Tune Effort Level Per Task, Not Per Model

Modern frontier models expose a tunable thinking/effort budget. The spread is enormous:

• Claude Fable 5.1 has five effort levels spanning 11x in output token usage (13.1M at low to 143.7M at max per Artificial Analysis). At the lowest effort it already scores ~26% on Terminal-Bench-Science at ~$11/task — beating Fable 5 at maximum effort (24.7%) which cost ~$44/task. The cheapest setting of the new model beats the most expensive setting of the old one at a quarter of the price.
• GPT-5.6 Sol exposes configurable reasoning effort (low → max / xhigh).
• Gemini 3.8 Flash exposes tunable thinking levels (low, medium, high).

Choosing the right effort level is now a real lever for anyone paying by the token — often bigger than the choice of model itself.

1. Default to a low/medium effort for exploration and scaffolding. Most turns in an agentic session are not the hard reasoning turn — they are tool calls, file reads, and small edits. Running those at max effort burns output tokens for no quality gain.
2. Reserve max effort for the long-horizon, high-stakes turn. The terminal-bench-class tasks where effort actually moves the score are a minority of turns. Identify them and spike effort there.
3. Measure cost-per-task, not cost-per-token. Fable 5.1 (max) costs $3.76 per Intelligence Index task; at xhigh it scores only 1 point lower (65 vs 66) for $2.72 — $1.04 less. A model that finishes in fewer tokens at lower effort can be cheaper per result while costing more per token.
4. Let the agent or harness pick. Some harnesses (Claude Code defaults Fable 5.1 to High; Claude Cowork and Claude.ai default to Medium). Know your harness's default and override it deliberately, not accidentally.

Effort	Fable 5.1 Intelligence Index	Cost per task
Low	58	lowest
xhigh	65	$2.72
Max	66	$3.76

The jump from low to xhigh buys 7 index points. The jump from xhigh to max buys 1 point for ~$1.04 more. For most production workloads, xhigh is the Pareto-efficient choice — not max.

• Don't set max effort globally. It inflates cost on the 80% of turns that don't need it, and the 75% cache-read price cut (where available) doesn't help with output token bloat from high effort.
• Don't assume lower effort always degrades quality. On short-horizon, well-scoped tasks (single-file IDE edits), the gap between effort levels is narrow — sometimes within noise.
• Do compare effort levels on your own workload, not just published benchmarks. The effort-vs-cost curve is task-dependent; a benchmark average can mislead on your specific mix.
• Remember effort ≠ model tier. A lower-effort setting on a frontier model can beat a higher-effort setting on a cheaper model on the same task. Compare at matched effort, not at default.

The single biggest cost lever available without changing models is the effort dial. The spread across effort levels on one model (11x token usage, 8 index points, ~$1.04/task swing) is often wider than the spread across models at fixed effort. Tune it deliberately, per task, and measure cost-per-result.