M.G. Siegler • #siri • Aug 23, 2025

Apple's Great AI Bake Off

Gemini enters the tent alongside Claude, ChatGPT, and Apple's own models to see who can bake the best Siri...

As the saying goes, keep your friends close, and your frenemies closer:

Apple Inc. is in early discussions about using Google Gemini to power a revamped version of the Siri voice assistant, marking a key potential step toward outsourcing more of its artificial intelligence technology.

The iPhone maker recently approached Alphabet Inc.’s Google to explore building a custom AI model that would serve as the foundation of the new Siri next year, according to people familiar with the matter. Google has started training a model that could run on Apple’s servers, said the people, who asked not to be identified because the discussions are private.

Yes, that's the business that OpenAI and Anthropic are also said to be competing for. And while Apple was thought to be favoring the Claude models in that regard, this report indicates that Apple may be balking at the price Anthropic wants to make this all happen. Could it start with a 'B'? Apple famously barely makes acquisitions at such heights, let alone technology deals for a company that famously aims to own and operate their whole stack. Of course, the AI situation in which Apple now finds themselves may demand it. At least until they can get their own models up to speed...

It's the case I've been making for the past year, well ahead of these reports. Siri, for lack of a better word, sucks. And she has for over a decade at this point. We were told this time would be different – again – and we were misled – again. It's clearly time for Apple to seek outside help here.

Naturally, they're still hesitant to do that. As they're also weighing if they can get their own foundation models up to a level so as to not have to outsource this work:

Internally, Apple is holding a bake-off to see which approach will work best. The company is simultaneously developing two versions of the new Siri: one dubbed Linwood that is powered by its models and another code-named Glenwood that runs on outside technology.

It's seemingly two bake offs at once: Apple's own models versus the external ones, and the three external ones versus one another.¹ If they're being honest in the assessment, it's hard to see how Apple could go with their own here right now; not only are they playing catch up with the other three when it comes to foundation models, but they're also losing key talent left and right – largely to Meta, who keeps poaching their players in their own bid to catch up in AI.

The other interesting wrinkle in working with Google here would be the fact that they may be on the verge of altering their search agreement. Certainly, it seems like the default placement is going to go away, so the question is how much it alters the $20B+ annual payments associated with that deal. The key question there: does the judge still allow Google to pay Apple a more standard traffic acquisition rate? Without that, this would be an absolutely wild swing from Google paying Apple billions to potentially Apple paying Google billions to use Gemini. That would really, really fuck up the Services growth narrative that Tim Cook has worked so hard to cultivate.

Or maybe Google would be willing to discount the models for Apple in exchange for some level of branding? Again, these would be Google models running on Apple servers, so the cost would be much less for Google than if they were running on GCP. Or what if Apple sends back data to Google to help improve their models – undoubtedly in a highly anonymized manner? Presumably these are all the elements being weighed right now...