Claude Fable 5.1 is Anthropic's most capable generally available model, sitting above the Opus family. If you've been vibe coding with Opus or Sonnet, most of your habits still work — but a handful of things are genuinely different, and knowing them up front saves a confused first week.

It thinks before every answer, always

On earlier models, extended thinking was something you switched on. On Fable 5.1 it's always on and can't be turned off. You control how much it thinks with an effort setting — low, medium, high, xhigh, max — rather than whether it thinks at all.

The practical effect: it takes longer to start answering, and the answers are more considered. For quick back-and-forth, turn effort down. For "build this whole feature", leave it high and go and make a coffee.

Turns are longer

A single request on a hard task can run for many minutes. A fifteen-minute turn where it reads your code, plans, builds and checks its own work is normal, not a hang. If you're used to interrupting after thirty seconds of silence, you'll need to unlearn that.

This is the biggest change in feel. It's also where most of the extra capability lives.

It does more without being asked — in both directions

It's better at long autonomous runs and needs less hand-holding. It's also more likely to fix a nearby bug you didn't mention, add tests you didn't ask for, or rewrite a whole file when a small edit would do. Explicit boundaries about what to leave alone work well; see keeping changes surgical.

Prompts written for older models can make it worse

Step-by-step scaffolding that helped Sonnet 4 stay on track tends to constrain Fable 5.1 and lower the quality of what comes back. State the goal and the constraints; let it work out the steps. More in de-prescribing your prompts.

It costs more

Roughly double Opus 5 per token. That's worth it for hard problems and long autonomous runs, and often not worth it for a quick CSS fix. The cost guide covers when to reach for it.

Its refusals look different

Safety classifiers can decline a request outright rather than answering cautiously. It's rare in ordinary building, but there are specific phrasings that trigger false positives, and knowing them helps. See when Fable refuses.

The short version

fable-quickstart.txt
Working with Claude Fable 5.1, in one paragraph:

Say what you want and why, not how to do it. Set effort to match the
task — low for chat and small edits, high for real work. Let long turns
run; interrupting wastes the thinking. Tell it what to leave alone.
Commit before every agent task. Ask for a plain-English summary at the
end. And if it declines something ordinary, rephrase — it's a classifier,
not a judgement.

The single biggest mistake is treating it like a faster Sonnet. It's a different kind of collaborator: slower to start, far more capable of finishing.