I had Fable 5 for exactly 24 hours. In that time it hit 6.4% of my total model usage – across all models, over months.

That’s not a complaint. That’s context.

How I used it

I didn’t run it for routine tasks. I pulled up projects that had been stuck – things where Sonnet did the work but something wasn’t adding up. Wrong assumptions, bad architecture decisions, edge cases nobody caught. Fable’s job was to find the problem and fix it.

That kind of session burns tokens. You’re not generating content, you’re thinking out loud with a model that actually keeps up. Long context, multiple files open, back and forth until it clicks. Of course the quota went fast.

What made the difference

Fable held more in context without losing the thread. On complex debugging sessions – the kind where you have to trace a problem across several layers – it stayed accurate longer and needed less hand-holding.

The 3.2M output tokens in the chart tell the story. I was in long sessions, not quick queries.

24 hours is too short

I would have liked a week. Specifically to test the mixed setup – Fable for planning and root cause analysis, Sonnet or Haiku for execution. Keep the expensive model on the hard parts, offload the rest.

That experiment didn’t happen. Fable was gone before I got there.