Apple Intelligence shipped with iOS 18 / macOS 15 and has matured significantly with iOS 27 / macOS 27. The pitch for developers: call Apple's Foundation Models API and run language model inference for free, on-device, using Apple Silicon. No API key, no per-token charges, no egress.

For business applications targeting Apple users, this is a genuine cost shift — not a marketing story.

What Apple Intelligence Actually Runs On-Device

The key distinction is on-device vs. Private Cloud Compute (PCC). On-device inference runs on the A18 chip (iPhone 16+) or M1+ (Mac, iPad). PCC sends requests to Apple's server infrastructure for heavier tasks.

Both are free to app developers. Neither sends data to OpenAI or any third-party model provider (unless you opt in to ChatGPT extensions, which Apple routes through their privacy proxy).

What runs on-device:

What gets routed to PCC:

Apple doesn't publish the exact routing thresholds, but in practice, PCC handles requests where the on-device model would produce notably worse results. Still free to you.

The Foundation Models API

Apple's Foundation Models framework (available since iOS 18.1, significantly expanded in iOS 27) exposes a Swift API for on-device inference. There's no Python SDK — this is a native Apple platform API.

import FoundationModels

// Text summarization
let model = SystemLanguageModel.default
let session = LanguageModelSession(model: model)

let prompt = "Summarize the following customer feedback concisely:\n\n\(feedbackText)"
let response = try await session.respond(to: prompt)
print(response.content)

For structured extraction:

import FoundationModels

@Generable
struct CustomerIntent: Codable {
    let category: String
    let urgency: String
    let sentiment: String
}

let session = LanguageModelSession()
let intent = try await session.respond(
    to: "Classify this support message: \(message)",
    generating: CustomerIntent.self
)
print(intent.category) // "billing", "technical", "general"

The @Generable macro handles schema generation for structured output. This is Apple's equivalent of JSON schema / structured output from OpenAI and Anthropic.

The Developer Cost Calculus

If you're building an iOS or macOS app with any of these features, the comparison is stark:

Without Apple Intelligence: A document summarization feature for 10,000 daily users, each summarizing one document per day (~2,000 tokens input, 300 tokens output):

With Apple Intelligence: $0/month. Free. The compute runs on the user's device.

The obvious caveat: quality. Apple's on-device model is roughly in the GPT-3.5-turbo territory for complex tasks, better than that for its optimized use cases (summarization, extraction, writing). For features where that quality bar is sufficient, the economics are unbeatable.

Practical Integration Patterns for Business Apps

Email triage apps: Classify and summarize incoming email on-device. Zero cost, zero latency (no network round-trip), works offline. This is a perfect fit — the on-device model handles classification well.

CRM and sales tools: Summarize call notes, extract action items, suggest follow-ups. These tasks fall squarely in Apple Intelligence's wheelhouse.

Document review: Flag relevant sections, generate summaries, answer questions about document content. Works well for moderate-length documents; longer documents may route to PCC.

Customer support clients: Draft initial responses, classify ticket type, suggest knowledge base articles. On-device inference means response suggestions appear instantly as the user types, with no API latency.

Code review tools for Xcode: Apple Intelligence integrates with Xcode's code completion. For internal dev tools on Mac, this is a free tier for basic code suggestions.

Where It Falls Short

Complex reasoning tasks. Multi-step analysis, comparing multiple documents, nuanced argumentation — Apple's on-device model isn't competitive with GPT-4o or Claude Sonnet 4.5. You'd use Apple Intelligence for the first pass and escalate to cloud APIs when depth is required.

Non-Apple users. If your business app runs on Android, Windows, or web, this is irrelevant. Apple Intelligence is Apple Silicon only. This matters enormously for enterprise deployments that aren't Apple-first.

API maturity. The Foundation Models API has evolved rapidly but still has rough edges. Structured output support matured in iOS 26. Custom fine-tuning of on-device models isn't currently supported. The API is opinionated about what it can do.

Offline and low-resource edge cases. A18 chip devices handle inference well under normal conditions. Inference can be throttled by iOS when the device is low on battery or thermal headroom is limited. Don't assume you'll get consistent low-latency responses in all conditions.

The Hybrid Strategy for Apple Platform Apps

The most cost-effective pattern for Apple-platform business apps in 2026:

  1. Default to Apple Intelligence for all summarization, classification, extraction, and short-form generation tasks
  2. Escalate to cloud only when the user explicitly requests a more detailed or complex response, or when task type requires it
  3. Cache cloud results aggressively — if 10,000 users all ask similar questions, serve from cache rather than paying per-query
  4. Track escalation rate — if users are hitting the cloud fallback more than 20% of the time, either your on-device capability assessment was wrong, or you need to tune your escalation thresholds

A well-implemented hybrid setup on an iOS business app can realistically run 85–90% of LLM calls free via Apple Intelligence, paying only for the 10–15% that genuinely require frontier model capability.

The business case is clearest for apps with large user bases doing routine NLP tasks. A CRM app with 50,000 daily active users doing on-device summarization instead of cloud calls isn't saving $100/month — it's saving $50,000/month against GPT-4o pricing.