Which iPhones
Offhand runs on Apple Intelligence. Making tasks from a note needs an A17 Pro chip or newer, and that is a separate requirement from iOS 26. It is the one that decides.
Speech recognition is by SpeechAnalyzer and task extraction is by Foundation Models, both on the iPhone. Neither has a server behind it, which is why the chip matters: the model that reads your note has to fit on the phone in your hand.
We measured the difference rather than assert it. On the hardest of our own test sets, 136 written notes, the model gets the number of tasks right 89% of the time where code alone gets 69%, and the day and time right 86% against 64%. That is what the chip buys.
It is smaller than it sounds for short notes: 28% of everyday notes, 73 of 261 in our everyday set, are short enough that no model is asked on any phone. Buy milk is one task on every iPhone.
What the model does not do is make things up. In our launch study, no task was given a clock time its note never said: 0 of 980 notes, across both languages measured and both generations of the model. Every figure here was decoded on a Mac, not on an iPhone; the model is managed by the OS, and nothing here measures how long anything takes on a phone.
Nearly always Apple Intelligence is off. Check Settings › Apple Intelligence & Siri. If you turned it on recently the model may still be downloading; recording keeps working and tasks start appearing once it lands. If the phone is an iPhone 14 or 15, see above: one task per note is what it does, and it is doing it.