The next great leap in consumer AI could be courtesy of an unlikely alliance between two of the world’s biggest tech rivals: Apple is finalising a landmark deal to license Google’s most-advanced artificial intelligence model, a 1.2 trillion-parameter version of Gemini, in an annual deal costing about $1 billion that will redefine the capabilities of Siri.
The scale of Gemini is unprecedented within the Apple ecosystem: while the currently deployed cloud-based Apple Intelligence runs on a 150 billion-parameter system, the architecture of Gemini is approximately eight times more complex. Parameters are one way to describe how large-scale language models learn weights that define how a system processes and generates information. More generally, more parameters enable richer contextual understanding, deeper reasoning, and nuanced responses. At 1.2 trillion, Gemini is considered one of the most computationally demanding AI systems in production, on par with the frontier models from OpenAI and Anthropic.
The move comes after months of internal benchmarking against ChatGPT and Claude. Although Anthropic’s Claude reportedly fared rather well in some tests, Apple settled for Google’s product, swayed by both technical advantages in integrating the product and pre-existing commercial relationships-Google already pays Apple more than $20 billion annually to be the default search in Safari. The positioning of the partnership is such that Gemini powers select high-complexity Siri features, while it keeps its core functionality on Apple’s in-house models.
Technically speaking, Gemini will handle the “summarizer” and “planner” layers of Siri. The former enables advanced summarisation of text, audio, and web content to generate things like notification digests, webpage summaries, or even contextual writing suggestions, while the latter is essentially a decision engine that identifies optimal execution paths for multi-step tasks. This would likely involve orchestrating app actions via Apple’s App Intents framework, integrating personal data such as calendars or photos, and stitching together online and local resources. Knowledge search, however, remains with Apple’s own LLMs on-device, keeping general trivia queries, among other things, strictly under Apple’s control.
Integration at this scale is a non-trivial engineering challenge. Executing a trillion-parameter model demands enormous computational throughput and memory bandwidth, far beyond what today’s consumer devices can sustain locally. Apple will host Gemini on its Private Cloud Compute infrastructure, a privacy-preserving architecture that leverages encrypted, stateless processing to ensure user data never touches Google’s servers. This kind of deployment requires specialised AI server hardware, which Apple has reportedly already allocated, that’s capable of sustaining the inference workloads while meeting latency requirements for real-time voice interaction.
Performance benchmarks for trillion-parameter systems suggest dramatic improvements in reasoning and multi-modal understanding. Models of this scale can sustain coherence for long dialogues, perform multi-step logical chains, and incorporate cross-domain knowledge with minimal degradation in accuracy. They also require sophisticated guardrail engineering to avoid undesired outputs. Apple will be putting its own content constraints on Gemini’s responses when they are summoned within Siri, in effect “white-labeling” the model to operate under Apple’s rules rather than Google’s. That difference is crucial for brand coherence and to avoid public association with some of the more controversial capabilities that Gemini might provide in other contexts.
The deal is not expected to be permanent. Apple continues to invest heavily in its AI research, seeking to close the gap with frontier models. Internally codenamed “Glenwood,” the Siri revamp comes from Vision Pro creator Mike Rockwell and software engineering chief Craig Federighi, while the voice assistant itself-planned for iOS 26.4-is codenamed “Linwood.” All that said, the Gemini integration is considered an interim measure that accelerates feature deployment, with a public launch aimed for spring 2026.
For Google, the deal secures a strategic position within Apple’s ecosystem, possibly barring its competitors like OpenAI and Anthropic. For Apple, it is access to a model developed using more than $100 billion of cumulative AI investment at a fraction of that cost. Pragmatically, it could mean a Siri that is quicker, more context-aware, and able to perform complicated tasks with a fluidity that simply wasn’t possible earlier-while keeping the supplier behind it invisible to end users.
