Which AI Models Sign a BAA? OpenAI, Claude and Gemini for Healthcare Apps
Key Takeaways
- OpenAI, Anthropic, Google, Microsoft and AWS all offer a BAA, but only for named products and configurations, never for free consumer accounts.
- OpenAI’s API BAA needs Modified Retention on your organization, and ChatGPT Business has no BAA at all.
- On a HIPAA-ready Claude API organization, the Batch API, Files API, Skills API, Code Execution and Web Fetch aren’t covered, and the API blocks them with a 400 error.
- Google’s BAA covers Gemini in Workspace and on Gemini Enterprise Agent Platform (formerly Vertex AI), but AI Studio and the Gemini Developer API aren’t on its covered list.
- A signed BAA covers the vendor’s side only, so risk analysis, access control, logging and human review stay with your team.
There’s no such thing as a HIPAA-compliant LLM. A model is just a model. What can be compliant is the way you deploy it: the contract you sign, the product tier you use, the features you turn on, and the data you send.
That distinction catches a lot of teams building their first AI feature. Someone tests a prompt with real patient notes on a personal account; it works beautifully, and only later does anyone ask whether the vendor signed anything. If you’re scoping healthcare app development with an LLM inside, this is one of the first questions to settle, because the answer can change your provider, your architecture, and your retention settings.
We’ve compared the main vendors below from a builder’s side. Each section covers what the BAA includes, what it leaves out, and what you need to switch on before PHI touches the API. Terms in this space change often, so we’ve linked each vendor’s own documentation. Treat that documentation and your signed agreement as the final word.
Which AI Models Will Sign a BAA in 2026?
All five major routes offer a BAA. The differences sit in which plans qualify and what the agreement excludes.
| Vendor | Covered route | BAA available? | Retention terms | Main exclusions |
|---|---|---|---|---|
| OpenAI | API with Modified Retention | Yes. Self-serve for eligible organizations | Modified Retention required unless OpenAI says otherwise | Check the eligible endpoint list |
| OpenAI | ChatGPT for Healthcare, Enterprise with Regulated Workspace, ChatGPT for Clinicians | Yes | Set by product and contract | ChatGPT Business has no BAA. Event-triggered tasks aren’t covered |
| Anthropic | Claude API on a HIPAA-ready organization | Yes. Self-serve for eligible organizations | Standard retention or ZDR, depending on setup. Covered Models need 30 days | Batch, Files, Skills, Code Execution, Web Fetch |
| Anthropic | HIPAA-ready Claude Enterprise | Yes. Click-to-accept for eligible organizations | Standard retention | Claude Code and Cowork need an extra configuration |
| Gemini Enterprise Agent Platform, Gemini Enterprise | Yes, under the Google Cloud BAA | Follow the HIPAA implementation guidance | AI Studio and the Gemini Developer API aren’t listed | |
| Gemini app and Gemini in Workspace | Yes, under the BAA with a Workspace account | Workspace controls apply | Gemini in Chrome. Consumer accounts | |
| Microsoft | OpenAI models sold through Azure in Microsoft Foundry | Yes, included through the Product Terms | Abuse monitoring may store prompts unless you’re approved to modify it | In-scope services only |
| AWS | Amazon Bedrock | Yes, under the AWS BAA | Your AWS configuration controls logging and storage | AWS’s eligible services list applies |
Before you pick a row, read the HHS guidance on HIPAA and cloud computing. HHS treats a cloud provider that handles ePHI as a business associate even when it only stores encrypted data. An LLM API receives readable text, so the case for a BAA is even simpler.
What Does an OpenAI BAA Cover?
For API work, the OpenAI BAA no longer needs an enterprise agreement. Eligible organizations can accept the standard Business Associate and Healthcare Addendum in API Platform settings, under Organization and then General. Self-serve enrollment needs an established history of API usage. If your account shows “Not eligible yet,” try again later or email OpenAI’s BAA team.
Two details matter more than the signature:
- Modified Retention is required. OpenAI’s HIPAA eligible products page says API eligibility needs an executed BAA, and your organization must also be provisioned with Modified Retention unless OpenAI specifies otherwise.
- The switch is one-way. Once HIPAA compliance support is on, you can’t turn it off in API Platform settings. If you share one organization across production and experiments, set up a separate one first.
So, is OpenAI HIPAA compliant? For the API, it can support your compliance when both conditions are met and you stick to the eligible endpoints. Check that list against your design, especially if you plan to use newer agent features.
