Apple’s Gemini-powered Siri: 1.2T parameters, $1 billion a year, and a temporary admission of AI weakness

Apple is preparing to plug a custom 1.2 trillion parameter version of Google’s Gemini into a rebuilt Siri, in a deal worth around $1 billion a year. On the surface this looks like a simple “better Siri” story. In reality it is Apple renting frontier model capacity from a direct rival while it tries to catch up with Microsoft plus OpenAI and Google in general purpose AI.

Glenwood, Linwood, and what is actually changing

Internally, the Siri overhaul is split into two efforts. Glenwood is the project to wire a custom Gemini model into Apple’s infrastructure. Linwood is the new Siri experience planned for iOS 26.4. The work is being led by Vision Pro veteran Mike Rockwell and software chief Craig Federighi.

Today, Apple’s cloud side Apple Intelligence features reportedly use a model of around 150 billion parameters. The custom Gemini version is 1.2 trillion parameters. Parameter count is not a quality guarantee, but it does track capacity and context window limits. Moving from 150 billion to 1.2 trillion is not a fine tune, it is a step change in the size of the engine behind Siri’s most demanding tasks.

Apple is not handing the entire assistant to Google. Under the current plan, Gemini will handle Siri’s summariser and planner roles. Those are the parts that decide what information matters and how to break a complex request into actions. Other Siri functions will continue to use Apple’s own models. This lets Apple bolt a stronger reasoning core into the system while keeping some behaviour anchored to its existing stack.

Why Apple ended up choosing Google

Bloomberg’s reporting says Apple did not start from a “Google by default” position. It evaluated models from OpenAI (ChatGPT), Anthropic (Claude) and Google (Gemini) before deciding earlier this year. That testing likely covered accuracy, latency, safety controls and practical integration work.

The money is significant but manageable. Roughly $1 billion a year for a custom Gemini variant is a large ongoing expense. It is still smaller and more flexible than the capex and opex needed to train and serve a comparable Apple model at global Siri scale on Apple owned infrastructure. Renting buys time at a predictable cost and offloads some model R&D to Google’s team.

There is also the simple performance argument. Gemini 2.5 Pro sits at or near the top of most public leaderboards today. Leaderboards move, but in the window when Apple had to commit, Gemini appears to have offered the best balance of capability and maturity for this use case.

Private Cloud Compute and how data flows

The sensitive part is not only which model Apple uses, but where it runs. Apple has spent years telling users that their data stays on device or in tightly controlled Apple data centres. Sending Siri traffic to a Google model is an obvious tension with that message.

Apple’s answer is to keep Gemini off Google’s own infrastructure entirely. The custom model will run on Apple’s Private Cloud Compute servers, using AI hardware that Apple has already set aside. In that design, Google supplies model weights, tooling and technical guidance. Apple operates the runtime environment.

Even with that setup, Apple still has to be precise about three things.

Which types of request are allowed to go to the Gemini tier, and which are forced to stay on device or on Apple only models.
How much personal context is stripped out or anonymised before any text or embeddings are sent to the cloud.
How long logs, traces and telemetry are kept, and whether any of that can be used to improve the model rather than just for reliability and abuse handling.

If Apple can present Gemini as a stateless service that receives minimised, de identified inputs and returns answers without keeping user level history, it can argue that its privacy posture is intact. If the implementation is opaque, users and regulators will assume the worst.

Temporary crutch or new long term dependency

Apple is publicly positioning the Gemini deal as an interim step. Its own models team is working on a 1 trillion parameter cloud model that it hopes to ship into consumer products as early as next year. The stated goal is to reach comparable quality and then replace Gemini.

That is easy to describe and hard to execute. Apple has lost senior AI staff to Meta and other companies, and Google is still improving Gemini in parallel. If Apple’s own trillion parameter system launches and feels even slightly weaker than Gemini in real use, the temptation to keep renewing the Google deal will be strong.

