The dial

Fable 5 effort levels, explained — low, medium, high, xhigh

One parameter decides how hard Fable 5 thinks, how long you wait, and how much you pay. Anthropic calls effort the primary dial on this model — here's what each setting buys, and the two ways people get it wrong.

Why effort matters more on Fable than anywhere else

On Fable 5, effort is the primary dial trading intelligence against latency and cost. Other prompting choices nudge the model; effort decides what kind of worker shows up. Anthropic's prompting playbook treats it as the first decision you make for any task — before the prompt, before the tools.

The four settings

low

Quick questions, rewrites, interactive back-and-forth where every extra second of thinking is a second you're waiting. Don't read "low" as weak: low effort on Fable often beats maximum effort on prior models.

medium

Routine work with a little substance — ordinary edits, short analyses, everyday coding turns. The same caveat applies: still strong, and fast enough to stay conversational.

high

The recommended default. Real engineering, real writing, real analysis — the setting Anthropic suggests you start from and adjust away from only with a reason.

xhigh

Capability-critical work where quality is non-negotiable: the hardest debugging, tasks where one wrong answer costs more than the tokens, long autonomous runs nobody will be supervising. This is where Fable's best verification and reasoning behavior lives.

The two failure modes

Too high on routine work

Fable will happily over-deliberate a simple job — longer waits, wasted tokens, no better answer. The tell: a task completes correctly but takes too long. That's your signal to reduce effort, not to rewrite the prompt.

Too low on the work that matters

Higher effort also brings the model's best verification and reasoning behavior — the double-checking that catches its own mistakes. Starving your hardest jobs to save latency buys you a fast wrong answer. If quality is the point, pay for the thinking.

The effort picker

Task typeSuggested effort
Quick questions, chat, interactive back-and-forthlow
Routine edits, rewrites, everyday coding turnslow / medium
Serious engineering, analysis, substantial writinghigh (the default)
Whole-codebase review, hardest debuggingxhigh
Long autonomous runs, unsupervised buildsxhigh
Anything where one wrong answer costs more than the tokensxhigh

Two practical notes

  • Long runs are normal. Hard tasks at higher effort run for many minutes in a single request. That's the design, not a hang — adjust your client timeouts and go do something else; don't stare at spinners.
  • Steer with one sentence, not a rulebook. On Fable, one brief instruction changes behavior more reliably than the long rule lists you wrote for older models — over-specified prompts can actually degrade output. The full playbook translation, with Anthropic's copy-ready instructions, is Recipe 07.

Effort answers how hard Fable should work; it doesn't answer whether the job belongs on Fable at all. For that triage — what earns Fable's rate and what a cheaper model does just as well — see what to run on Fable 5.

Moral: match the effort to the stakes, not to the model's reputation — Fable at low effort is still Fable, and Fable at xhigh is only worth it when being wrong is expensive.

Keep reading

Live