Apple made on-device AI free for small studios. Here is the catch.Apple made on-device AI free for small studios. Here is the catch.

At the 2026 Platforms State of the Union, Apple said developers with fewer than two million first-time App Store downloads get Private Cloud Compute access to its Foundation Models at no cost. The framework also gained image input, a Dynamic Profiles system for multi-agent workflows, and the ability to call third-party models such as Claude and Gemini through the same Swift API.

For a studio our size that is a genuinely significant change, and not for the reason it is usually reported.

What actually changed

The barrier to shipping an AI feature in a client app was rarely the model. It was the bill, and the conversation that came with it.

Adding a hosted model to a client's app used to mean introducing a per-request operating cost to a project quoted as a fixed build. Somebody then owns that cost forever. For a startup client it is a line item they had not planned; for an enterprise client it triggers a procurement review that can outlast the build itself.

Removing the infrastructure cost removes that conversation. A summarise button becomes a feature decision rather than a budget decision, which is a different category of problem.

The threshold is the interesting part

Two million first-time downloads is a generous ceiling. Most apps never approach it. But note what it is measured in: downloads, not revenue, not requests.

That has an odd consequence. A free app with wide reach and no monetisation can cross the line while earning nothing. A high-value B2B app with four thousand installs will never come close. The threshold correlates with popularity rather than with ability to pay.

If you are building something with viral potential, model what happens on the other side of that line before the architecture assumes free inference. If you are building an internal tool or a niche professional app, treat it as effectively unlimited.

Design for the day the free tier ends, even if that day never arrives.

The same API for third-party models is underrated

Calling Claude or Gemini through the same Swift interface matters more than it sounds, because it turns the model into a configuration detail rather than an architectural commitment.

The usual failure mode with AI features is coupling: prompt construction, response parsing and error handling get written against one provider's quirks and spread through the codebase. Changing provider then means touching thirty files. A single interface keeps the decision reversible, which matters because the price and quality ranking of these models has changed several times a year, every year.

What has not changed

• Models still produce confidently wrong output, and your interface has to be designed for that rather than around it.
• You still need an answer for no connectivity, because Private Cloud Compute is not on-device.
• Latency is still a design problem. A two-second pause needs designing for, not apologising for.
• You still have to tell users what leaves their device and when, in language they understand.

The cost was never the whole barrier. It was the most visible one, and the one that killed proposals before the interesting problems were reached.

What we would build with it

The features worth building here are the small, unglamorous ones. Summarising a long note. Extracting structure from pasted text. Drafting a reply the user then edits. Cheap to build, they degrade gracefully when the model is wrong, and they do not ask the user to trust output blindly.

Avoid the features where a wrong answer is expensive and invisible - anything touching money, medical information or legal text without a human confirming the result. The economics of inference changed this year. The reliability did not.