skill
fable-sonnet-orchestrator
Claude Code skill: route subagent workers by task profile. Frontier reasoning for planning, judging and synthesis; commodity pricing for execution. Based on Anthropic's benchmarked Fable-orchestrator/
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/juniper234/fable-sonnet-orchestrator itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026fable-sonnet-orchestrator
Now bundled: this skill ships alongside its two siblings in the claude-orchestration plugin — one install, all three skills, same content. Prefer that for new installs. This repo stays up and works fine if you only want this one skill.
A Claude Code skill that tells Claude which model to route each subagent to when it fans out work: frontier reasoning for planning, judging and synthesis; commodity pricing for execution.
Without it, a Fable 5 session fanning out 12 files of audit work leaves every worker
inheriting Fable 5 (USD 10 in / 50 out per MTok) for pure execution — because the
Agent tool's own guidance says to omit model unless "highly confident" a different
tier fits, and nothing in context supplies that confidence. This skill is that
confidence: a routing table with the economics behind it.
The pattern
Anthropic benchmarked two ways to keep frontier quality without frontier cost:
| Pattern | Result | Benchmark |
|---|---|---|
| Fable 5 orchestrator + Sonnet 5 workers | 96% of all-Fable quality at 46% of cost (86.8% vs 90.8% accuracy, USD 18.53 vs 40.56 per problem) | BrowseComp |
| Advisor: cheap executor consulting a smarter model ~once per task | ~92% of quality at 63% of cost (84% vs 91.5% pass rate, USD 1.40 vs 2.25 per task) | SWE-bench Pro |
Sources: anthropics/claude-cookbooks → managed_agents/CMA_plan_big_execute_small.ipynb,
and Anthropic's advisor strategy post.
Figures are Anthropic-reported, not independently reproduced here.
Worker routing
| Task profile | model | Why |
|---|---|---|
| Decompose, dispatch, judge, synthesize | (orchestrator keeps these — never delegate) | Planning errors compound downstream; the 96% figure assumes Fable planning |
| Implementation, multi-file edits, tests, code review, verification | "sonnet" | Near-Opus execution quality, 1M context, USD 3 in / 15 out per MTok vs Fable's 10/50 |
| Mechanical scan, classify, extract, format | "haiku" | USD 1 in / 5 out per MTok; 200K context suffices for a narrow slice |
| Security/bio-adjacent worker tasks | "opus" | Fewer safeguard false-positives on benign security work; skips refusal-retry loops |
The savings come from workers. Never downgrade the orchestrator to save money, and never delegate synthesis.
Install
Clone into your skills directory:
git clone https://github.com/juniper-tc02e/fable-sonnet-orchestrator \
~/.claude/skills/fable-sonnet-orchestrator
Restart Claude Code and it loads as fable-sonnet-orchestrator@skills-dir — confirm
it appears in your available skills, or invoke /fable-sonnet-orchestrator directly.
This is a skill directory, not a plugin: SKILL.md sits at the repo root, so there
is no .claude-plugin/plugin.json and claude plugin validate does not apply to it.
When it does not apply
Single-thread deep reasoning stays all-Fable — there are no workers to route. Work one Grep covers shouldn't fan out at all. ZDR organisations can't use Fable 5 at all (it requires 30-day retention), so those run an all-Sonnet fleet.
Testing
TESTING.md has the red/green evidence: the baseline failure this was
written against, the with-skill result, and the measured delta. Eval cases live in
evals/ and are re-runnable:
claude plugin eval ~/.claude/skills/fable-sonnet-orchestrator
Limitations are stated there rather than hidden: one rep per arm, planning-exercise format, and benchmark figures cited as reported.
Pricing note
Prices are USD per million tokens and were verified against Anthropic's published model reference in July 2026. Sonnet 5 carries an introductory 2/10 rate through 2026-08-31. Prices change — check the current reference before relying on them.
Author
Ong Jun Kai — LinkedIn
Licence
Apache 2.0. See LICENSE.