We're used to hearing that Apple is losing the AI race. The company has no frontier-level foundation model of its own that can compete with Gemini, GPT, or Claude. And even before modern LLMs appeared, Siri was more often a punchline than an example of a good assistant. The partnership with Google was widely seen as Apple's final admission of defeat.

But I wouldn't write Apple off just yet. It has arrived late to markets before, then assembled a more coherent product — that was the story with Apple Watch and AirPods. Siri's history shows that this doesn't always work. Still, perhaps we're measuring the AI race by the wrong metric.

We're Measuring the Wrong Race

Leadership is usually reduced to the quality of the foundation model: who scored higher on a benchmark, solved a harder problem, or handled a longer context better. That matters to model developers. Users don't need the winner of a leaderboard; they need a product that helps them every day.

The most powerful LLM doesn't automatically become the most useful assistant. It still needs context, tools, access to apps, and a clear permissions system. The model is central, but it's only one part of the product.

So I see two races: one for the best frontier model, and another for a personal AI layer between people, their apps, and their devices. Apple is behind in the first one. In the second, it starts from an unusually strong position.

Gemini as Optimization, Not Surrender

In January 2026, Apple and Google announced a multiyear partnership: the next generation of Apple Foundation Models will be based on Gemini models and Google's cloud technologies. It's easy to read that as a white flag: Apple couldn't do it, so now it's buying intelligence from a competitor.

I find another explanation more convincing. Developing and running frontier models requires enormous computing resources. If Google already does that part better, Apple doesn't necessarily need to burn money owning every layer of the system. The partnership could lower costs and risks, shorten development time, and let Apple focus on the part that directly shapes the product.

That's only my interpretation; the financial terms of the deal haven't been disclosed. At the same time, Apple hasn't abandoned its own development — it calls the new family built in collaboration with Google, not a simple call to somebody else's API. But if intelligence becomes an infrastructure component, integrating it well may matter more than developing everything in-house.

A Small Model May Be Enough

Raw power matters when a model has to solve a new and difficult problem. But a personal agent's actions are often more mundane: understand the intent, choose a tool, pass along the parameters, and ask for confirmation when necessary.

A small language model (SLM) can run right on the device in this setting — fast, offline, and without sending personal context elsewhere. More complex requests can go to a more powerful model in the cloud. This isn't a choice between a small and a large LLM; it's a hierarchy. Each one steps in where its particular capabilities are needed.

Apple already has a technical foundation for this approach. Its local model, with roughly 3 billion parameters, supports structured output and tool calling, as the company described in an update to Apple Foundation Models. It isn't frontier-level, but an agent often benefits more from reliably triggering the right action than from writing a perfect essay.

Qwen3.8-27B is another telling example. According to its model card, it comes close to much larger systems on some tasks. 27 billion parameters is too much for an ordinary phone, and the results were published by the developer itself. This isn't a ready-made mobile solution. It's a sign of where things are heading: smaller models are gradually handling increasingly difficult tasks.

AI Shouldn't Live in a Separate Window

Today, AI features often feel bolted onto the main product. You have to open ChatGPT, Claude, or Gemini, bring the context there, get an answer, and move it back yourself.

Apple could integrate AI more deeply because it controls the path from chip to interface: Neural Engine, the operating system, system apps, permissions, frameworks, and the rules developers follow. It already has Spotlight, Shortcuts, quick actions, and App Intents, through which apps describe the actions and entities available to the system.

These pieces don't yet form a single agent, and an action triggered by holding an icon can't magically become a safe AI tool. But the necessary vocabulary is already emerging: apps tell the OS not only about their screens, but also about what can be done inside them. Apple now needs to connect those capabilities with a shared model, context, and access rules.

This is where vertical integration becomes more important than another percentage point on a benchmark. An assistant developer that doesn't control the platform has to negotiate with the owners of the OS, apps, and devices. Apple often gets to negotiate with itself.

An Agent That Follows the Person

Picture a morning. I head out for a run with my Apple Watch and earbuds, ask about the day's plans, and ask the agent to prepare materials for my first meeting. I get home and use HomePod to clarify one of the items. When I sit down at my Mac, the agent continues the task, opens the documents I need, and shows a short summary before the meeting starts.

For now, that's a fantasy, not a working feature. Individual pieces exist, but Apple doesn't have a reliable agent that carries an unfinished task across all of its devices. Few companies, though, have the phone, watch, earbuds, speaker, tablet, and computer integrated so tightly.

Continuity isn't an endless chat. The interface changes with the situation while the intent stays the same: on a run, the agent answers by voice; at home, it continues the conversation; at the desk, it shows the documents. AI becomes a property of the whole system.

Without Trust, None of This Works

An assistant like this would need access to your calendar, email, location, documents, and authenticated sessions. The more useful the agent, the more data and authority you have to entrust to it — and the more dangerous its mistakes become. Users need to understand what the agent can see, what it can do, and where it will stop without confirmation.

Apple is already trying to address the cloud side of this problem through Private Cloud Compute. The company says requests are processed on attested nodes, data isn't stored, and the software images are available to researchers for inspection. That's more concrete than the usual promise to “value your privacy.”

Mark Zuckerberg recently described a similar future. In his manifesto The Future is for Everyone, he talks about a personal agent that understands a user's goals and works in their interest “24/7.” On the same day, Meta released the open weights for Muse Glimmer-30B under Apache 2.0. Meta calls the model optimized for “always-on local agent workflows.” It's a local model for a powerful Mac or PC, not a ready-made mass-market assistant, but the direction is quite specific.

The idea itself sounds reasonable, even to the author of a critical column. Russell Brandom wrote in TechCrunch that he agrees with Zuckerberg “in broad strokes.” His problem is the messenger: society doesn't trust technology-company executives to decide the consequences of new technologies. That skepticism toward Meta wasn't unfounded. In 2019, Facebook agreed to pay a record $5 billion fine to settle FTC allegations that it had violated a 2012 privacy order.

Apple may have an advantage here: it backs up its privacy claims with the concrete architecture of PCC and a reputation built over many years. But PCC doesn't make an agent safe automatically. It won't protect against a misunderstood command, excessive permissions, prompt injection in an email, or a bug in an app. And trust in a brand can't be earned once and then spent forever.

Why Apple Could Still Lose

The strongest argument against everything above is Apple itself. The company spent years failing to make Siri genuinely useful. In March 2025, Apple acknowledged a delay to the personal-context and in-app-action features it had promised a year earlier. In June 2026, it introduced the new Siri AI, but it was still a beta at the time of publication. Control over the stack is useless if the team can't connect its parts into a working product.

There's a strategic risk, too. Google is now not only a partner but also a supplier of critical technology, while it develops Gemini, Android, and its own device ecosystem. Google and Samsung already have Gemini with actions across multiple apps, so Apple's head start is far from unique.

A real agent also needs third-party developers to participate. They'll have to describe their actions, work with permissions, and provide ways to undo mistakes. Safety adds confirmations and limits; autonomy raises the cost of an error. Finding the balance is harder than putting on a pretty demo at WWDC.

In Conclusion

The winner of the consumer AI race may not be decided on a benchmark leaderboard. It will be whoever makes an agent useful across devices and apps, while also making it safe enough for real-world actions. Apple has an operating system, system tools, on-device models, a private cloud, and users who have lived inside its ecosystem for years.

The hardest part remains: bringing it all together. If Apple succeeds, its current lag in foundation models will turn out not to be a death sentence, but just one detail. If it doesn't, vertical integration will become another attractive advantage Apple failed to use.