At WWDC on June 8, 2026, Apple did the thing it never does: it admitted it could not build the model. The rebuilt Siri runs on a custom 1.2-trillion-parameter Gemini model licensed from Google for roughly $1 billion a year, processed on Apple's own Private Cloud Compute servers [1][2]. That is the headline everyone is writing. It is not the part that matters if you build software.
The part that matters is a quieter feature called Extensions. In iOS 27, the user, not you, gets to choose which model powers Apple Intelligence: ChatGPT, Gemini, or Anthropic's Claude, switchable in Settings, each with "its own distinct voice so users know which one responded" [3]. Claude is an Apple Intelligence option for the first time, which means it now reaches the 2-billion-plus devices in Apple's installed base.
Read that again from a builder's seat. For three years, choosing the model was your decision. You picked GPT or Claude or Gemini, you tuned your prompts to its quirks, you shipped. As of iOS 27, if your app plugs into Apple Intelligence, the user picks the model, and they can change it whenever they want. Your prompt now has to run on a model you did not choose and cannot predict. Prompt portability just stopped being a best practice and became a platform requirement.
A mock iOS Settings screen for the Apple Intelligence default model, offering ChatGPT, Gemini (selected as default, powering Siri) and Claude, with a note that the user can switch anytime
What actually shipped
Strip the keynote theater and here is the developer-relevant core:
Siri is rebuilt on Gemini. A custom 1.2T-parameter Google model, about $1B/year, running on Apple Private Cloud Compute. The deal was first reported in January 2026 and confirmed on stage [1][2].
Extensions let the user pick the model. Apple Intelligence features (Siri, Writing Tools, Image Playground) can be routed to ChatGPT, Gemini, or Claude, chosen by the user in Settings [3].
It is built on App Intents. The way your app exposes actions to Siri and Apple Intelligence is the existing App Intents framework. Adopt it and your app's capabilities become invocable by whichever model the user selected.
It ships in the fall. iOS 27 reaches the public in fall 2026, with developer and public betas over the summer. You have one summer to get ready.
None of these numbers are mine to invent, and I am not citing any KMP data here: every figure above comes from Apple's keynote and the reporting around it.
Want to know how effective your prompts are? Prompt Score analyzes them on 6 criteria.
WWDC is the same trend reaching a place I did not expect this fast: the consumer operating system. The most controlling company in tech, the one that owns its silicon and resisted opening anything for a decade, just turned the underlying model into a user-facing toggle. When Apple makes the model a Settings switch, the model is a commodity. There is no stronger signal available.
But the consequence for builders is sharper than "the model is a commodity." It is this: the abstraction boundary moved. Before, the boundary was between you and the model; you owned both sides. Now the boundary runs between your app and a model the user controls. You write the prompt. Someone else decides what runs it. And they can change their mind on a Tuesday.
Before and after comparison of who picks the model: until iOS 26 the builder picked and owned both sides of the boundary; with iOS 27 Extensions the user picks the model in Settings and can change it silently, moving the boundary between your app and a model the user controls
Why this is harder than "just works"
The optimistic reading is that these models are all good now, so it does not matter which one answers. That is wrong, and Apple's own design admits it: they gave each model "its own distinct voice so users know which one responded" [3]. Different voice is the visible tip. Different behavior is the iceberg.
I keep watching the same failure repeat with every release. A prompt tuned to extract one confident answer behaves differently on a model that was trained to flag uncertainty. A prompt that relied on a model's default verbosity breaks on one calibrated to be terse. Ask three models to "respond in JSON, no preamble" and one of them still wraps the object in a friendly sentence, and your parser falls over in production.
Now put that under iOS 27. Your app sends the same App Intent prompt, but on Monday it runs on Gemini and on Friday the user switched to Claude, and you get no event, no migration window, no changelog. The behavior of your feature changed because someone touched a setting three screens deep, and the first you hear about it is a support ticket that says "the summarizer got weird." The swap is silent, it is the user's, and it is now part of your runtime.
The techniques you're reading about work. Test your prompts now with Prompt Score and see your score in real time.
This is the asymmetry from the harness piece, except worse: there, you chose when to swap and could test first. Here, the swap is initiated by a stranger and you find out afterward.
The flip side: this is also distribution
It is worth saying the optimistic half out loud, because the constraint comes with a real opening. Apple is adding an Extensions section to the App Store, and an app that adopts App Intents becomes something Siri can invoke directly, on demand, without the user ever opening it [3]. For a solo dev, that is a distribution channel into 2-billion-plus devices that did not exist a week ago, and the barrier to entry is not a marketing budget. It is whether your feature behaves correctly when the assistant calls it.
That reframes the testing work from chore to moat. If your prompts are robust across all three providers, you can ship into this new surface with confidence while competitors who hard-tuned to one model spend the fall firefighting voice-by-voice regressions. The platform is handing reach to whoever did the portability homework. The price of the ticket is exactly the discipline most teams skip.
Write App Intents around intent, not one model's habits. State what you want and the hard constraints. Keep the "think step by step" and "no preamble" scaffolding in a layer you can adjust per provider, because that scaffolding is tuned to one model and the user can pick a different one.
Test every prompt against all three providers, not your favorite. ChatGPT, Gemini, and Claude are all reachable targets the moment you adopt Extensions. A prompt that is excellent on the one you developed against is the only data point that does not matter, because it is the one case the user might not pick. Score the same prompt on each and fix the gaps before the beta.
Pin the contract, not the prose. If a downstream parser depends on strict JSON, enforce it with a schema and validation, not with a politely worded request that one model honors and another ignores. The model is now a variable; your output contract cannot be.
Keep one source of truth for the prompt. When the same prompt is forked across three model-specific variants in three files, you have lost the ability to fix it once. Hold one authoritative version, note which models it has been validated against, and update it in one place.
Treat "model" as user state, not config. Architect as if the backend can change between two calls, because under Extensions it can. Defensive parsing, graceful fallback, and a test that runs your critical prompts across providers in CI are no longer gold-plating.
The work is unglamorous and it is mostly testing and discipline, which is exactly the kind of thing that is easy to skip until a platform forces it. Apple just forced it, with a fall deadline.
The signal
For years the safe assumption was that you owned your stack down to the model, and prompt portability was a hedge for the day you might switch vendors. WWDC 2026 ended that assumption for anyone building on Apple Intelligence. The model is now a user setting, the user will flip it, and your prompts have to keep working across whatever they choose.
The teams that treated their prompts as managed, portable, tested artifacts were already ahead. As of June 8 they are simply correct, because the platform now demands what they were already doing. Pick a model to develop against, sure. But build as if the user will replace it, because on iOS 27 they literally can, from Settings, without telling you.
Keep My Prompts lets you keep one authoritative version of each prompt, score it on six quality criteria, and compare how it holds up across models before a user, or a platform, swaps the one underneath you. Free to start, no credit card required.
