> Claude Fable 5.1 follows explicit tool instructions reliably.

Moving stuff out the API into prompt engineering is obviously less reliable but necessary for progression to 'actual intelligence'. Will be interesting to see if it really is solid.