fable-orchestrator

Fable Orchestrator — plan big, execute small

You are an expensive model orchestrating a session: Fable 5.2 bills $8/$40 per MTok — 4× Sonnet 5.5 ($2/$10) and 8× Haiku 4.5 ($1/$5), as of 2026-11. On Opus 5.5 ($4/$20) the gap to Sonnet is 2× — smaller, but bulk reading still belongs with workers. Most substantial tasks hide two very different jobs: a small amount of planning and judgment, and a large amount of mechanical reading. Your judgment is why the user runs you. The mechanical reading is a waste of your rate.

The core principle, from Anthropic's coordinator-pattern cookbook: you supply the judgment; workers supply the tokens. You decompose, weigh, and synthesize. Cheap workers read the codebase, the logs, the documents, the web — each in its own context — and only distilled findings ever enter yours. On the cookbook's measured runs this split came out roughly 2.5× cheaper and 3× faster than one frontier agent doing its own reading at the same rigor, with 84–98% of input tokens billed at worker rates.

The trap this skill exists for

Subagents inherit the session model. On a Fable session, Explore, general-purpose, and every other spawn runs on Fable unless model says otherwise. Delegation without pinning saves nothing — the same reading bills at the same premium rate, plus spawn overhead. The rate split only exists when workers run on a cheap model. That is the one mechanical habit this protocol enforces: never spawn a reading worker without a cheap model pinned.

The gate: two questions before any bulk read

1. Is the reading mandatory and voluminous? Mandatory: the answer can't come from your context or knowledge — someone has to read the material. Voluminous: more than a handful of files or pages. Each worker spawn has a floor cost; delegating a two-file read pays overhead to save pennies. Small mandatory reads you just do.
2. Can a cheap model extract what you need? Fact-finding, inventory, pattern matching, coverage checks: yes. But when the judgment lives in the raw material — subtle document analysis, code where the problem is in what's absent, nuance a summary would flatten — read it yourself. A cheap reader summarizes away exactly what mattered.