OpenAI announced ChatGPT Health yesterday (leading to my rough attempt at “Sam Altman in the Jon Hamm meme” above), as many pundits and reporters have jumped on with fervor and a groundswell of “breaking news” posts.
These are, of course, just a sampling - if your feed on LinkedIn or Twitter is at all like mine (god have mercy on your soul if so), then there was a veritable litany, an unending torrent. A good time to go outside and touch grass, just to feel something and escape this cursed echo chamber of takes.
Alas. You are here instead. My good friend Mendel and others prompted me for thoughts, so let’s break this puppy down and see what novel insights may lurk beneath the surface.
The Non-Announcement Announcement
The lay-up observation is that this is not news at face value. It’s been well-known that OpenAI would launch a consumer health offering in implicit ways:
However, all the implicit signal in the world is just gravy compared to explicit, overt, unambiguous signal, like them literally signing a pledge in a room with Trump to create a consumer health app, like they did in July! We should not be surprised that they made this move, given we talked about it in August.
The Long Arc of Patient Access
Patient access to health data is not new. We have been on a path of continued iteration and improvement since HIPAA, as detailed in HTD’s 2024 webinar on the topic:
This fifth generation of patient access, powered by digital identity rather than logins, pushes us to a point where perhaps it’s easy enough for most people to use. In a world where individual access to health data has moved from hard and frictionful to easy (thanks to TEFCA Individual Access Services and the CMS Health Tech Ecosystem), it is obvious that existing consumer-facing touchpoints should embed the capability. From OpenAI Builds a PHR?:
Dominant tech aggregators like Google, OpenAI, and Apple will own the middle of the bell curve by virtue of meeting those patients in the consumer experience they already are. For that population, embedded health or “healthcare that shows up just when you need it” will be the manifestation that wins.
That future is not contingent on building a new app. It’s contingent on putting health data in the one they already have. That’s only possible on healthcare infrastructure that’s easy and ubiquitous
So the actual newness of this announcement is less the what (that they are doing a PHR), since that was already known, but the how (the implementation details) and with whom (what partners did they choose to use).
How OpenAI Implemented ChatGPT Health
In August’s article, I also took a stab at guessing how OpenAI’s health offering would manifest:
Brittany Trang of STAT News released a follow-up piece to the CMS’s Make Health Tech Great Again pledge that did a deep exploration of the conversational AI.
I appreciated it digging into some of the more obscure companies on the list more comprehensively than I did. My main pushback: the article misses the key blocker for OpenAI and other tech giants like Anthropic, Google, Apple, or Meta. The implication in her piece is that the main challenge for OpenAI under this pledge is to create a new health-focused app. But they don’t need that.
That framing reminded me of something Nikhil Krishnan of Out-of-Pocket posted, not necessarily about the CMS pledge, but about the broader consumer and healthcare intersection:
These consumer technology companies don’t need to build an app - they already have one! OpenAI has ChatGPT. Anthropic has Claude. Google has Gemini, AI Search, Bard, and whatever else they’re doing. Their strategic play isn’t to start from scratch, but to embed health data into the products that millions of people already use every day. Launching a separate health app would squander their biggest asset: the massive, built-in daily active user (DAU) base that makes distribution a solved problem.
…
Nothing in it precludes OpenAI from using ChatGPT as their entry into the conversational AI assistant category. They’re already doing the back half of the pledge, as exemplified by the story of the cancer patient in the GPT-5 announcements yesterday. So they just need to connect to CMS Aligned Networks, collect consent, and continue what they’re already doing. Surfacing a “link your health records” button in ChatGPT is trivial, but only once the underlying connectivity to do so is ubiquitous and easy.
So…was I right? Did they build a separate app or did they embed the capabilities into the existing experience?
Once you have access
Select ‘Health’ from the sidebar menu in ChatGPT.
Desktop ChatGPT interface with a left sidebar showing navigation items including New chat, Search chats, Library, Health (highlighted), Codex, and GPTs, while the main panel displays a Health view with a prompt asking how the user is feeling and options to connect medical records or understand lab results.
Structurally, the new Health option is a part of the same ChatGPT UI, unlike other initiatives like Sora. However, the health data itself is stored separately underneath the hood:
Due to the sensitive nature of health data, Health builds on this foundation with additional, layered protections—including purpose-built encryption and isolation—to keep health conversations protected and compartmentalized. Conversations in Health are not used to train our foundation models.
The design here is intentional - keep the ChatGPT user experience, but ensure it’s partitioned.
Where the Data Comes From (and What It Doesn’t Do Yet)
OpenAI handwaves away how data is brought in, listing out multiple partners for data from wearables, providers, and more:
Bring your information in
Bring your medical records and the apps you use to track your health and wellness into Health. You can upload files directly, connect from tools (+) or “Apps” in Settings.