On the ChatGPT side, the eligible products are ChatGPT for Healthcare, ChatGPT for Enterprise with Regulated Workspace, ChatGPT for Clinicians and ChatGPT FedRAMP. OpenAI states plainly that it doesn’t offer a BAA for ChatGPT Business, which surprises many small practices. Individual clinicians on ChatGPT for Clinicians have their own in-product BAA flow, but that’s a personal workspace, and it won’t support an app you’re building for patients.
Also keep PHI out of OpenAI support tickets, screenshots and attachments. The company tells customers not to include it, and engineers tend to forget this during a production incident.
Is Claude HIPAA-Compliant?
The short answer to “is Claude HIPAA compliant” is that it can be, on the right route. Anthropic offers a BAA for its first-party API and for HIPAA-ready Enterprise plans, and its BAA documentation lists coverage feature by feature.
For an app, the API route is the one you care about. Eligible organizations can enable HIPAA readiness in the Claude Console by accepting the standard BAA. Like OpenAI’s setting, it’s permanent once enabled, so keep test work in another organization.
What makes Anthropic’s setup useful for builders is that the API enforces the limits. Send a request using a feature that isn’t eligible and you get a 400 error, so a developer can’t quietly route PHI through it. Here’s where the line sits:
- Covered: the Messages API, including prompt caching, structured outputs, memory, web search, the bash tool and the text editor tool. Computer Use and the Token Counting, Models, Org Management and Compliance APIs are also covered.
- Not covered: the Batch API, Files API, Skills API, Code Execution and Web Fetch.
- Your risk: External MCP is available, but data you send to third parties through it isn’t covered by Anthropic’s BAA.
Retention needs a closer look. Anthropic’s newer Covered Models, such as Claude Fable 5, require 30-day data retention and can’t run on organizations or workspaces with zero data retention (ZDR). If your compliance plan assumes ZDR, that decision limits which models you can use.
HIPAA-ready Enterprise is a separate product for staff using Claude directly, and the BAA is click-to-accept for eligible organizations. Claude Code and Cowork need an extra HIPAA configuration before they’re covered. One catch for older customers: an API BAA signed before December 2, 2025 doesn’t extend to the Enterprise plan.
Claude is also available through Amazon Bedrock and Google Cloud, where your agreement sits with the cloud provider. Through Microsoft Foundry, you need a BAA with Anthropic plus the HIPAA configuration on your Azure subscription. Confirm the partner route with your account team, because a BAA with one vendor doesn’t carry over to another.
Is Gemini HIPAA Compliant?
Gemini HIPAA compliant answers depend on which Gemini you mean, since the name covers several products with different terms.
For an app, use Gemini through Google Cloud. Google’s HIPAA page lists Gemini Enterprise, Gemini Enterprise Agent Platform and Generative AI on Agent Platform under the Google Cloud BAA. If you’re reading older guides, Agent Platform is the product many of them still call Vertex AI. Don’t assume a Google service is covered unless you can find it on that list.
The trap is the developer route. AI Studio and the Gemini Developer API aren’t on that list, and many prototypes start there because signing up takes two minutes. The paid Gemini API has decent data terms, but decent data terms don’t equal a BAA. Its unpaid tier is worse for this purpose, since human reviewers may read your input and output.
If your prototype runs on a Developer API key, plan the move to Agent Platform before any real patient data enters the system. Expect some rework around authentication, regions and quotas.
For staff tools, the Gemini app and Gemini in Google Workspace are in the BAA for Workspace customers. Consumer Google accounts have no BAA, and Google says Gemini in Chrome isn’t covered, so admins may want to turn it off. Google also notes that some Gemini app features can be blocked for customers who have signed the BAA.
Can You Get a HIPAA Compliant LLM Through Azure or AWS?
Yes, and for many teams this is the easiest route, since the BAA already sits with a cloud provider you use for hosting.
Microsoft Foundry. Microsoft includes its BAA through its Product Terms for covered entities and business associates, so there’s no separate document to sign. OpenAI models sold by Azure don’t send your prompts to OpenAI. Watch abuse monitoring, though. Microsoft’s documentation says flagged prompts and completions can be stored for human review unless you’ve been approved for modified abuse monitoring. Apply early if your risk analysis needs that off.
Amazon Bedrock. Bedrock is on AWS’s HIPAA eligible list, and AWS says model providers can’t reach your prompts or completions. You still have to sign the AWS BAA and configure the service properly. Also keep PHI out of resource names and tags, since AWS warns that those fields may end up in billing or diagnostic logs.
Self-hosting an open-weight model on your own HIPAA-eligible cloud setup is the other option. No model vendor sees your data, but you now run the GPUs, patching, scaling and monitoring yourself. That usually only pays off at high volume or with strict data residency needs.
