# Effort Is a Cost-Quality Dial That Subsumes Prompt Complexity On Claude Fable 5, Anthropic names **effort** (low / medium / high / xhigh) as "the primary control for the trade-off between intelligence, latency, and cost." This is a quiet but real shift: the knob that used to be implicit in *how elaborately you wrote the prompt* is now an explicit, first-class parameter. You no longer coax more reasoning out of the model by inflating the prompt; you set the dial. The guidance even says lower effort settings "often exceed `xhigh` performance on prior models" — capability is decoupled from prompt elaborateness and attached to a setting. The operating rule is to default `high`, reserve `xhigh` for genuinely capability-sensitive work, and drop to `medium`/`low` for routine tasks — explicitly to buy speed and lower cost when the task doesn't need depth. That makes effort a **FinOps dial**, not just a quality dial: every notch is a cost decision applied uniformly across requests, far easier to govern than per-prompt verbosity. <!-- atomicity: 1 idea ✓ --> ## Why This Matters For **tokenomics**, effort is org-level cost control in a way prompt discipline can't be. [[3. Resources/Permanent Notes/Vague Prompts Are Expensive Specificity Cuts AI Error and Cost Together]] is the *request-level* lever — write it well to spend reasoning tokens efficiently. Effort is the *dial-level* lever — set depth-to-cost once, route routine work cheap. The two are complementary altitudes of the same bill, which is exactly the multi-altitude FinOps framing in [[AI Cost & Tokenomics MOC]]. For **prompt engineering**, it deflates a habit: padding prompts to force more thinking. With an explicit effort control, prompt verbosity for its own sake is pure cost with no upside — clarity still matters, elaborateness doesn't. ## Counterpoints to Hold - **One global dial is coarse.** A single effort setting per call can't express "think hard about step 3, coast on the rest"; real tasks have uneven depth needs the dial flattens. - **Default-high is a soft upsell.** Recommending `high` as the default means most traffic runs at a higher cost tier unless teams actively tune down — a default that benefits the vendor's revenue. - **Vendor-specific.** Effort is a Fable-5 control; the abstraction may not port across providers, so building process around it is a small lock-in. ## Links - [[3. Resources/Permanent Notes/Vague Prompts Are Expensive Specificity Cuts AI Error and Cost Together]] — the request-level cost lever; effort is the dial-level lever - [[3. Resources/Permanent Notes/Long-Horizon AI Shifts the Bottleneck From Model Capability to Memory and Effort Scaffolding]] — effort is one of the two long-horizon levers (with memory) - [[3. Resources/Permanent Notes/Rewarding AI Usage Manufactures Usage Not Value]] — a default-high dial inflates spend the way usage incentives inflate adoption - [[3. Resources/Literature Notes/Lit · Fable 5 prompting — effort and memory]] — the Literature source tying the explicit effort dial to external memory for long-horizon work - [[3. Resources/Permanent Notes/A Declared Capability Costs You Every Turn Not Every Use]] — contrasts on altitude (added 2026-08-01): effort is a dial you set per request, while declared tool schemas are a fixed overhead charged before any request is made. Both are cost levers above the model, and conflating them hides the one you can't tune away by prompting better. ## Open Questions - Can effort be set per-step inside an agent loop, not just per-call? - Where's the crossover — at what task complexity does `medium` on a new model beat `xhigh` on the prior one, and how do you find it without burning the savings testing? ## Source - [[3. Resources/Readwise/Articles/Prompting Claude Fable 5|Prompting Claude Fable 5]] — effort as the primary intelligence/latency/cost control; default high, xhigh for hard work, medium/low for routine; lower effort can exceed prior-model xhigh - See also: [[AI Cost & Tokenomics MOC]]