New: Medical Records for lab results, visit summaries, and clinical history
New: Apple Health for health and fitness data, including movement, sleep, and activity patterns (must be on iOS to sync)
New: Function for lab test insights, nutrition ideas, and taking action on your health
New: MyFitnessPal for nutrition advice, macros, and recipes
New: Weight Watchers for GLP-1 personalized meal ideas, recipes, and food guidance
AllTrails to help you find your next hike
Instacart to turn meal plans into shoppable lists
Peloton for suggested workout classes or guided meditations
Look closely at how several data sources are labeled New, but three are not. AllTrails, Instacart, and Peloton are all already part of the ChatGPT ecosystem by virtue of being “Apps”:
Sad Peloton noises
If you use these apps, you simply link the external account via username and password and then treat those parties as an external data source:
Wait, so does ChatGPT Health already exist? The catch here is that the current implementations today do not reference personal information like your prior hikes or workouts from Peloton and can only execute point-in-time actions rather than draw upon any longitudinal history:
In other words, today’s ChatGPT Apps can call Peloton’s APIs to suggest a class, or AllTrails’ APIs to find a nearby hike, but they do so without importing, storing, or reasoning over your past behavior. They operate as standalone, real-time agents, not as contributors to a unified personal health record.
The incentives of building an App are straightforward: every invocation should lead to some downstream conversion, utilization, or engagement that the partner can attribute back to ChatGPT. Instacart wants grocery orders. AllTrails wants trail discovery that leads to premium subscriptions. Peloton wants class starts. Function, MyFitnessPal, and Weight Watchers want sustained program adherence and funnel expansion. In other words, these aren’t neutral data pipes - they are conversion-oriented endpoints wrapped in an SDK.
So that perhaps explains why these apps aren’t contributing longitudinal data into the Health partition today. That kind of deep, bidirectional integration provides little incremental value to their core business objectives and would require them to expose user history in ways that don’t directly support conversion. A lightweight, transactional interaction model is both cheaper to implement and more aligned with how they monetize.
From the partners’ perspective, the assistant is a top-of-funnel channel. There is no clear revenue incentive to let ChatGPT reason over a user’s entire workout history, nutrition logs, or behavioral patterns if the only guaranteed outcome is that the assistant absorbs more context while the partner absorbs more engineering and compliance burden.
You can make your own call, but my feeling is that’s ultimately not all that useful for what most people would like to do with those services. Hopefully, that changes with ChatGPT Health. I want to pull in my AllTrails hike and Peloton workout histories. The text in the launch announcement (“AllTrails to help you find your next hike” and “Peloton for suggested workout classes or guided meditations”) doesn’t really suggest it.
The Real Connectivity Constraints
Taken together, this means the “Health” product is less a wholly new integration surface and more a reclassification of existing and emerging partners under a walled-off data and policy domain. They are using the same App SDK, just funneled into the distinct, compartmentalized “Health” data partition.
Conspicuously, we see the actual patient linking process from both the OpenAI announcement (they handwave past it and don’t touch on app linking at all) or demos posted elsewhere, which just focus on the post-link experience in ChatGPT.
The main paths available today for patient access to data are TEFCA Individual Access Services and Patient Access API aggregation. This reflects the best available connectivity patterns today, but both have their problems:
TEFCA IAS: Avoids the usability pain of portal credentials, but is lacking on coverage as only major EHRs like Epic, athenahealth, MEDITECH, and eClinicalWorks are live
Patient Access API aggregation: Has excellent coverage of providers given certification requirements, but has the usability pain of portal credentials. Additionally, providers on non-certified EHRs will be missing.
An inconsistent experience of mixed usability and coverage holes will be a dissatisfier for some of the initial users once ChatGPT Health is more broadly available (like it is for all PHRs today). This isn’t a knock on any particular vendor - it’s just the state of things today. You can try Fasten’s TEFCA / Patient Access beta to get a rough equivalent of the experience while we wait for OpenAI.
Partners and Power Dynamics
Based on best available knowledge from press releases, documentation, and educated guesses/wishful thinking
Circling back on the partners listed above, it’s worth thinking about how the technical and business arrangements between OpenAI and different partners might be structured. OpenAI ideally wants Apps to be low-friction extensions of the ChatGPT experience that increase stickiness without ceding platform control.
Instacart, AllTrails, and Peloton all followed this paradigm. Of the list of new partners, I imagine Function, Weight Watchers, and MyFitnessPal followed suit there. They will not be the last - there will be a flood of competing solutions that follow suit there. As soon as ChatGPT becomes a distribution surface with measurable intent and conversion pathways, every adjacent wellness, nutrition, fitness, chronic-condition, and diagnostics company will feel obligated to participate.
