Apple made the model a setting

Apple spent forty years controlling its whole stack — chip, OS, app store, the screws. So it means something that at WWDC on June 8 the company handed Siri's brain to Google, rebuilding the assistant on Gemini under a deal reportedly worth about a billion dollars a year. And then it went further: in iOS 27 you can pick Claude or ChatGPT instead, and even route different jobs to different providers — Claude for code, Gemini for research.

Read that twice. The most vertically integrated company on earth just turned the AI model into a dropdown menu.

The model became a config value

For a year the industry argued about whose model is smartest. Apple just answered a different, better question: how much should that even matter to your product? Their answer is not enough to marry one. Siri is the interface, the personal context, the on-device privacy boundary — and the model is a swappable component plugged in behind it. If the most controlling product company alive won't hardcode a single model, why would you?

This isn't Apple being indecisive. It's Apple being right. The model is the fastest-moving, fastest-depreciating part of your whole system. Pinning your architecture to one vendor's frontier model in June is a bet that expires in July.

This is model-agnostic architecture, shipped to a billion phones

Builders keep relearning the same lesson, and Apple just taught it to 1.4 billion devices at once: don't let the rest of your system know which model answered.

The way I build, the model sits behind a thin adapter. Everything upstream — the prompts, the orchestration, the tool calls, the UI — talks to an interface, not to OpenAI or Anthropic or Google. The provider is configuration, not architecture. When a better model ships in six weeks (and on the current cadence, one will), switching is a config change and a fresh eval run, not a rewrite.

The teams that wired their product directly to one provider's SDK, quirks and all, are the ones who now treat every model upgrade as a migration. The teams that put a boundary there treat it as a Tuesday.

What "swappable" actually takes

Apple makes it look like a settings toggle. Under the hood, model-agnostic isn't free — but it's cheap if you build it in from the start:

• An adapter at the boundary. One interface your app calls; one place per provider that implements it. Nothing else in the codebase names a vendor.
• Evals that run across models. You can only switch on evidence if you can measure each candidate on your tasks. Without that, "swappable" just means "guessing in a new font."
• Prompts that aren't married to one model's tics. The more you over-fit to one provider's exact formatting, the higher the cost to leave. Keep the contract general.

Do those three things and the model genuinely becomes a setting — you change it when the numbers say to, not when a migration project clears.

The bottom line

Apple didn't pick a winner in the model wars. It refused to play, and made the model a part you can replace without touching the product around it.

If Apple — the company that controls everything — treats the AI model as a swappable component, that's the strongest possible endorsement of building model-agnostic. Put an adapter at the boundary, keep cross-model evals, and don't marry one provider's quirks. The model is the part most likely to change. Build so it can.