Hi friends,
This one's a day late, Friday instead of the usual Thursday. Anthropic launched Claude Fable 5 on Tuesday, and I've spent the last few days running it on real work so I could share my thoughts.
My take on Fable 5
It's very good at the long and complex tasks. I pointed it at a migration I'd been putting off for weeks, the kind that touches 30 files and it just worked through the whole thing.
Two things to watch. The first is cost. It's $10 per million input tokens and $50 per million out (~2x Opus). And you feel it fast. So it's not a daily driver. You need to choose the kind of task it excels at.
The second is the deadline. Fable's only on the Pro, Max, Team, and Enterprise plans through June 22. After that have to pay API price for it.
So how I am thinking about Fable: Don't rebuild your whole setup around a model that’s temporary. I am using it for the next 2 weeks to get a sense of the what the frontier looks like but not getting my hopes too high.
And everyone started saying "loop engineering"
The other big thing this week wasn't a model, it was a phrase. It was all over X. This is the next phrase after prompt engineering, context engineering and harness engineering.
The short version: instead of prompting Claude a step at a time, you give it a goal and a way to check its own work. Then let it run on its own, write, test, read the error, fix, and go again until the check passes. Boris Cherny says he barely prompts it directly anymore, his whole job now is writing the loops.
Under the hood this isn't new, and it's the part to be careful with. The bare-bones version even has a name, the "ralph loop," from Geoffrey Huntley: you run the agent in a while true loop on the same prompt and let the work pile up on disk between runs.
Before you point one at your repo, know two things. It only works when the task is well-defined and the agent has a real way to check itself. And it burns a lot of tokens, because it keeps rerunning until that check goes green. Great for a boring, well-specified grind, risky and expensive for anything fuzzy. If you want to go deeper, Addy Osmani's writeup is the clearest one I've found.
I put these two together for a reason. Fable makes the agent good enough that you want to step back from the code, and a loop is how you actually step back, as long as the task is tight enough to trust it. The Claude Code team summed it up: we used to verify that Claude did the work right, now we verify that it's doing the right work.
The memes
That's the week. I would love to know how your experience with Fable has been so far. Hit reply and let me know!
Abhishek