Apple lets developers plug into Apple Intelligence

Adding AI features to an app usually comes with two big headaches: brutal monthly cloud bills for API tokens, and angry users the second their phones lose cell service.

Apple just handed developers a way around both. At WWDC, the company unveiled the Foundation Models Framework, a brand-new API that gives third-party developers direct access to the on-device large language models powering Apple Intelligence.

Offline smarts with zero API bills

Instead of spinning up external cloud servers or bundling heavy open-source models into an app download, developers can now plug straight into Apple's local models.

According to Craig Federighi, Apple's senior VP of software engineering, the models behind Apple Intelligence have grown more capable and efficient. Opening them up directly is meant to kick off a wave of privacy-first, on-device intelligence across everyday apps.

The perks are huge: features run entirely on-device, user data stays private, everything works offline, and developers do not pay a cent in inference costs.

Apple pointed out a couple of immediate use cases. An education app can scan a user's notes and generate quizzes completely on-device. An outdoor navigation app can handle natural language searches deep in the backcountry with zero cell reception.

The framework also ships with built-in guided generation and tool calling, making it straightforward to plug generative features and app actions into existing codebases.

Developers can start testing the framework today through the Apple Developer Program at developer.apple.com, with a broader public beta rolling out next month on beta.apple.com.

Why it matters

For indie developers and smaller software studios, zero inference costs is an absolute game-changer.

Up until now, adding generative AI meant either charging a recurring subscription to cover OpenAI or Anthropic API bills, or praying power users would not bankrupt the app. Offloading inference to the user's own iPhone chip completely rewrites the unit economics.

We think this is one of Apple's smartest ecosystem plays yet. Instead of trying to invent every single killer consumer AI feature inside Cupertino, Apple is turning its local models into primitive device hardware, just like GPS or the camera. It lets millions of outside developers do the heavy lifting of figuring out what consumer AI is actually good for.

The catch, of course, is model capacity. An on-device model constrained by phone thermals and battery life cannot match the raw reasoning of a frontier cloud model. It will excel at structured tasks like parsing notes, sorting data, and local search, but developers will have to carefully manage user expectations when tasks demand deeper reasoning.