30% OFF Pro PlanClaim Now

Medical AI market data

Medical AI Models: Where to Run Them and What They Cost (2026)

Last verified: ยท 35 models ยท Download CSV

Short answer

Most medical AI models are released as open weights, so there are three ways to run them:

  1. Hosted API, billed per token or by subscription โ€” usually the cheapest way to start, but only a few medical models are hosted (for example Baichuan's official API, and several open models on Featherless).
  2. Cloud endpoint, billed per hour โ€” Google Cloud Model Garden, Microsoft Foundry, or an AWS Marketplace listing deploys the model for you on a GPU you pay for while it runs.
  3. Self-host the open weights โ€” download from Hugging Face and run on your own GPU; the model is free, the GPU is not.

Always check the license: several popular models are non-commercial. Do not assume any model is cleared for clinical use; many sources explicitly say it is not, as-is.

Models, access routes and licenses

Each row links to the primary source used for verification. Prices are list prices before taxes and discounts.

Google โ€” Health AI Developer Foundations (HAI-DEF)

MedGemma 4B
Google ยท 4.3B ยท Text + medical images
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

MedGemma 1.5 4B
Google ยท 4.3B ยท Text + medical images
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

MedGemma 27B (multimodal)
Google ยท 28.8B ยท Text + medical images
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

MedGemma 27B (text-only)
Google ยท 27.0B ยท Text
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

MedSigLIP
Google ยท 0.88B ยท Medical image encoder (embeddings, zero-shot)
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

MedASR
Google ยท 105M ยท Medical speech recognition
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

TxGemma 27B (chat)
Google ยท 27.2B ยท Text (therapeutics development)
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

HeAR
Google ยท Health acoustics embeddings (e.g. cough)
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

Path Foundation
Google ยท Histopathology image embeddings
Health AI Developer Foundations Terms of Use
Commercial with conditions

Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

CXR Foundation (legacy)
Google ยท Chest X-ray embeddings
Health AI Developer Foundations Terms of Use
Commercial with conditions

Google marks it legacy and recommends MedSigLIP instead. Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

Derm Foundation (legacy)
Google ยท Dermatology image embeddings
Health AI Developer Foundations Terms of Use
Commercial with conditions

Google marks it legacy and recommends MedSigLIP instead. Commercial use allowed under the HAI-DEF terms; you must obtain any required health-regulatory authorization before clinical use.

Google DeepMind

AlphaGenome
Google DeepMind ยท DNA sequence โ†’ regulatory predictions
AlphaGenome Terms of Use
Commercial with conditions

API is free for non-commercial use; commercial use goes through Google Cloud. Research tool, not for clinical decision-making.

Microsoft

MedImageInsight
Microsoft ยท Medical image embeddings (multiple modalities)
Microsoft Foundry catalog terms
Commercial with conditions

Microsoft states the healthcare models are for research and development, not for clinical use as-is. A closed-weight "Premium" serverless version is listed; no price was published on the Azure Foundry pricing page on the verification date.

CXRReportGen
Microsoft ยท Chest X-ray โ†’ draft findings report
Microsoft Foundry catalog terms
Commercial with conditions

Research and development only, not for clinical use as-is. A closed-weight "Premium" serverless version is listed without a published price.

RAD-DINO
Microsoft ยท 86.6M ยท Chest X-ray image encoder
MIT
Commercial OK

Model card: shared for research purposes only, not meant for clinical practice.

Baichuan

Baichuan-M3
Baichuan AI ยท 235B (open weights) ยท Text (medical reasoning)
Apache-2.0 (open weights)
Commercial OK

Apache-2.0 covers the open weights; the hosted API is governed by Baichuan's platform terms.

Baichuan-M3-Plus
Baichuan AI ยท Text (medical reasoning + medical search)
Proprietary API
Commercial with conditions

Usage rights are governed by Baichuan's platform terms; check them before commercial use.

  • Hosted APIBaichuan official API$0.74 input / $1.34 output per 1M tokens (ยฅ0.005 / ยฅ0.009 per 1K), plus ยฅ0.03 per automatic medical search
Baichuan-M2
Baichuan AI ยท 32.8B (open weights) ยท Text (medical reasoning)
Apache-2.0 (open weights)
Commercial OK

Apache-2.0 covers the open weights; the hosted API is governed by Baichuan's platform terms.

Baichuan-M2-Plus
Baichuan AI ยท Text (medical reasoning + medical search)
Proprietary API
Commercial with conditions

Usage rights are governed by Baichuan's platform terms; check them before commercial use.

