Claude Fable 5 dropped this week, and the benchmark making the rounds is Every's Senior Engineer coding eval (91 for Fable 5, 63 for Opus 4.8).
But benchmarks only go so far. So I ran it on my own work.
It feels like the first model that did what I was expecting it to do. I wrote the brief, prompted it, came back later, and the work matched what I MEANT, not just what I WROTE.
Anthropic also published their official prompting guide for Fable (link in first comment), and it confirms my hunch:
Most of what we've been taught about prompting works against this model.
Here are three prompting habits we're going to have to unlearn:
1. Stop chopping work into small prompts. We all learned to break tasks down because models lost the thread halfway. Anthropic's guide says give it the whole job, not just the first step.
2. Stop iterating turn by turn. Write the full plan upfront, point it at the spec, and let it run uninterrupted, for hours if needed. It checks its own work before handing it back.
3. Stop defaulting to the strongest model. Fable is 2x cost of Opus 4.8 per token, and slower. Daily work stays on Opus/Sonnet. Fable is for the one long job a week you'd otherwise have to do yourself.
It's free on paid subscription plans until June 22 before it runs on credits. Go test it on something real.
What are you pointing it at first?