There is also integration cost. Once Siri’s higher level logic, prompts and safety filters are tuned around the behaviour of a particular model class, dropping in a different frontier model is not just a weight swap. It is a retuning exercise across the assistant stack. Every extra year that Siri runs on Gemini pushes Apple deeper into that dependency.

How this alters the Apple and Google relationship

Apple and Google already share one of the most valuable deals in tech. Google pays Apple large sums to be the default search engine in Safari on iOS and macOS. That arrangement ties search share, ad revenue and platform distribution together.

Adding Gemini to Siri takes the relationship deeper.

Google becomes both the default search provider and a core AI supplier for iOS. That increases technical and business lock in.
Competition authorities who already dislike the search deal now have a new line of inquiry. They can ask whether Google is becoming the default AI provider across both Android and iOS and closing the door to other model vendors.
Any overlap between Siri queries and search like intent will blur the line between assistant responses and search results. Apple has an incentive to keep that boundary clear if it wants to argue that it is not quietly handing more control of user journeys to Google.

It is notable that Apple reportedly does not plan to advertise the Gemini partnership heavily. Where the search deal is public and obvious, the Gemini contract is meant to sit in the background. The message to consumers will be that Siri has improved. The branding of the underlying model is meant to stay almost invisible.

China forces a different stack

The Glenwood and Linwood blueprint does not work in China. Google services are blocked there, and Gemini is not an option. For the Chinese market, Apple is pursuing a different architecture.

A tailored version of Apple Intelligence will use Apple’s own models, combined with a filtering layer developed by Alibaba. That filter is designed to adjust content to comply with Chinese government rules. Apple has also explored a partnership with Baidu for AI in China.

This has two consequences. First, Apple now has to maintain two distinct AI stacks: a Gemini assisted Siri for most markets and a local stack built around Apple models plus Chinese partners. Second, China becomes a live test of whether Apple’s own trillion parameter model can match or approach the Gemini backed experience. If the China only Siri feels similar in quality to the Gemini version, the case for eventually dropping Gemini globally becomes stronger.

What end users are likely to notice

If the projects ship roughly as described, Siri should change in visible ways for everyday use.

Complex multi step tasks, such as combining information from several apps and services, should succeed more often because the planner is stronger.
Summaries of long messages, web pages and documents should be more coherent and less prone to missing key points.
Conversational requests and follow ups should feel less brittle, with fewer failures on rephrased commands.

At the same time, Siri will pick up the usual generative AI weaknesses. A 1.2 trillion parameter model can still hallucinate, misread ambiguous prompts and produce confident but wrong answers. Apple will have to decide when Siri should admit uncertainty, ask follow up questions or fall back to deterministic behaviours instead of guessing.

There is also the risk of uneven behaviour. Some queries routed to Gemini backed functions will feel impressively smart. Others that stay on smaller Apple models may still feel like the old Siri. That inconsistency is part of the cost of a hybrid architecture that splits work across several model tiers.

Where this fits in the wider AI platform race

The Siri and Gemini plan underlines where the platform race stands in late 2025. Microsoft is tied closely to OpenAI across Windows, Office and Azure. Google is building around Gemini for Android, Workspace and its cloud customers. Apple is now in a mixed position, renting Gemini for heavy lifting while trying to scale its own models underneath.

The common pattern is that a handful of very large models sit at the centre of mainstream products. Platform owners then differentiate through routing, privacy rules, on device acceleration and how they expose AI to developers. In that world, Apple using a Google model behind Siri is less of an outlier and more of an early example of how alliances form around shared AI engines.

The open questions are how long Apple is willing to let a rival supply the top tier of Siri’s intelligence, and whether its own trillion parameter model can genuinely close the gap. For now, the plan to buy a 1.2 trillion parameter Gemini while building an in house alternative is both a major upgrade for users and a quiet admission. Apple knows that its current models are not enough on their own, and it is prepared to pay Google to fill that gap while it rebuilds.