Two things changed recently and they are easy to misread together. Your training data can now reach ChatGPT by two different mechanisms, and the choice between ChatGPT and Claude turns on which of those is open to you rather than on which model writes better prose. One is a connection a device maker built and you add inside the assistant. The other is a health sync that travels through your iPhone. They are open to different people and they carry different things. It is also still only half the question, because nothing on any route holds what you are training for.
If you are choosing between ChatGPT and Claude
Most comparisons argue about plan quality. That is the least decidable part, and the least useful one to decide on: both write competent, periodised plans, and the gap between them is smaller than the gap between a plan built on your real numbers and one built on guesses. What you can check before committing is what each assistant connects to.
Claude reaches Strava. The connector sits in Claude's own directory, so it works on the free plan, and it covers any device that syncs to Strava, which is most of them. It is read-only, it needs a Strava subscription, and Strava says it launched with Anthropic and is looking at other clients later. Claude has no route into Apple Health.
ChatGPT reaches Apple Health. Through ChatGPT Health it can read whatever Garmin, WHOOP, Oura and Strava write there, which is the widest set of data on offer and the only route that carries sleep and daily recovery. The limit is who qualifies: United States accounts, aged 18 or over, on an iPhone. Custom connectors in ChatGPT also need a paid plan, where Claude's free plan is enough.
COROS connects directly to both, so a COROS athlete can decide on other grounds entirely.
Setting COROS aside, the short version is geographic. Outside the United States, or without an iPhone, Claude currently reaches more of your training. Inside it, with an iPhone, ChatGPT Health reaches more of your physiology. Neither reaches what you are training for.
COROS: the one direct connection
COROS is the only device maker Athmex has verified as publishing a direct MCP endpoint for ChatGPT. You add it in ChatGPT's settings as a custom app, sign in to your COROS account, and the assistant can read your training directly. On which plan you need, the two sources currently differ: OpenAI documents read and fetch MCP connections in Developer Mode for Pro users and does not list Plus, while COROS's own guide refers to ChatGPT Plus. Check what your own plan actually offers before counting on it. The step-by-step setup for both assistants is on the COROS page.
COROS also lists an app directly inside ChatGPT's plugin directory and Claude's connector directory, discoverable without pasting the server URL yourself. Installing from ChatGPT's plugin directory does not require turning on Developer Mode, unlike the manual custom-connector route described below.
What comes through is substantial: activity records and lap data, sleep, HRV, resting heart rate and stress, training load, VO2 max, race prediction times, threshold pace and pace zones, and your training calendar.
Two limits are worth knowing before you lean on it. Requests for the underlying activity files, the ones carrying GPS traces and split-by-split detail, are capped at fifty per calendar day. And cycling power sits outside the standard query set, so a power curve or an FTP estimate is not something the connection can produce, even with a power meter recording.
COROS warns that Developer Mode turns memory off
Connections like this go through Developer Mode, and COROS's own guide is blunt about a cost: it says Developer Mode disables memory. That is COROS's statement, not OpenAI's. OpenAI's Developer Mode documentation does not mention memory at all, and its Memory FAQ does not list Developer Mode among the things that turn memory off. Treat it as a caution worth testing on your own account rather than a settled platform behaviour.
For a training assistant that is an awkward trade. Memory is what stops you re-explaining yourself, and ChatGPT's memory has become genuinely capable at carrying context between conversations. If COROS is right, turning memory off to gain your activity history leaves you with an assistant that can see every run you have done and still has no idea what you are training for.
COROS recommends Claude for this reason, and that is fair advice rather than a slight. It is worth weighing if you have not already settled on one assistant.
Apple Health is the second route, and it is not a connector
Garmin, WHOOP, Oura, Strava and Apple Watch still publish no ChatGPT connection of their own. What changed is at OpenAI's end. ChatGPT Health can read Apple Health, and each of those apps can write into Apple Health, so their numbers can reach ChatGPT without any of them shipping a connector.
