A model too good to just kick the tires with

Two threads today, and they ended up framing the same question from opposite ends.

The first is steady progress on Relay Notes, the iPhone app I started last week. Transcription is now working through Apple's built-in engine, and I've started wiring in the option to swap that for an open-source speech-to-text model, with Whisper as the first target. Giving the app a choice between the on-device default and something I can run myself feels like the right shape this early.

The second thread was kicking the tires on Claude Fable 5. It is genuinely impressive, but it would burn through my Max subscription in no time. And Fable only sits on the subscription plan for a couple of weeks before it moves to API pricing, which changes how I have to think about it. A model at that tier is no longer in the "let me try something out" space, at least not on my budget. Using it well now means having an ROI plan before I open the session, not after.