- Apple is making Foundation Models inference on Private Cloud Compute free for developers in the Small Business Program whose apps have under two million first-time lifetime downloads, removing per-token cost for most App Store apps.
- The Foundation Models Swift API now acts as a server-side router to multiple backends, including Claude and Gemini, positioning Apple as the default integration layer rather than just a model provider.
- The framework expansion (image input, Dynamic Profiles for multi-agent workflows, and a separate Core AI for custom on-device models) signals a broader platform push, not a minor pricing tweak.
- Free inference changes the decision from whether to ship AI at all to which model to use and how much dependency you’re taking on via Apple’s APIs and infrastructure.
- The convenience of a unified Apple interface can create lock-in, so developers should treat it as a default path, not necessarily a permanent one.
Apple now gives developers with under two million lifetime App Store downloads free access to its Foundation Models on Private Cloud Compute, which removes per-token cost as a reason not to ship an AI feature for most apps on the store.
The same Swift API now routes to Claude and Gemini server-side, so Apple is positioning itself as the layer your app calls rather than the model you call. The free tier is the on-ramp to that position.
Free changes the question from "can we afford inference" to "which model, called through whom, and how locked in are we." Treat the Swift API as a convenient default, not a permanent home.
At the 2026 Platforms State of the Union, Apple did something that sounds like a footnote and is actually a positioning move. Developers enrolled in the App Store Small Business Program, whose apps have fewer than two million first-time downloads, can now call Apple Foundation Models running on Private Cloud Compute at no cloud API cost. Re-downloads and updates do not count toward the threshold. By Apple's own framing, that covers the vast majority of apps on the store.
Read that again with a builder's eye. For most of the App Store, the cost of running a model just went to zero.
What actually shipped
The headline is the free tier, but the framework picked up three other things that matter more than the price.
The Swift API now does server-side routing to third-party models. You can call Claude or Gemini through the same interface you use to call Apple's own model. One API, several backends.
The framework gained image input, so it is no longer text only. And there is a new Dynamic Profiles system for building multi-agent workflows, plus a separate Core AI framework for running your own custom models on device with ahead-of-time compilation. Apple also confirmed the Foundation Models framework will go open source later this summer.
So the package is: a free model for most developers, a unified way to reach the expensive models when you need them, vision, and a multi-agent story. That is not a feature drop. That is a platform trying to become the place your app asks for intelligence.
Why free is the lever
Per-token cost has quietly killed more AI features than any technical limit. Not because inference is expensive in absolute terms, but because it turns a feature into a variable cost that scales with usage you cannot fully predict. A solo developer with a note-taking app does the math on summarization at scale, sees a line item that grows with every active user, and ships the boring version instead.
Apple just deleted that line item for that developer. The summarizer, the smart reply, the on-device search that actually understands the query, all of it stops being a cost center and becomes a capability you turn on. When the marginal cost of a model call is zero, the calculus that produced the boring version no longer holds.
That is the real effect of the free tier. It is not that Apple's model is the best one available. It almost certainly is not, for hard tasks. It is that free removes the single most common reason a small team talks itself out of shipping.
The part that is a router play
Look at the move underneath the generosity. The same Swift API that gives you a free Apple model also routes to Claude and Gemini. Apple is not just handing out free inference. It is making itself the default place your app goes to get a model, and the free tier is the on-ramp.
Think about what that does to your code. You wire up the Foundation Models framework once. Free Apple model for the cheap, high-volume calls. A Claude or Gemini call through the same API for the hard ones. It feels clean, and it is clean, right up until you want to leave. At that point you discover how much of your app's intelligence is expressed in Apple's interface, with Apple's profiles, behind Apple's Private Cloud Compute, billed and metered on Apple's terms.
This is the aggregation pattern every platform reaches for. Make the easy thing free, make the unified interface so convenient that routing through you becomes the default, and the platform sits between the developer and every model on the market. Apple is good at this. The free tier is not charity, it is distribution.
None of that makes the offer a trap. Free private inference with a credible privacy story is a genuinely good deal for a small team, and Private Cloud Compute is one of the more serious attempts anyone has made at verifiable private inference. The point is to take the deal with both eyes open about what it is.
How to take the deal without getting stuck
The discipline is the same one that applies to any hosted model, and I have written it before in the context of a model going dark. The model is a component you rent, not one you own, and the interface in front of it is doing more than it looks like.
Do not let the Swift API become the only place your prompts live. The actual asset in an AI feature is the prompt, the routing logic, and the evals that tell you it still works. Keep those in your own code, in a form you could point at a different backend tomorrow. The Apple API should be a thin adapter you call, not the place your product's intelligence is defined.
Use the free Apple model for what it is good at, which is high-volume, latency-sensitive, privacy-sensitive work that does not need a frontier model. Route to Claude or Gemini for the hard tasks, and notice that the moment you do, you are paying again and you are paying through Apple. Know which calls are which, and watch the ratio, because that ratio is your real bill and your real dependency.
And run your evals against more than one backend. If your whole feature is tuned to one model's quirks, the free tier has quietly made that model load-bearing. The time to find out whether your prompts survive a swap is not the day Apple changes the terms or your download count crosses two million.
The wider read
The number that matters is two million downloads. Cross it and the free tier ends, which means Apple has built a pricing cliff exactly where a small developer becomes a real business. Most apps never get there, so most developers get free inference forever. The ones who succeed get a bill that arrives precisely when they can least afford a surprise. That is worth planning for before you are anywhere near the line.
The bigger shift is that a major platform just declared the base model a free utility and kept the meter on everything above it. That is probably where all of this is heading. The commodity model becomes free and bundled, and the money, the lock-in, and the leverage move to the layer that routes, profiles, governs, and meters. Apple got there first for app developers, with a genuinely good free tier as the door.
Walk through it. Just keep your prompts, your routing, and your evals on your side of it, so the door swings both ways.
If you’re under the download threshold, revisit AI features you previously shelved due to variable inference costs and prototype them using Apple’s free tier for high-volume tasks. At the same time, design your app so model calls can be swapped out later—keep prompts, agent logic, and routing decisions as portable as possible rather than deeply embedding them in Apple-specific profiles and abstractions. Watch how Apple’s routing, metering, and future open-sourcing evolve, because the biggest long-term risk isn’t model quality—it’s how hard it becomes to move once your app’s “intelligence” lives behind Apple’s interface.