All good in theory - MCP seems like a cool technology to me. However, has anyone seriously tried the existing Apps available today? I had not, so I took a pass running through five or six of the most popular and applicable.
Google Calendar answered basic questions with data from 2024:
OpenTable simply failed on basic queries:
After struggling to use the App, ChatGPT made statements about my Spotify musical taste that have no basis in fact, including a Paulius Mui lookalike that has 179 monthly listeners:
Despite repeated attempts to use the Booking.com App with it activated, ChatGPT responded with web searches and direct links to hotel websites:
This is all anecdotal, but if the experience of using ChatGPT Health with the ChatGPT App SDK is at all like this, it’s going to be problematic! Inaccuracy, brittleness, and inconsistent invocation pathways undermine the fundamental premise that Apps are a seamless extension of the core assistant. When the model falls back to generic search, hallucinates stale data, or ignores the user’s explicit choice to use an integrated App, it indicates that the orchestration layer between the model, the MCP server, and the UI is not yet stable.
In a consumer health context, that gap becomes untenable. Errors are not just annoying; they create liability, erode trust, and sharply constrain what partners will be willing to expose. In other words: the ChatGPT App SDK is being asked to support a category that demands much higher reliability, determinism, and input fidelity than what the current generation visibly delivers.
I am throwing darts half-blindfolded here, but I would bet OpenAI developed to the specifications of the two remaining higher sensitivity and fidelity data classes in Health. An easy way to derisk is avoid inheriting the brittleness of third-party MCP implementations. When OpenAI forces everything through the App ChatGPT, it becomes downstream of every partner’s quirks and failure modes for a new, unfamiliar paradigm (MCP). If instead OpenAI develops up to established external APIs, it controls far more of the reliability surface, although they have to do more work.
Beyond that, I would be shocked to learn that Apple built up to OpenAI’s specifications. Directionality of API development generally reflects who needs whom more. Apple already operates from a position of structural dominance in consumer health data via Apple Health and device-level capture. It seems far more plausible that the OpenAI Health team developed to the Healthkit API than the other way around.
Where the plot truly thickens, though, is Medical Records. That integration is powered by b.well. To borrow the framing from my earlier article, CMS’s Health Tech Bombshell:
b.well: At first blush, most people think of b.well less as a network and more as a tool for providers, payers, and life sciences to build better consumer experiences (as that’s how they market themselves). They’ve always been heavily invested into and strong proponents of interoperability, but the shift of messaging to being recognized as a network is a good evolution (for them and for any point solution).
This matters here because OpenAI is not integrating a lifestyle App like Peloton or AllTrails. They are integrating a clinical data network. That places b.well in a fundamentally different category from every other launch partner.
OpenAI is a much larger company than b.well - at least 10x in headcount and many multiples higher in terms of revenue, capitalization, and prestige. Again, this is not a knock on b.well. We’re literally talking about a clinical data aggregator and a top-tier AI platform company. The asymmetry is structural, not evaluative. Leverage here goes to OpenAI at a glance.
And thus the twist:
OpenAI is leveraging b.well’s SDK for Health AI to activate this connectivity and accelerate the delivery of AI-driven health experiences.
Now that right there? That’s interesting. From b.well’s website:
More context from the press release announcing the SDK last month:
Backed by connectivity spanning 2.2 million providers, 300+ health plans, and the broadest ecosystem of pharmacy, lab, device, payer, and patient-generated data, the Health AI SDK unifies real-world health information into a single AI-ready dataset. With this foundation, AI assistants can offer personalized guidance based on a consumer’s complete medical history and take action on their behalf, such as scheduling appointments or transferring prescriptions.
“The b.well SDK for Health AI does the back-end data and API integration work so health systems, insurers, life sciences organizations, pharmacies, and others can efficiently, accurately, and securely build AI technology that goes beyond providing answers to recommending the right action — all backed by clinical evidence and detailed medical history,” said Imran Qureshi, Chief AI and Technology Officer, b.well Connected Health.
So what is it? Upon further digging, the b.well API documentation returns us back to where we started - an MCP server:
It’s telling that of the partners, this particular offering is the only one listed in a whitelabeled way:
New: Medical Records for lab results, visit summaries, and clinical history
So my guess is still that OpenAI did the development, but in a way that’s very congruous with their own App SDK (and thus likely lower lift), given it’s all MCP.
That wraps the (already long) perspective on how ChatGPT Health is being assembled structurally, technically, and through the incentives of its earliest partners. For subscribers, we’ll dive a bit deeper on monetization, exclusivity, and winners and losers after the break.
Keep reading with a 7-day free trial
Subscribe to Health API Guy to keep reading this post and get 7 days of free access to the full post archives.
