I waited a few days to write this up. The keynote was the 8th, the State of the Union the 9th, and the developer sessions ran through Friday. Most of the take-pieces went up within twelve hours of the keynote and most of them were wrong, or at least incomplete, because the actual story wasn’t in the keynote — it was in the framework documentation and the partner-tier disclosures that landed on Wednesday.
Here is the short version: Apple opened the Foundation Models framework to every developer, shipped a meaningfully capable 7B on-device model called Apple Foundation 2, and picked Anthropic — not OpenAI — as the cloud partner for the “Private Cloud Compute Plus” tier that powers the new Siri. Any one of those would have been the story of a normal WWDC. Together they are the most coherent platform bet anyone has placed in the AI era.
The on-device model is actually good now
Apple Foundation 1, shipped in iOS 18 two years ago, was a ~3B model that could rewrite an email and summarise a notification. It was fine. It was not a model anyone built a product against. The benchmark scores were embarrassing in a way Apple’s PR carefully avoided quantifying.
Apple Foundation 2 is a different animal. 7B parameters, MoE architecture (Apple is calling it “selective activation” because Apple), trained on a dataset that — based on the technical report — looks substantially larger and cleaner than the 1.x training corpus. The published evals put it roughly at Gemini 2.5 Flash and slightly behind Haiku 4.5 on reasoning, ahead of both on instruction-following in Apple’s chosen English/Japanese/Mandarin/German benchmark mix. It runs at ~40 tokens/sec on an iPhone 16 Pro and ~75 tokens/sec on the M4 iPads.
That is the first on-device model from Apple that I would actually build against if I were a developer in their ecosystem. The capability floor has moved from “novelty” to “useful default.”
The framework is the actual product
The Foundation Models framework — quietly previewed at WWDC 2025, now generally available — is a LanguageModelSession API that gives any third-party developer direct access to the on-device model with a few lines of Swift. Structured outputs via @Generable Swift macros. Streaming. Tool calling. A reasonable token budget per session.
No API key. No billing. No rate limit. No network.
This is the thing the AI coverage has been undercounting for a year. Apple is not trying to win at chatbot. Apple is trying to make “intelligence” a free, unmetered, on-device primitive that every iOS, iPadOS and macOS app gets to use the same way they get to use Core Image or Metal. If you are a small developer shipping a productivity app, you no longer need to decide between paying OpenAI per token, paying Anthropic per token, or shipping nothing. You ship against the framework. The model is the platform’s responsibility, not yours.
The strategic shape of this is the same shape as 2008 Core Location or 2014 Metal. Apple commoditises a layer that the rest of the industry is still trying to charge for. The developers downstream get richer. The dependency on Apple gets deeper. The competing API businesses get a little harder to defend on the consumer-facing edge.
The Anthropic deal is the buried lede
The keynote slide listed “Private Cloud Compute Plus” as the new tier for Siri queries that exceed on-device capability — long-context summarisation, multi-step reasoning, image-grounded conversation. The slide showed an Apple silhouette of a server. No partner logos.
The Wednesday session on Siri architecture, attended by maybe four hundred developers in person, named the partner directly: Anthropic’s Claude Opus 4.7, served through Apple’s Private Cloud Compute infrastructure, with Anthropic having zero access to the prompts or outputs. Apple operates the hardware. Anthropic provides the weights and the inference stack. The data confidentiality contract is the same posture Apple has held since the original PCC announcement in 2024.
This matters for three reasons:
1. It is not OpenAI. The 2024 deal that put ChatGPT inside Siri as a fallback was a stopgap, and was widely expected to be renewed. It was not renewed for the premium tier. ChatGPT remains the “ask the chatbot” affordance on the home screen; Claude is what actually powers the new Siri when the on-device model defers upward. The relationship between Apple and OpenAI just got asymmetric in a way Sam Altman cannot be enjoying.
2. It validates Anthropic’s “trust” positioning commercially. Apple chose Anthropic specifically because the contractual posture around training data, prompt confidentiality, and audit rights was the one Anthropic was willing to sign. That is a six-figure-per-month Claude deal at minimum, but the dollar number is the small part. The reference customer is Apple. Every enterprise procurement conversation Anthropic has for the next eighteen months gets easier because of that slide.
3. It quietly answers the “what is Apple doing about frontier models” question. Apple is not training a frontier model. Apple is buying access to one, on its own infrastructure, under its own privacy guarantees, while owning the on-device layer outright. That is the divide. Frontier capability is rented; ambient capability is owned. As a capital-allocation strategy it is more disciplined than anything the trillion-dollar competitors are doing.
What this means for the rest of the stack
A few second-order effects I think are underpriced right now.
Google’s Android Gemini story just got harder. Apple now has a free, unmetered, on-device model that every iOS developer can call, plus a frontier cloud fallback. Google has Gemini Nano on-device (similar idea, less consistent rollout) and Gemini Pro/Ultra in the cloud (paid, metered, increasingly product-tied to Google’s own apps). The developer story on Android is messier, and Apple just made theirs cleaner in one keynote.
The “AI feature” startups built on OpenAI’s API have a new floor under them. If your product is a Mac menu-bar app that rewrites text, your unit economics just changed. The on-device model is free. You will either need to be obviously better than it, or you will need to be doing something the on-device model demonstrably cannot do. The middle of that market — “thin wrapper over a frontier API for a common workflow” — looks structurally less viable on Apple platforms starting this fall.
The compute-pre-buy story Anthropic has been telling investors gets a non-Anthropic-controlled validation event. The Apple deal is presumably running on a chunk of the cloud capacity Anthropic has been booking through Google Cloud, AWS, and the Oracle slice. The capacity is now publicly tied to a marquee enterprise contract, not just internal Claude usage. The $900B valuation conversation gets one more anchor.
What I am watching
Three things over the next two quarters.
-
Adoption. How many third-party apps ship Foundation Models integration by the iOS 27 GA in September? My over/under is two thousand at launch and ten thousand by year-end. If it’s higher than that, the platform thesis is working faster than I expect.
-
The on-device benchmark drift. Apple has historically updated on-device models in dot releases without much fanfare. Foundation 2.1 in iOS 27.2 is the one to watch — if Apple is keeping the model in step with frontier capability via in-place upgrades, the framework gets compoundingly more valuable to build against.
-
Whether OpenAI gets a second chance. The Siri fallback to ChatGPT still ships. The premium tier is Anthropic. If OpenAI ships a model in the next twelve months that beats Claude on the specific evaluation suite Apple uses internally — and we know they have one — the partner slot is contestable again. It usually isn’t.
The thing I keep coming back to: for two years the consensus take on Apple was “they missed it.” That take was always more vibes than analysis. WWDC 2026 is the moment the analysis catches up. Apple did not miss AI. Apple was waiting to ship the version that fits its actual platform strategy. They just did.