If an LLM, apple or otherwise, runs on a phone, with audio input and output, then airpods and AliExpress five dollar earbuds should both be able to perform the I/O. I don't see a technical reason the latter is impossible. Indeed, it seems like it should work with the phone mic and speaker and no headset, too.

This isn't rocket science: audio goes into mic => STT engine => translation model => TTS engine => audio comes out of speaker. As a fellow hacker here, you could piece together something like this in a weekend on your computer for fun.

As for your question though: they can charge a subscription for using their LLM if they want, or charge for this specific app/feature of iOS. Or just be like me: whenever I'm about to execute on a business plan, I ask myself: "Is this business plan economically feasible without breaking the law?" And if it is not, then I do not do that plan. So far I haven't been cited for illegal conduct by any unions of dozens of countries, so it appears this tactic works.