Those two routes behave differently. A connection like COROS's is a live link to your account: ChatGPT asks it a question and gets whatever the account holds. The Apple Health route is a copy of whatever each app has chosen to write into your iPhone's health store. OpenAI says what arrives varies by app and by metric, and that some proprietary scores may not transfer. Recovery, Body Battery and Readiness are vendor calculations, and if the app does not write that exact value into Apple Health, ChatGPT will not see it.
Turning it on
First switch on Apple Health sharing in the device app. OpenAI documents the path for each one: in WHOOP, More, then App Settings, then Integrations, then Apple Health. In Oura, the menu icon, then Settings, then Apple Health under Data Sharing. In Garmin Connect, More, then Settings, then Connected Apps, then Apple Health. Then open Health from the ChatGPT sidebar, choose Connect Health and pick Apple Health.
Who can actually use it
This is the part worth checking before you plan around it. OpenAI documents Health as available to logged-in Free, Go, Plus and Pro users in the United States who are 18 or over, on web and iOS, and says it is rolling out gradually. Connecting Apple Health requires an iPhone. If you are outside the United States, or on Android, this route is not open to you yet, and the Strava position below is still the one that applies.
For Claude, every other device still goes through Strava
For Claude, nothing has changed. Garmin, WHOOP, Apple Watch, Polar, Suunto, Wahoo, Zwift and Peloton have no connection of their own, and each reaches Claude by syncing into Strava first. Strava's connector supports Claude only. Athmex will not point you at an unofficial workaround, and the full list of routes checked against first-party documentation is on the official connectors page.
Three options, depending on where you are
None of these is a workaround. They are different trades, and which one fits depends on where you live, what you wear, and what you actually ask an assistant to do.
Connect Apple Health, if it is available to you. On a United States account with an iPhone this is the shortest path to ChatGPT knowing what your body did, and it covers Garmin, WHOOP and Oura equally, because the sync happens at the Apple Health layer rather than device by device. Check which categories your app actually writes before assuming a particular score made the trip.
Add Claude alongside ChatGPT. Claude's free plan is enough for the Strava connector, because it comes from Claude's own directory rather than a custom connection. You keep ChatGPT for everything you already use it for and open Claude when a question needs your real sessions. Beyond the Strava subscription the connector requires, this costs nothing.
For the exact setup steps, see how to connect Strava to Claude.
Or stay on ChatGPT without activity history. This is more reasonable than it first sounds. If what you mainly want is a plan shaped around your constraints, a taper sketched out, or a session explained, the assistant does not need to read every run to help. It needs your numbers, and those you can supply yourself. That is what Athmex does, and it connects to ChatGPT whether or not your device ever gets an activity route.
What no connector supplies, on either route
Every route above has the same shape. It reports what your training did. None of them has a field for what you are training towards.
Your goal race and its date, the time you are chasing, an injury that changes what you can do this month, a threshold you tested yourself and want treated as current: no activity connector has anywhere to put any of it. That holds on the COROS route as firmly as the Strava one.
For a ChatGPT user on the COROS route it may bite harder still, if COROS is right that Developer Mode also costs you memory. The assistant would end up with your data and no sense of what it is for.
That is the half Athmex holds. You enter your thresholds, PBs, goals, races, injuries and equipment once, each with the date it applied, and connect Athmex to ChatGPT or Claude. Athmex does not read your activities or calculate anything. A value changes when you change it, and stays as you entered it until then. See how Athmex works.
What to actually do
If you use COROS, add the COROS connection for your training history and Athmex for the context around it. Between them ChatGPT has both what you did and what it is for, which matters all the more if Developer Mode has also cost you memory.
If your device routes through Strava and Apple Health is open to you, try that first: it is fewer moving parts than adding a second assistant. If it is not, decide whether adding Claude is worth it. If it is, Claude gets your activity history and your context together. If it is not, connect Athmex to ChatGPT and accept that you are supplying the numbers rather than the sessions. That is still well ahead of retyping your PBs into every new conversation.
If you use Oura, Apple Health is the route that applies to you, and Oura is one of the apps OpenAI names by name. There is still no Oura connector for either assistant, and on Claude there is no Oura route at all.