Fable 5.1 and GPT Astra Changed How I Prompt. TheEffort Dial Was the Wrong Place to Reach First.

Cranking a coding agent up to its highest effort setting and hoping is how a week of tokens disappears in one night. What works with this generation of models is sorting the task before you type anything: is this a job you walk away from, or a job you sit at the desk for?

The word I stopped typing

In December 2025, half of everything I typed at a coding agent ended with "ultrathink," the magic word that made Claude Code think as hard as it could on that turn. I have not typed it in three months. I grepped my own logs expecting to find that I had learned to prompt better. That is not what I found.

Claude Fable 5.1 and GPT Astra changed how I prompt, and the change is not what I expected. The old move was to crank it up and hope. With these two, crank it up and hope is exactly how you burn through your weekly allowance before the week is half over. Anthropic now documents the successor to the word I used to type: a setting called ultracode that sends the highest effort level and lets Claude spin up as many agents as a task needs (Claude Code model configuration). More thinking is one setting away. That is precisely why reaching for it first is the mistake.

What works now is deciding, before I type anything, which of two kinds of task I have.

The walk-away task

One prompt, then I leave, and the model finishes without me.

My logs have 167 sessions in the last three months where I typed once and came back to commits and green tests. Every one of them had the same shape: the model could check its own work. Tests it wrote and ran. A build it kicked off. A screenshot it could look at and compare. The times I was the only check, I came back to a prototype and an afternoon of finishing it myself.

Astra is my walk-away model. It does not ask questions. It picks something and iterates, which is exactly what you want when nobody is waiting and a wrong attempt costs nothing but a retry. It is also eager, and eager with no finish line means it iterates all night. That is how I learned about my weekly cap.

My best Astra run went 48 hours at medium effort and came back with 103 browser checks passed and a commit waiting for my approval. Medium effort. The effort setting was never the difference. The stopping rule was: it had a finish line it could check, and an instruction to stop and ask before pushing.

If you want to try one tonight, this is the prompt shape that has worked for me:

Build me [a small tool I would use]. Write the tests first and run them. Install it and use it once. Do not say done until all three pass. If one will not pass in three tries, stop and tell me why.

Every clause is a check the model can run without you. That is the whole trick.

The sit-at-the-desk task

This is the thinking partner, and it is most of my week. I describe the goal, we argue about it, and the plan comes out of the conversation rather than going in as a spec.

Here I want the model that asks. Fable 5.1 asks more questions than anything I have used, and it reads what I meant when the prompt was incomplete, which is most of my prompts. Nate B Jones ran a clean version of this comparison: the same five-line brief for a Mac clipboard manager to both models. Fable asked which edge of the screen he wanted the panel on before it built anything. Astra picked one and built (his video). Both were right, for different tasks.

The catch is that Fable is rationed, on purpose. On Anthropic's price list Fable 5.1 is 10 dollars per million input tokens to Opus 5's 5, and on the Max plans Fable is capped at half your weekly limit while Opus 5 is not capped at all. So I spend Fable on the one problem where a wrong answer costs most, and when the cap hits, Opus 5 holds the same conversation. I am still running both Fable and Astra and will write again when I have more than three months of this.

The takeaway

With Fable and Astra, the effort dial is the wrong place to reach first. Sort the task, then pick.

Before your next prompt, ask: is this walk-away or sit-at-the-desk?

Walk-away: give the model a way to check itself and a rule for when to stop, hand it to the eager model, and leave the effort at medium.

Sit-at-the-desk: use the model that asks questions, and save the rationed one for the job where being wrong is expensive.

What is the task you would never walk away from?