Why Is ChatGPT Health Not a HIPAA Route for Your App?
“Is ChatGPT Health HIPAA compliant” is the wrong question for a builder. ChatGPT Health is a consumer feature for people on Free, Go, Plus and Pro plans. It doesn’t appear on OpenAI’s list of HIPAA eligible products, and it isn’t something you can build a patient product on.
A patient can choose to upload their own records to a consumer tool. That’s their decision. It’s very different from your clinic or app sending patient data to a vendor on someone’s behalf, which is the situation a BAA exists for. If a user says “just connect it to ChatGPT,” the answer is the API route above, with the right agreement in place.
What Does Your Team Still Own After Signing a BAA?
A BAA moves some responsibility to the vendor. It doesn’t make your app compliant. Every vendor above says so in some form, and Google’s page calls it shared responsibility outright.
Here’s what stays with you:
- Risk analysis. Your LLM feature needs its own entry, covering what data goes in, where it’s stored and who can see the logs.
- Your full vendor chain. Your vector database, observability tool and prompt logger each need a BAA if they touch PHI. A careful BAA vendor review often finds more gaps here than at the model provider.
- Access control and audit logs. Log who sent what to the model, without creating a new unprotected copy of PHI in the process.
- Feature discipline. Someone has to check each new model feature against the eligible list before release.
- Human review. Clinical output needs a defined human oversight for HIPAA AI process, and a healthcare LLM evaluation plan that runs before and after launch.
If your team built its prototype with AI coding tools, add one more check. A HIPAA audit checklist will usually find issues in auth and logging long before anyone looks at the model.
How to Keep PHI Out of the Prompt Anyway
The safest PHI is the PHI you never send. Even with a BAA, HIPAA’s minimum necessary standard pushes you to send only what the task requires.
A few practical habits help:
- De-identify before the call. Strip or tokenize names, dates and record numbers, then restore them after the response. Data de-identification for healthcare AI works well when the task doesn’t depend on identity.
- Send fields, not whole charts. Pass the five facts the task needs instead of the full record.
- Control logging. Turn off prompt logging in your own stack by default, or redact before you write to logs.
- Keep test data synthetic. Developers don’t need real patient data to tune a prompt.
- Separate your environments. Use one HIPAA-enabled organization for production and another for experiments, with no PHI in the second.
These habits matter most in patient-facing products. A HIPAA-compliant AI chatbot collects free text, and patients type things you never planned for, so filter before the model call.
How Do You Choose a HIPAA Compliant LLM for Your App?
Choosing comes down to where you already run. These rules cover most cases:
- Already on AWS, Azure or Google Cloud: start with that cloud’s model service, because your BAA and security reviews are in place.
- Need a specific model’s quality: go first-party with OpenAI or Anthropic, and budget time for retention settings.
- Need strict zero retention: check model availability first, because the newest models may not support it.
- Staff tools only, no custom app: look at ChatGPT for Healthcare, HIPAA-ready Claude Enterprise or Gemini in Workspace.
Tech Exactly helped a client build a HIPAA-compliant AI app for autism caregivers, and the model decision on projects like that one is rarely the hard part. The hard part is the data flow around the model. Teams that hire AI developers with healthcare experience usually map that flow first and choose a provider second, which avoids moving a working prototype to a new vendor months later.
The best HIPAA compliant AI models for your app are the ones you can put under a signed agreement, with retention settings your risk analysis can defend.
Frequently Asked Questions
No. HIPAA applies to organizations and how they handle data, and HHS doesn't certify software. An LLM can only be part of a compliant setup when a BAA, the right product tier and your own safeguards are all in place.
Yes. OpenAI offers a BAA for the API, where eligible organizations can accept it in settings, and for ChatGPT for Healthcare, Enterprise with Regulated Workspace and ChatGPT for Clinicians. It doesn't offer one for ChatGPT Business.
Anthropic's HIPAA-ready offering is built on Enterprise plans and the API. Don't use individual or Team plans with PHI unless Anthropic confirms coverage in writing for your account.
Not through AI Studio or the Gemini Developer API, since neither is on Google's covered list. Use Gemini through Google Cloud's Agent Platform with the Google Cloud BAA in place.
No. Retention depends on your configuration and the model. Some Claude models need 30-day retention, OpenAI uses Modified Retention, and Azure abuse monitoring can store flagged content unless modified. Read the retention terms alongside the BAA.
Data that is properly de-identified under HIPAA's Safe Harbor or Expert Determination methods is no longer PHI. Removing names alone doesn't meet that bar, so have the method reviewed before you rely on it.
