iOS 27 Is Here: The WWDC 2026 Features That Change Mobile App Development

Apple used its WWDC 2026 keynote on June 8 to do something it rarely does: put developers at the centre of the story. iOS 27 and the Apple-silicon-only macOS 27 — named Golden Gate — arrive this autumn with a redesigned, AI-powered Siri on top. The change that actually rewires how mobile app development works, though, sits underneath the marketing: a rebuilt on-device model, a new protocol that lets you swap in Claude or Gemini without touching your code, and App Intents promoted to the single front door for Siri.

What happened

At WWDC 2026 Apple announced iOS 27, iPadOS 27, macOS 27 (“Golden Gate”), watchOS 27, visionOS 27, and tvOS 27, all shipping to the public in the autumn after a summer of developer and public betas. The keynote’s headline was a redesigned Siri — a more personal, more contextual assistant powered by a new on-device model, available both as a standalone app and woven through the system, with access gated behind a waitlist in the developer betas and limited to iPhone 16 and later.

For builders, the more important announcement was the one Apple published to its developer newsroom the same day: a batch of new intelligence frameworks and tools that open Apple’s on-device model to third-party apps (Apple’s developer newsroom post laid out the full set). Apple says iOS 27 carries “over 250 changes,” from a personalisation slider for the Liquid Glass interface to a rebuilt systemwide search. macOS Golden Gate, which Apple framed in the performance-first spirit of Snow Leopard, is the first version of macOS to run exclusively on Apple silicon and the last to ship full Rosetta 2 support (MacRumors’ WWDC 2026 recap catalogued the consumer-facing list).

Why it matters for builders and founders

Until now, putting an AI feature in your app meant signing up with a cloud provider, routing every request to their servers, and paying per token forever. iOS 27 changes that maths in three concrete ways. First, the on-device Foundation Models framework is free at the call level — your app talks to the model already on the user’s phone. Second, developers enrolled in the App Store Small Business Program (under two million first-time downloads) get access to Apple’s next-generation models running on Private Cloud Compute at no cloud API cost — a genuinely new free tier for indie shops and white-label app builders. Third, a new model-provider protocol means you are no longer locked to Apple’s model at all.

The App Intents change cuts the other way. App Intents is now the only supported path for Siri to call into a third-party app, and Apple gave SiriKit a formal deprecation notice with a two-to-three-year runway. That means migration work for any app with a Siri integration. The upside is distribution: once your actions live in App Intents, users can ask the new Siri to run them by voice or from the system, which is a discovery channel most apps have never had.

The details, in plain English

“On-device inference” means the AI model runs on the iPhone’s own chip with no network round-trip to a server. The trade-off is that the model is smaller than a frontier cloud model like Claude or GPT-5, because it has to fit in the phone’s memory and thermal budget. In iOS 27 Apple rebuilt that model from the ground up — better reasoning, better tool calling, tighter guardrails — and, for the first time, gave it vision. The Foundation Models framework now accepts images directly (UIImage, CGImage, CoreVideo pixel buffers, and file URLs), which unlocks on-device receipt parsing, visual product search, and barcode reading with no server call.

The pieces a developer actually touches in iOS 27:

• Foundation Models framework — the on-device model with streaming, structured output, tool calling, and now image input.

• The LanguageModel protocol — a common Swift interface that any model provider implements. Write your feature once and swap between Apple’s model, Claude, or Gemini without changing the surrounding code. Anthropic and Google integrations arrive with iOS 27 and macOS 27.

• Private Cloud Compute — Apple’s privacy-preserving cloud tier for prompts the on-device model cannot handle, free for Small Business Program developers.

• App Intents + View Annotations — a new API that maps your on-screen views to entities, so users can reference and act on what is on screen conversationally. An App Intents Testing framework validates these through real system pathways instead of brittle UI automation.

Where the on-device model still falls short, in 2026, is anything needing fresh world knowledge or long multi-step reasoning. For those, the discipline is unchanged: route to a cloud model and treat the on-device path as the fast, private, offline default. The earlier story of Apple’s on-device push is worth the context in our Apple Intelligence review.

The bigger picture

Apple is doing what only Apple can: bundling good-enough AI into roughly 1.5 billion devices people already own and trust, free, privacy-first, and letting third-party developers do the creative work. The most telling move is the LanguageModel protocol. By making it trivial to plug Claude or Gemini into an app through a standard interface, Apple is quietly conceding that it will not have the best model — and positioning itself as the integration layer instead of the model vendor. For founders choosing infrastructure, that is the friendliest possible posture: you get a free on-device baseline and a clean swap to a frontier model when you need one. The same swap-without-lock-in logic is reshaping cloud LLM choices too, which we covered in our Claude 4 vs GPT-5 comparison.

macOS Golden Gate carries a quieter mandate. Dropping Intel support and sunsetting Rosetta 2 means any Mac app still shipping an x86 build has a deadline now. The performance gains Apple advertised — quicker AirDrop, faster network file browsing, better Spotlight suggestions — are the carrot; the Apple-silicon-only requirement is the stick.

What to watch next

Three things to track between now and the autumn general-availability release. First, adoption: whether marquee apps ship Foundation Models and App Intents integrations during the beta, which will tell you how stable the developer ergonomics really are. Second, fragmentation: two of the flashiest features — more expressive Siri voices and the big systemwide dictation accuracy gain — require 12GB of memory, which limits them to the iPhone Air, iPhone 17 Pro line, and a handful of iPads and Macs. Building for “Apple Intelligence” will increasingly mean building for a capability tier, not a single OS version. Third, App Store policy: watch for the review and disclosure rules around apps that route to cloud models, which Apple has signalled but not fully detailed.

For builders, the practical move this summer is small and concrete: install the iOS 27 SDK, prototype one feature on the Foundation Models framework, and audit any SiriKit code for the App Intents migration. iOS 27 is the first release where shipping an AI feature does not automatically mean signing up for a recurring cloud bill, and that alone is worth a week of prototyping time.