every benchmark figure; 17–19 are the reasoning and billing documentation for Anthropic, OpenAI, and DeepSeek, including the counter-example to any claim that returning a reasoning trace is by itself unusual. Prices, parameters, and competitor capabilities change — check the source before you build against a number.

1. 1
   Source 1. EvidenceMD OpenAPI specificationMachine-readable schema for EvidenceMD API v2.1.0: the /chat/completions and /models endpoints, the x-api-key security scheme, the three model identifiers, and every documented request parameter and default quoted on this page.
2. 2
   Source 2. EvidenceMD developer dashboard and pricingSource for the credit economics used throughout: $0.05 per credit, 4 credits for fast and pro, 5 for deep, top-ups from $10 for 200 credits, free starter credits on email verification, and per-request billing in which invalid requests are rejected before they are charged.
3. 3
   Source 3. EvidenceMD medical AI API overviewProduct-level reference for the OpenAI-compatible endpoint, the three models, streaming and JSON modes, specialty conditioning, and the multilingual and chain-of-thought options described here.
4. 4
   Source 4. OpenAI chat completions API referenceThe request and response schema the EvidenceMD endpoint is compatible with, and the reference for the SDK usage shown in the quickstart.
5. 5
   Source 5. EvidenceMD benchmark resultsPublished performance on HealthBench Hard and evidence-based clinical reasoning benchmarks — the quality evidence behind the reasoning quality claimed on this page.
6. 6
   Source 6. PubMedThe primary biomedical literature index that inline citations in EvidenceMD answers most commonly resolve to, and the destination a clinician lands on when verifying a cited claim.
7. 7
   Source 7. What you can build with the medical reasoning APIThe use-case companion to this guide, working through ten product patterns — copilots, differential diagnosis, scribes, triage, EHR decision support, and agentic workflows — in more depth than the summary here.
8. 8
   Source 8. EvidenceMD enterprise and volume pricingWhere teams projecting sustained high volume, or needing an executed BAA, custom service levels, and specific deployment options, should start instead of the self-serve dashboard.
9. 9
   Source 9. OpenAI for HealthcareOpenAI's healthcare announcement, and the basis for two claims here: that eligible API customers can apply for a BAA to support HIPAA compliance, and that evidence retrieval with transparent citations is described as a ChatGPT for Healthcare capability rather than default API behaviour.
10. 10
    Source 10. Advancing Claude in healthcare and the life sciencesAnthropic's healthcare announcement, covering Claude for Healthcare, HIPAA-ready products, and the PubMed connector for biomedical literature — a Claude for Enterprise connector rather than default API behaviour.
11. 11
    Source 11. Claude thinking documentationAnthropic's reference for adaptive and extended thinking, and the source for the statement that thinking output is returned as a summary of the reasoning process on current models rather than the raw chain of thought.
12. 12
    Source 12. Prompting strategies and LLM performance on biliary tract cancer guideline queriesPeer-reviewed cross-sectional analysis of GPT-4o, Claude 3.5 Sonnet, and Llama 3 70b against ESMO guidelines across 40 clinical questions, evaluated by two independent senior physicians. Source for the finding that nearly 40% of the references the models produced were fabricated or misattributed despite otherwise strong accuracy.
13. 13
    Source 13. Fabrication and errors in the bibliographic citations generated by ChatGPTWalters and Wilder, Scientific Reports (2023). Independent corroboration that citations generated from model memory frequently carry real-looking authors, journals, and identifiers while not corresponding to real published work.
14. 14
    Source 14. EvidenceMD technical white paper: Clinical Reasoning Performance on HealthBench Hard (25 June 2026)The primary source for the 54.6 figure, the checkpoint date of 20 June 2026, the benchmark implementation used, and every condition and limitation listed in this section. Authored by Krishnakumar Srinivasan with Abhishek Shahi MD, Manjit Dhillon MD, and Shailesh Singh MD.
15. 15
    Source 15. HealthBench (Arora et al., arXiv:2505.08775)The benchmark paper. Source for the 5,000 examples, 48,562 unique criteria, 262 physicians across 60 countries, the construction of the 1,000-example Hard subset, and the finding that worst-case scores fall by roughly a third relative to the mean.
16. 16
    Source 16. OpenAI GPT-5 System Card (August 2025)The publisher-reported source for every OpenAI figure in the benchmark table, produced under the original public implementation of HealthBench rather than re-run by EvidenceMD.
17. 17
    Source 17. Anthropic extended thinking documentationSource for the statement that Claude 4 models return a summary of the thinking process, and that billing covers the full thinking tokens generated rather than the summary tokens returned, so the billed output count will not match the visible one.
18. 18
    Source 18. OpenAI reasoning models documentationSource for the statement that OpenAI reasoning models bill reasoning tokens while returning a summary of the reasoning rather than the raw chain of thought.
19. 19
    Source 19. DeepSeek thinking mode documentationCited as the counter-example to any claim that exposing a reasoning trace is unique: DeepSeek returns chain-of-thought in a reasoning_content field alongside the answer. What distinguishes EvidenceMD is that its reasoning is clinical and produced over retrieved literature, not that a trace is returned at all.