Apple is open-sourcing the Foundation Models framework that powers Apple Intelligence and, for the first time, letting third-party apps route prompts to outside models like Anthropic's Claude and Google's Gemini. Announced at WWDC on June 9 and set to open this summer, it is the clearest sign yet that Apple has stopped trying to win generative AI alone and is instead turning the iPhone into a neutral runtime other model makers plug into.
- The Foundation Models framework, Apple's on-device AI layer, is being open-sourced, alongside two runtimes: CoreAILanguageModel (Apple Neural Engine) and MLXLanguageModel (Mac GPU).
- Developers can now point the same API at Claude and Gemini, not just Apple's models, and pass image input in addition to text.
- Apple's newest Private Cloud Compute models are free to small developers, lowering the cost of shipping AI features on Apple platforms.
- The strategy shift: Apple monetizes the device and the runtime, not the model, and outsources the frontier race to companies already winning it.
What did Apple actually announce?
Three things that fit together. First, Apple committed to open-sourcing the Foundation Models framework, the same on-device layer that runs summarization, writing tools and other Apple Intelligence features, and to releasing CoreAILanguageModel and MLXLanguageModel so developers can run local models directly on the Neural Engine and Mac GPU. Second, it opened that framework to outside models, so an app that today calls Apple's small on-device model can, with the same code shape, call Claude or Gemini instead. Third, it added image input and made its newest Private Cloud Compute models free for small developers, removing a real cost barrier for indie apps that want cloud-grade answers without standing up their own inference.
RelatedOpenAI retracts SWE-Bench Pro after finding 30% broken
Why would Apple invite rivals onto the iPhone?
Because Apple is behind on models and ahead on distribution, and this trade plays to its strength. Apple's own models are competent at short, private, on-device tasks but are not competing with frontier systems from OpenAI, Anthropic and Google on reasoning or coding. Rather than spend years closing that gap, Apple is doing what it did with the App Store: own the platform, take a cut of the ecosystem, and let others supply the hits. If the best model on your iPhone is Claude or Gemini, Apple still wins, because the request runs through Apple's framework, on Apple silicon, inside Apple's privacy story. The model becomes a swappable component; the runtime is the moat.
How does this compare to the closed approach?
| Dimension | Foundation Models (2026) | Apple Intelligence at launch (2024) |
|---|---|---|
| Model choice | Apple, Claude, Gemini, local | Apple models only (plus opt-in ChatGPT hand-off) |
| Framework | Open-sourced this summer | Closed, Apple-internal |
| Inputs | Text and image | Mostly text |
| Local runtimes | CoreAILanguageModel, MLXLanguageModel | Not exposed |
| Cost to small devs | Private Cloud Compute free tier | No dedicated free tier |
The gap between the two columns is the whole story: in eighteen months Apple went from a walled model to a framework that treats its own AI as one option among several. That is an unusual admission from a company that prefers to control the full stack.
What it means for the market
The signal for investors is that Apple (AAPL) is repositioning from an AI model contender to an AI distribution layer, which is a more defensible place to stand given its device install base. It also strengthens Anthropic and Google as suppliers: every iPhone app that adds a Claude or Gemini option is incremental API demand routed through Apple's framework. The clearest loser is any startup whose pitch was a thin wrapper around a single model on Apple platforms, because Apple just made model choice a system feature. Watch whether Apple eventually charges a platform fee or rev-share on third-party model calls the way it does on in-app purchases; that is where the runtime-as-moat thesis turns into revenue.
RelatedGrok 4.5 lands: Opus-class claims, cheaper, unproven
What should developers watch before shipping?
- The license. Open-source can mean many things. Whether the framework ships under a permissive license or a restrictive one decides how much developers can actually fork and reuse.
- Latency and cost of routing. Falling back from a free on-device model to Claude or Gemini has real per-call cost. The router's defaults will shape which model most apps actually use.
- Private Cloud Compute limits. The free tier for small developers is generous on paper. The rate limits and definition of small will decide if it is usable in production.
- Rev-share. If Apple later taxes third-party model calls, the economics of building on the framework change overnight.
Our take
This is Apple playing to type, and it is smart. The company spent a year trying to be an AI model maker and quietly concluded that owning the runtime beats owning the weights. By open-sourcing the framework and welcoming Claude and Gemini, Apple turns a weakness, its middling models, into a platform advantage: the iPhone becomes the place where every good model is one API call away, and Apple sits at the toll booth. The risk is that opening the door invites the frontier labs to build direct relationships with users and route around Apple over time. But for now this is the rare move that helps developers, helps Anthropic and Google, and helps Apple all at once, and those are usually the moves that stick.
- OfficialApple Intelligence for developers Foundation Models framework overview
- ReferenceApple Machine Learning Research on-device and Private Cloud Compute models
- BenchmarkGenZTech AI Coding Leaderboard where Claude and Gemini rank on coding tasks
Original analysis by GenZTech. Figures current as of July 2026.