Apple Just Let Any App Hand a Prompt to Claude Through the Foundation Models Framework. If You Ship for iPhone, Read the Bring-Your-Own-Key Math First.

Apple spent the last two years keeping its on-device models walled off from the big cloud providers. At WWDC this month it stopped. The Foundation Models framework (the thing that lets your app call Apple Intelligence on-device) can now route a request out to an external model like Claude, with you supplying the API key. Xcode 27 picks up model choice too, so the agent inside your editor can be Claude, Gemini, or whatever implements Apple's new language-model protocol.

For a solo iOS developer this is genuinely useful and quietly expensive, in that order. The on-device model is free and private and runs on the user's silicon. The moment you escalate a request to Claude, the meter is yours. That split is the whole decision, and the App Store doesn't make it easy to pass the cost through.

What actually shipped

Let me separate what Apple announced from what's been reported, because the reporting on this got ahead of the release more than once this spring.

The concrete developer pieces, per Apple's own newsroom and Anthropic: WWDC 2026 turned the Foundation Models framework in iOS 27, iPadOS 27, macOS 27, and the rest of the 27 family into a single Swift API that covers on-device, Private Cloud Compute, and external models. The mechanism is a public LanguageModel protocol that any provider can implement. Anthropic ships an official Swift package for Claude; Google routes Gemini through its Firebase SDK. The nice part for builders is that your LanguageModelSession code stays the same across all of them: you swap providers by changing a dependency, not by rewriting your call sites, and you authenticate with your own provider key, your Anthropic key for Claude. Apple also said it will open-source the framework later this summer. Xcode 27 picks up the same model-choice idea for its coding assistant, so you're not locked to one vendor's agent inside the IDE.

There's also a consumer-facing side (an Extensions system that will eventually let Siri, Writing Tools, and Image Playground tap Claude, Gemini, or ChatGPT), but that's the part most likely to shift before it ships, and it's not where a solo builder makes money. The developer framework is the actionable piece.

The 27 releases land in the fall, on Apple's usual September cadence. So this is a "decide your architecture now, ship into it later" situation, not a "rewrite this weekend" one.

The pattern that makes sense for a small app

The right way to read this isn't "I can finally put Claude in my app." It's "I now have a free, private first tier and a paid, powerful second tier, and I get to decide what crosses the line between them."

The on-device Foundation Model is small, but it's real, and it costs you nothing per call. For a lot of in-app AI (classifying a note, summarizing a short message, extracting fields from text the user typed, rewriting a sentence) it's enough. You should treat it as the default and escalate to Claude only when the task genuinely needs the bigger model: long-context reasoning, multi-step tool use, anything where the on-device result would embarrass you.

That tiering does two things at once. It keeps the user's data on their device for the easy cases, which is a real privacy story you can put in your App Store listing. And it keeps your Anthropic bill tied to the hard requests instead of every tap.

Where the bill bites

Here's the part nobody puts on the WWDC slide.

When you bring your own Anthropic key, every escalated request is billed to you at API rates. Your heaviest, happiest, most engaged users (the ones hammering the feature you're proudest of) are also your largest line item. That's the opposite of how a flat-rate subscription app is supposed to work, where your best users cost you roughly nothing more than your worst.

And you can't cleanly pass it through. If you sell a one-time unlock or a flat monthly subscription through the App Store, Apple takes its cut and you eat the variable inference cost underneath a fixed price. There's no native "charge this user for the 4,000 tokens they just spent" primitive in StoreKit. So either you cap usage, or you price in a buffer for the power users, or you watch a small number of accounts quietly turn your margin negative.

None of that is a reason to skip Claude. It's a reason to instrument it from day one. Log token usage per request before you ship the feature, not after the first surprising invoice. Know your cost per active user, set a sane per-user ceiling on escalated calls, and decide deliberately whether the Claude tier is a free perk you're subsidizing for retention or a paid tier you actually meter.

The honest counter-take

I could be too cautious here. For a lot of apps the escalation rate will be low: most users do the easy thing, the on-device tier eats most of the volume, and the Claude bill stays small enough to be a rounding error against a healthy subscription. If that's your usage shape, the cost worry is overblown and you should just build the better feature.

The risk isn't average cost. It's the tail: the handful of accounts that use the expensive path constantly. If your app's whole value is the heavy AI feature, those users aren't the exception, they're the product, and the math gets real. Know which app you're building before you decide how much to worry.

What I'd actually do before the fall

If you ship iOS, spend an afternoon this summer on three things. Prototype the on-device Foundation Model against your actual use case and find out honestly how much of your AI work it can carry, it's probably more than you'd guess. Draw the line for what escalates to Claude, and put a per-user cap behind that line before a single beta tester touches it. Then decide your pricing posture deliberately: free subsidized perk, or metered paid tier, with the cost-per-active-user number in front of you instead of a vibe.

Do that now, while it's a calm design decision. The builders who get burned by bring-your-own-key economics won't be the ones who modeled the bill in July. They'll be the ones who shipped the magic feature in October and read about their own power users in a December statement.

Author

Lukas

@lukcombinator

Sources

• Apple aids app development with new intelligence frameworks and advanced tools (Apple Newsroom)
• Apple's Xcode now supports the Claude Agent SDK (Anthropic)
• iOS 27 will let you choose between Gemini, Claude, and more for AI features (9to5Mac)
• Apple bakes AI agents into Xcode with Anthropic and OpenAI (CNBC)