fable-baton

fable-baton 🪄

Fable 5 holds the baton. The orchestra plays.

A Claude Code plugin that makes Fable 5 the orchestrator. Fable keeps the judgment, tiered subagents on Opus, Sonnet and Haiku do the labor. Install once and every new session in every repo starts this way.

Why · Benchmark · How it works · What you'll see · Install · Day-to-day · Alternatives

Why

• Fable time is precious. It is the strongest model on a Claude subscription and the most expensive one to burn on grep runs and boilerplate. fable-baton routes each piece of work to the cheapest tier that can do it well, so you can keep Fable 5 as your daily model.
• No more switcheroo. When Fable time runs dry, Opus quietly takes over your session. With fable-baton, Fable spends its tokens on judgment only, Opus does the heavy work below it as a subagent, and Fable stays the one holding the context.

The tiers are Opus, Sonnet and Haiku today. Later this should open up to other models and structures.

Benchmark

Numbers from a small controlled test, run 2026-07-13 on Claude Code 2.1.197 with plugin v1.3.0. The project was a zero dependency Node.js library with a test suite. Three tasks: fix two seeded bugs, implement a feature from a spec in TODO.md, and review the codebase for bugs. Each task ran headless via claude -p, twice with the plugin off and twice with it on, same session model (Fable 5) in both. Success was checked from outside the session: the test suite had to pass (for the feature task against hidden tests the session never saw) and the review had to find both seeded bugs. All 12 runs passed.

task	plugin off, total cost	plugin on, total cost	plugin on, cost split	Fable output tokens, off / on
bugfix	$0.96	$1.15	Fable $0.92, Sonnet $0.19, Haiku $0.04	2.4k / 2.2k
feature	$1.52	$1.50	Fable $1.05, Sonnet $0.39, Haiku $0.06	6.1k / 3.4k
review	$2.13	$2.71	Fable $1.61, Opus $1.05, Haiku $0.05	13.7k / 8.5k

Values are per run averages over the two reps. Read the table honestly. Total API cost comes out about the same and sometimes higher, because orchestration adds coordination on top of the work. What changes is where the tokens land. Fable's own output tokens drop 44 percent on the feature task and 38 percent on the review, and its share of the cost drops with them. That work moves to Sonnet, Haiku and Opus. If you pay per token through the API and only care about the total, the plugin will not save you money on small tasks. If you are on a subscription where Fable quota is what runs out and triggers the mid session model switch, this is the tradeoff you want: Fable stays available for judgment much longer and the session keeps its conductor.

Keep in mind this is n=2 per cell on one small project, and variance between reps was real (one review run cost $3.25, the other $2.16). Treat the numbers as directional. The benchmark harness is not shipped with the plugin.

How it works

Four tiered agents

Agent	Model	Owns
scout	Haiku	Discovery, reading files/logs, summaries, simple checks
executor	Sonnet	Scoped implementation, tests, routine edits, local refactors
architect	Opus	Complex implementation, deep debugging, high-risk work, reviewing cheaper agents
verifier	Haiku	Evidence checks: tests green, diff matches plan, no regressions

One policy

A SessionStart hook injects the orchestration policy into every new session. It tells Fable what to keep (intent, architecture, tradeoffs, review) and what to route down (labor), with anti-waste rules: no pointless fan-out, focused context per agent, no delegation for genuinely trivial single steps.

Three layers of enforcement

A one-time policy is not enough. Models drift back to doing everything inline as a session goes on. We watched it happen in real sessions.

Layer	Hook	What it does
Policy	SessionStart	Loads the full orchestration policy when the session starts
Reminder	UserPromptSubmit	Re-asserts the delegation rules on every prompt
Counter	PostToolUse	Counts consecutive inline tool calls and injects a delegation notice once a streak crosses the threshold (default 4, set with FABLE_BATON_TRIPWIRE, resets whenever an agent is used)

The model can still ignore a notice. But ignoring a fresh instruction mid-streak is much harder than forgetting something from page one.

Special handling