A Claude Code skill that cuts the cost of your most expensive model by routing token-heavy work to lighter ones — while keeping apex-tier judgment where it counts. The apex seat binds to the most capable model you can call: Fable 5 (GA again 2026-07-01) in subscription Claude Code and for frontier-hard work; Opus 4.8 as the metered-API default.
Created by Jon Orozco
The apex tier is the most capable — and most expensive — model you can call. That's Fable 5 ($10/$50 per million tokens, redeployed 2026-07-01) with Opus 4.8 ($5/$25) as the metered-API default seat. Either way it costs far more than Sonnet 5 ($3/$15, intro $2/$10 through 26-0831) or Haiku ($1/$5).
Most real-world tasks aren't pure reasoning. They're a mix of grunt work (scanning repos, reading logs, running tests, making bounded edits) and judgment calls (architecture decisions, conflict resolution, risk assessment). Without orchestration, Fable spends premium tokens on work that cheaper models handle just as well.
JRO-Fable turns the apex model into an orchestrator — a tech lead who delegates the legwork and focuses on the decisions that actually need apex-tier intelligence.
| Work Type | Model | Why |
|---|---|---|
| File search, grep, log scanning, inventory | Haiku ($1/$5) | Pattern matching — no reasoning needed. 10x cheaper. |
| Bounded edits, test execution, code review | Sonnet ($3/$15) | Strong code generation at 3.3x cheaper. |
| Architecture, synthesis, conflict resolution, final review | Apex — Opus 4.8 (self) | This is what you're paying for. |
Each preset has an exact output contract so subagent results inject cleanly into Fable's context without wasting tokens on unstructured prose:
- Scout (Haiku) — file search, grep, codebase inventory
- Builder (Sonnet) — bounded code edits, refactoring, implementation
- Tester (Sonnet) — test execution, browser checks, log reduction
- Reviewer (Sonnet) — code review, security scans, doc accuracy checks
Real numbers from a typical 10-call research workflow:
| Setup | Cost | Savings |
|---|---|---|
| All apex (Opus 4.8) | ~$0.21 | — |
| Apex orchestrator + Sonnet workers | ~$0.10 | ~55% |
| Apex orchestrator + Haiku workers | ~$0.04 | ~80% |
Illustrative (10 calls, 2K in + 500 out each); lib/pricing.json is authoritative and the numbers scale with whatever the apex tier binds to.
Prompt caching (0.1x read price after first call) compounds these savings further.
This skill started as the open-source efficient-fable skill from Builder.io's skill registry. I then ran a multi-agent audit and rewrite process using Fable itself:
- Research phase — Four parallel agents gathered signal on Fable 5 pricing, orchestration best practices, cost optimization strategies, and gap analysis against my best-performing skills.
- Gap analysis — Identified 9 specific weaknesses: missing model routing table, no tool/API delegation syntax, no output contracts, no anti-patterns section, imprecise triggers, no pre-flight integration, missing cost data, no progressive disclosure, and no context-awareness for my workflow.
- Complete rewrite — Fable synthesized all research into a ground-up rewrite of SKILL.md (the core skill) and a new DELEGATION.md (progressive disclosure companion with named presets and workflow patterns).
- Validation — Every finding was verified against the actual Claude Code Agent tool behavior, current model pricing, and real delegation overhead measurements.
The result is a skill that practices what it preaches — built efficiently using the exact orchestration pattern it teaches.
git clone https://github.com/JRO424/jro-fable.git /tmp/jro-fable
cd /tmp/jro-fable && bash install.sh# Copy the skill into your Claude Code skills directory
cp -r jro-fable ~/.claude/skills/jro-fableThen add this to your ~/.claude/CLAUDE.md:
# jro-fable
- **jro-fable** (`~/.claude/skills/jro-fable/SKILL.md`) - Fable cost optimizer
When the user types `/jro-fable`, invoke the Skill tool with `skill: "jro-fable"` before doing anything else.In Claude Code, running on your apex model (Fable 5 / Opus 4.8):
/jro-fable
Or just start a token-heavy task — the skill auto-triggers on phrases like:
save tokens, delegate, fan out, orchestrate, too expensive, token budget
| File | Purpose |
|---|---|
SKILL.md |
Core skill — model routing, delegation rules, handoff packets, anti-patterns, cost reference |
DELEGATION.md |
Named presets with output contracts + workflow patterns for multi-stage orchestration |
install.sh |
One-command installer — copies files and wires up CLAUDE.md |
assets/ |
Excalidraw diagram source + rendered PNGs (light/dark) |
- Claude Code with access to a flagship model for the apex tier (Fable 5 / Opus 4.8)
- The
Agenttool (built into Claude Code) for spawning subagents - Optional: the
Workflowtool for multi-stage orchestration with 3+ agents
Improvements welcome! This skill is designed to evolve as Claude's model lineup and pricing change. Areas where contributions would be especially valuable:
- New delegation presets — if you find a recurring work pattern that benefits from a named preset and output contract, open a PR.
- Cost data updates — model pricing changes; keep the tables current.
- Workflow patterns — real-world
pipeline()andparallel()patterns from your own orchestration work. - Anti-pattern discoveries — if you find delegation scenarios that consistently lose money vs. doing the work directly, document them.
Fork the repo, make your changes, and open a pull request. Please keep the same structured style — concise, evidence-backed, no filler.
MIT — use it, modify it, share it. Attribution appreciated.