  • Hosted APIBaichuan official API$1.49 input / $4.46 output per 1M tokens (ยฅ0.01 / ยฅ0.03 per 1K), plus ยฅ0.03 per automatic medical search

Academic & community open models

Llama3-Med42-70B
M42 Health ยท 70.6B ยท Text (clinical)
Llama 3 Community License
Commercial with conditions

Commercial use under the Llama 3 license terms.

Llama3-OpenBioLLM-70B
aaditya (OpenBioLLM) ยท Text (biomedical)
Llama 3 Community License
Commercial with conditions

Commercial use under the Llama 3 license terms.

Meditron-7B
EPFL ยท 6.7B ยท Text (medical)
Llama 2 Community License (gated)
Commercial with conditions

Commercial use under the Llama 2 license terms.

PMC-LLaMA 13B
axiong (PMC-LLaMA authors) ยท Text (medical)
OpenRAIL
Commercial with conditions

Use-based restrictions apply under OpenRAIL.

Virchow2
Paige ยท 631M ยท Pathology image embeddings
CC BY-NC-ND 4.0 (gated)
Non-commercial only

GPU reference prices for self-hosting

Single-GPU list prices on Hugging Face Inference Endpoints (billed by the minute). Google Cloud and Azure price their own GPU machines separately.

GPUCloudPer hourAlways on (730 h)
NVIDIA T4 (16 GB)AWS$0.50$365
NVIDIA L4 (24 GB)AWS$0.80$584
NVIDIA A10G (24 GB)AWS$1.00$730
NVIDIA L40S (48 GB)AWS$1.80$1,314
NVIDIA A100 (80 GB)AWS$2.50$1,825
NVIDIA H200 (141 GB)AWS$5.00$3,650
NVIDIA T4 (16 GB)GCP$0.50$365
NVIDIA L4 (24 GB)GCP$0.70$511
NVIDIA A100 (80 GB)GCP$3.60$2,628
NVIDIA H100 (80 GB)GCP$10.00$7,300

What a month usually costs

Assumptions: one GPU kept running 730 hours a month, list prices, no autoscaling or discounts. At 16-bit precision the weights alone take about 2 bytes per parameter: roughly 9 GB for a 4B-class model and 54 GB for a 27B-class model, before runtime memory. The examples below use a single 24 GB L4 and a single 80 GB A100; quantization, multiple GPUs or offloading are alternatives.

4B-class model on one L4

$511โ€“$584

per month, Hugging Face Endpoints on GCP or AWS

27B-class model on one A100 80 GB

$1,825โ€“$2,628

per month, Hugging Face Endpoints on AWS or GCP

AWS Marketplace MedGemma listing

$1,095 + GPU

per month: $1.50/host-hour software fee, plus the SageMaker instance

Hosted options avoid the idle-GPU bill: Featherless API access starts with its Developer plan (from $50/month in credits, billed per token; the $25 Chat plan targets chat apps), and Baichuan's official API is billed per token. Rates for Baichuan are converted at 1 USD = 6.7227 CNY (open.er-api.com, 2026-09-24 00:02 UTC) (source).

Retired or not publicly priced

  • MedLM (built on Med-PaLM 2)

    Retired from Vertex AI. Google Cloud announced on May 14, 2025 that access ends on or after September 29, 2025, so it is no longer available through Vertex AI.

    Source
  • AlphaGenome on Google Cloud (commercial)

    Available to allowlisted users via a request form; no public model price. You also pay for the A100/H100 80 GB endpoint.

    Source
  • MedImageInsight Premium / CxrReportGen Premium (Azure serverless)

    Listed by Microsoft as closed-weight serverless models; no price appeared on the Azure Foundry models pricing page on the verification date.

    Source

Methodology and limitations

  • Every fact was checked on September 24, 2026 against a primary source: official model cards and license metadata on Hugging Face, vendor documentation, and vendor price pages. Items that could only be found on third-party aggregators were left out.
  • "Hosted API" means a provider serves the model for you. "Cloud endpoint" means the model is deployed on a GPU billed per hour. "Self-host" means you download the weights and run them yourself.
  • Available to buy does not mean approved for clinical use. Google's HAI-DEF terms, Microsoft's healthcare model documentation and several model cards state that the models are for research and development and require your own validation and any regulatory authorization before clinical use.
  • Prices change often and exclude taxes, storage, networking and discounts. Check the linked source before you buy.
  • Parameter counts come from Hugging Face model metadata and are left blank where the metadata does not state them.

Want one API instead?

Dr7.ai offers its own OpenAI-compatible medical AI API with per-token pricing. See the Dr7.ai API.

Corrections and updates

Found an outdated price or a missing provider? Email [email protected]. We re-check the full table every quarter.

  • โ€” First version: models, access routes, licenses and GPU reference prices verified against primary sources.