Reasoning Effort Control

Treat a model's reasoning effort as a dial to match to the task, not a fixed maximum. Modern models expose graded effort/thinking levels (the source names low / medium / high / x-high), and higher effort means longer runs and higher cost — individual requests on hard tasks "can run for many minutes at higher effort settings" while the model gathers context, builds, and self-verifies. Effort control is the discipline of matching the level to the task's value: the source relays Anthropic's recommendation to use high as the default, x-high for the most capability-sensitive workloads, and medium/low for routine work. The corollary is model and effort selection as a cost lever — reaching for the most expensive model at max effort "for everything" is "almost 100% overkill"; the source suggests realistically needing the top model only ~5–15% of the time.

A related, adjacent habit: let the model act once it has enough rather than forcing exhaustive up-front planning. On a model that can run for minutes, endless option-surveying "just burns time and money on choices it will never use" — so "when you have enough information to act, act" is effort control applied to deliberation, not just to a setting. (This half corroborates Agentic Simplicity.)

Claims

Match reasoning effort to the task's value; higher effort trades latency and cost for capability, so spend it only where the task earns it. principle — durable: effort is a cost/quality tradeoff, and the same "pay for complexity only where it pays off" discipline that governs architecture governs the effort dial.
Default to high; use x-high for the most capability-sensitive workloads and medium/low for routine work. best practice — context: the source attributes this ladder to Anthropic's Fable 5 guidance; the specific level names and default are model/vendor-specific and will drift, so the ladder is the durable part, not the exact labels.
Using the most expensive model at high effort for everything is overkill — select model and effort per task (the source estimates reaching for the top model only ~5–15% of the time). best practice — context: cost-sensitive operation of a premium model billed by usage; the percentage is an illustrative estimate, not a measured constant.
"Act when you have enough" beats exhaustive up-front planning on a model that can run for minutes — don't over-plan or narrate options you won't pursue. best practice — context: agentic/long-running tasks; over-deliberation burns budget on unused choices. Corroborates Agentic Simplicity; the operator's own "plan mode until ready to act" replaces a forced full-plan step. See Agent Loop.
The source's FrontierCode accuracy-vs-cost chart claims Fable 5 at low effort is "similar to Opus 4.8 at x-high, but cheaper." observation — a benchmark comparison reproduced as presented; the numbers are groundable and are flagged for a later verification pass, not adjudicated here.
A second, independent source (Chase AI) corroborates the effort ladder with fresh numbers: on the DeepSWE agentic benchmark it shows Fable 5 at low scoring 60% at $3.76/task — beating Opus 4.8 at max (59% at $13) — while Fable at max costs $22 for ~70%, i.e. dropping to low is an ~80% cost cut for a ~10-point pass-rate loss (60→65→69→70% across low/med/high/x-high). observation — a distinct benchmark (DeepSWE) converging on the same "low effort is the value pick, max barely beats x-high" conclusion; figures reproduced as the source's chart presents them, flagged for grounding, not adjudicated. That source also shows the effort selector exposing an "ultracode" tier above max (labeled "x-high + workflows").

Agentic Simplicity — "act when you have enough" and "don't over-plan" are the simplicity thesis at the deliberation level; this concept is already cross-linked from there.
Concise Prompting — effort control tunes how hard the model thinks; concise prompting tunes how much you say. Both are "lighter touch" for a capable model.
Agent Loop — the operator's own iterate-until-ready loop is where "act when you have enough" lives, replacing a mandatory plan-mode step.
Safety Routing Fallback — a cheaper backup model answering a routed request is another axis of the same cost/capability tradeoff.
Model-Tier Routing — the same tradeoff generalized from one model's effort dial to which model tier runs each role in a task.
Distillate: How Anthropic Engineers Actually Prompt Fable 5 — six habits for a smarter, lighter touch
Distillate: Make Fable 5 80% Cheaper (& Other Usage Cheat Codes) — five levers for spending a premium model less — independent DeepSWE + FrontierCode numbers corroborating "reduce effort to cut cost; low is the value pick."

Reasoning Effort Control

Claims

Related

Linked from