Apple's 27 wave lands: on-device model APIs, iPad windowing, fewer excuses

What happened

Apple pushed iOS 27, iPadOS 27, and macOS 27 ("Sequoia's successor," internally still called Cascade in some framework headers) to general availability today. The rollout hits the usual matrix: iPhone 13 and later, iPad Air M1 and later plus iPad Pro from 2022 onward, and every Apple-silicon Mac. Intel Macs are officially off the boat — macOS 26 was the last release to carry them, and the deprecation notice in the developer docs is now a hard cutoff.

The headline API is the Foundation Models framework, which exposes Apple's ~3B-parameter on-device model to third-party apps through a Swift-native `LanguageModelSession` type. Structured output, tool calling, and streaming are all first-class. Apple is also shipping a smaller "Writing Tools" model and a vision encoder that developers can call directly rather than routing through the system-level Writing Tools UI. Crucially, on-device inference through Foundation Models carries no per-call cost and no network dependency, which flips the unit economics of shipping AI features in indie apps.

On iPadOS, the long-rumored windowing overhaul is real: freely resizable overlapping windows, a persistent menu bar in the top chrome, and Exposé. Stage Manager is still there but demoted to "one option among several." macOS 27 gets a redesigned Spotlight (now with Apple Intelligence-powered actions), a rebuilt Passwords app with passkey sync improvements, and — quietly — a new `com.apple.private.security.no-container` entitlement gate that tightens what unsigned developer tools can do outside of Xcode's blessed paths.

Why it matters

Strip away the keynote choreography and this release is really about one bet: Apple is trying to make on-device AI the default deployment target for consumer app features. For two years the pattern has been "call GPT-4/Claude/Gemini from your iOS app, eat the token cost, hope the user's connection holds." Foundation Models says: for a large class of features — summarization, extraction, classification, structured JSON generation, short-form drafting — you can do it locally, for free, offline, with no privacy review to write.

The model itself is not going to beat frontier hosted models on reasoning-heavy benchmarks. Apple's own developer session was refreshingly honest about this: the ~3B on-device model is positioned for "focused tasks with clear inputs and outputs," not open-ended chat. But that's most of what apps actually ship. A recipe app doesn't need Claude Opus to parse ingredients into a shopping list; it needs something that runs in 200ms on an iPhone 15 and never gets rate-limited.

The iPadOS windowing story is the other big shift, and it's more consequential than the tech press is treating it. For a decade the standard developer joke has been that the iPad is a beautiful screen attached to an OS that refuses to let you do work on it. iPadOS 27 doesn't fully close that gap — Xcode is still not there, Terminal is still not there — but the operating system now behaves like a computer instead of a large phone. If you write cross-platform tools, your iPad-first users just became a real segment instead of a rounding error.

Compare the trajectory to what's happening on the other side of the fence. Google's on-device Gemini Nano is technically more capable per parameter, but Android's fragmentation means you're targeting a distribution where maybe 30% of active devices can actually run it well. Apple's install base is the opposite: within twelve months, roughly 70% of active iPhones will be on iOS 27, and every one of them has the Neural Engine silicon to run the framework. That's the leverage. A capability that ships on 800 million devices in a year isn't a feature — it's a platform assumption you can build against.

The less-glamorous story is the tightening on the Mac side. The new sandboxing entitlements and the expanded Gatekeeper checks are going to bite anyone shipping developer tooling outside the Mac App Store. If your product is a CLI wrapped in a `.app` bundle, or an Electron app that spawns unsigned helper binaries, budget time this quarter for notarization audits. Apple hasn't broken anything yet, but the direction is unambiguous.

What this means for your stack

If you ship an iOS or Mac app with AI features today, run the numbers on moving your low-risk calls to Foundation Models. Summarization, tagging, entity extraction, and simple rewriting are the obvious wins. Keep your hosted model for anything that needs real reasoning, long context, or up-to-date world knowledge — the on-device model has a knowledge cutoff and a context window measured in low thousands of tokens, not hundreds of thousands. A hybrid router that tries local first and falls back to Claude or GPT on low-confidence outputs is the pattern most shops will land on within six months.

If you build cross-platform mobile tooling, iPadOS 27 is the version where "iPad support" stops meaning "it runs, technically." Real window management means users will drag your app next to Safari, next to Notes, next to a video call — and they'll notice if your layout only has two breakpoints. Test at 320pt-wide floating windows. Test with the menu bar consuming vertical space. The apps that get this right in the first six months will pick up the switchers.

If you ship developer tools for macOS, read the entitlements changelog before you read anything else. The direction of travel — more sandbox, more notarization, more explicit user grants for filesystem and network access — is going to make some existing patterns untenable. Homebrew-installed CLI-in-a-GUI wrappers are the most exposed. Start the notarization work now; don't wait for a macOS 27.2 point release to break your users' installs.

Looking ahead

The interesting question isn't whether Apple's on-device model is good enough today — it's what happens when the pricing gravity of "free local inference" compounds against "$3 per million tokens" over two or three years of app development. If you're an indie developer or a small team, Foundation Models is the first Apple API in a long time that materially changes what you can afford to build. Ship something with it this quarter; the review-time competitive advantage of being early on a new framework has historically been six to twelve months on the App Store, and this one has more surface area than most.