Siri has a daily allowance now.

Apple seeded the iOS 27 release candidate on 9 September, the same day as the iPhone 18 Pro event, and the public release is Monday 14 September. The headline feature is the new Siri AI, with its own app and a Dynamic Island invocation. The more useful document for anyone who builds software went up alongside it, a support page titled Learn about Apple Intelligence usage limits.

It lists five things that now have daily limits. Siri AI. The intelligent photo editing tools, including Clean Up, Extend and Spatial Reframing. Image Playground. The AFM 3 Cloud and AFM 3 Cloud Pro models in Shortcuts. And the line we read twice: developer apps using Apple Foundation Models in Private Cloud Compute. Limits vary by feature, request complexity, system demand, system policies and other factors. Hit one and the feature comes back after a time out period. Increased access will be available for a fee, with no price given yet.

The meter was always coming.

None of this is a scandal. Server-side inference costs real money on every request, and nobody absorbs that bill for ever across every iPhone that asks a cloud model for something. Every AI product with a free tier has an allowance somewhere, usually buried. What is new is that the operating system itself now has one, written down, and that it reaches into third-party apps. If your feature calls the platform's cloud model, your user's allowance is now part of your feature.

Look at what the page does not say. It publishes no numbers. It does not say whether a long Siri session draws on the same budget your app does. And one of the listed variables is system demand, which means the limit can tighten on the busiest day of the year. That is exactly the day a retail or events client needs the thing to work.

How we are treating it.

We already design around rate-limited APIs, and a cloud model with an allowance is one more of those. These are the rules we are applying to client work from this week.

• 01The feature does its job without the model. AI shortens the path, it is never the only path. A receipt still gets filed if the summary fails, and a booking still goes through if the suggestion never appears.
• 02On-device first for the common case. The limits page covers the features that run on Apple's servers. Short, frequent, predictable tasks go to the local model, and the cloud model is kept for the occasional heavy request where a wait is acceptable.
• 03Say what happened in plain words. A spinner that never resolves is the worst possible response to a time out. Tell the person the smart option is unavailable for now and hand them the manual one.
• 04Nothing front-of-house depends on it. A check-in desk, a queue screen, a kiosk on a shop floor. If system demand can switch a feature off, that feature stays out of the critical path.
• 05Test the limit on purpose. Nobody running a release candidate will ever hit a production allowance, so we stub the time out and walk the whole flow with the model gone.

The cost moves to the customer.

There is a commercial side too. Calling the platform model is attractive precisely because the inference looks like Apple's cost rather than the client's. Once increased access is something the user pays for, that cost has not gone away. It has moved onto the customer's own subscription, and the heaviest users of a product feel it first. For a consumer app that can be fine. For a staff tool that is open all day, it is a support ticket waiting to happen, and we would rather pay for our own inference and know exactly where the ceiling is.

The practical advice for this week is short. If you have an app on the store that calls Foundation Models in Private Cloud Compute, find out on Monday what it does when the answer is not now, before your users find out for you.