Whenever Apple hands developers something for free, it pays to read the terms twice before opening the champagne. iOS 27 landed on 14 September, and this year's gift is a big one: your app can talk to Apple's own AI model, either on the phone or on Apple's servers, without you paying a cent or pasting an API key anywhere. If you have been putting off "that AI feature" because of the server bill, this is worth a coffee and ten minutes of your time.
The phone does the thinking (up to a point)
The on-device model has been open to apps since iOS 26, through a framework with the very modest name of Foundation Models. It runs on the iPhone itself, works in airplane mode, costs nothing per request and never sends your users' text anywhere, which is exactly the kind of privacy story your lawyer likes to hear. For iOS 27, Apple rebuilt it: it is better at logic, better at calling your app's own functions, and it can now look at images, so "what is on this receipt?" becomes a perfectly reasonable question to ask it.
What it is not is a genius with a large memory. Its context is short, a few pages of text at most, so it shines at summarising a note or sorting a list and is rather less brilliant at digesting a forty-page contract.
Then the cloud, free, with an asterisk
Here is the real news. Third-party apps can now send requests to Apple's bigger model running on Private Cloud Compute, with a 32K-token context and several levels of reasoning. There is no account to create and no key to leak in a public repo, because the user's device takes care of authentication. Each user gets a daily allowance of requests, and when they run out, Apple will kindly suggest a bigger iCloud+ plan (Apple being Apple).
You, the developer, pay nothing, as long as three things are true: you are enrolled in the App Store Small Business Program, your apps add up to fewer than 2 million first-time downloads, and you have asked Apple for the entitlement.
Now the asterisk. If one of your apps goes past 2 million, or you leave the Small Business Program, Apple lets you know and gives you six months to move to something else, and the page mentions no paid plan that would let you stay. In other words, success gets you evicted. It is the nicest problem an app can have, but it is still one you want to plan for on day one rather than on the morning your app goes viral.
The escape hatch is already built in
Apple seems to know this, because the framework now has a common protocol that any model can plug into, and Apple announced that Anthropic and Google publish Swift packages for it (those you pay per token, to them, so the free lunch ends there). Your screens and your prompts stay the same when you swap the engine underneath, and moving from the on-device model to the cloud one is, quite literally, one line:
func makeSession() -> LanguageModelSession {
if #available(iOS 27.0, *) {
let cloud = PrivateCloudComputeLanguageModel()
if cloud.isAvailable {
return LanguageModelSession(model: cloud)
}
}
return LanguageModelSession() // on-device model
}
The sensible route looks like the diagram above: try the model on the phone first, step up to Private Cloud Compute when the task needs more context or some real reasoning, and when the network is gone, the daily quota is used up or the device simply cannot run Apple's AI, fall back to the local model or to a screen that does without it. Apple itself recommends retrying on the phone when a cloud request fails for lack of network, and showing the quota in the screen rather than in yet another alert nobody reads.
The guests who were not invited
This is the part the keynote does not linger on. Both Apple models only run on devices that support Apple Intelligence, meaning the iPhone 15 Pro and the iPhone 16 or later, in a supported region and language. Someone on an iPhone 13 gets nothing, and your Android users were never on the guest list to begin with. So the AI feature has to stay a bonus, or you need a server solution behind it for everyone else. The check itself is short, and skipping it is how you ship a lovely button that does nothing on half your users' phones:
import FoundationModels
func summarize(_ text: String) async throws -> String? {
let model = SystemLanguageModel.default
guard case .available = model.availability else {
// Older iPhone, Apple Intelligence turned off, or model still downloading:
// hide the feature or offer another path.
return nil
}
let session = LanguageModelSession(
model: model,
instructions: "Summarise the text in three short sentences."
)
let response = try await session.respond(to: text)
return response.content
}
And if you go with a third-party model, App Store guideline 5.1.2(i) asks you to say clearly when personal data goes to a third-party AI and to get explicit permission first. A polite sentence and a toggle, nothing heroic, but the reviewers do read it.
So what would I actually do?
I would pick one feature that saves people real time (not a shiny "AI" badge for the store listing) and try it on the on-device model first, which is also what Apple recommends before reaching for the cloud. If it falls short and you qualify, request the entitlement now and test through TestFlight, since test installs do not count towards the famous 2 million. I would also design the fallback screen before the pretty one, because that is the screen a good share of your users will actually see. And whatever you decide, the upgrade is coming for you anyway: from April 2027, App Store uploads will need the iOS 27 SDK, the same way the Xcode 26 rule caught a few people by surprise this year.
If you want help
I add AI features to iOS and Android apps, privacy and store rules included, and it is all on the AI features for mobile apps page. If you would rather know where your app stands before adding anything clever to it, the five-day health check is made for that.
Sources: What's new in the Foundation Models framework (WWDC26), Build with the new Apple Foundation Model on Private Cloud Compute (WWDC26), Adding server-side intelligence with Private Cloud Compute, Private Cloud Compute eligibility, Apple Developer news, App Review Guidelines, Apple Newsroom, 14 September 2026.