Fable 5 showed up without an announcement. I used it for 24 hours and burned through 6.4% of my total model usage in that time – across all models, over months.
Worth it.
What I used it for
Not content, not routine tasks. I went straight to the projects that had open problems – things where something wasn’t right but I hadn’t nailed down why. Fable’s job was diagnosis and fix, not execution.
Long sessions, deep context, multiple files. That’s why the token count went through the roof. The 3.2M output tokens weren’t waste – that’s what a real debugging session looks like.
Where it was clearly better
Context retention. On sessions where you have to hold several layers of a problem in mind at once, Fable stayed on track longer and needed less correction. Fewer iterations to get to the right answer.
What I didn’t get to test
Fable as the planning model, Sonnet or Haiku doing the execution. Keep the expensive model on root cause and architecture, offload the mechanical parts. I’m curious how far that stretches the quota.
Maybe next time it shows up.