Stop Wasting Your Fable 5 Tokens
You will have a CLAUDE.md that keeps Fable 5 on reasoning, UI and architecture, sends token-hungry work to cheaper models, and lets Claude Code hand back end coding to Codex through the Codex plugin.
Was Sie brauchen
Claude Code (or Cowork) with access to Fable 5
A CLAUDE.md file in your project
Optional: a Codex subscription (I find the $100 plan is enough for many use cases)
Optional: a Gemini API key if you want Gemini as another model
Schritt für Schritt
Know why your Fable 5 tokens run out
Your Fable 5 allowance is now half of what it was: only 50% of your weekly plan can go to Fable 5. It also consumes about twice as many tokens as Opus 4.8. Plan to spend Fable 5 only on the work that needs it, or you will hit your rate limits.
Expect Fable 5 to offload tasks on its own
The model now has stronger guardrails and tends to hand tasks to lower models such as Opus 4.8 or Sonnet 5. Fable 5 is strongest at taste and architecture, so you need to tell it which tasks it must do itself and which ones it should pass on.
Write your routing rules in CLAUDE.md
Open your CLAUDE.md. List the tasks you run (for example revising a plan or creating a UI) and name every model Fable 5 can use and when to pick each one. Example: Fable 5 for plans and UI, Opus 4.8 for coding tasks.
Make Fable 5 the orchestrator at high effort or lower
Select Fable 5 as your main model and set its effort to high at most. Effort controls how long the model thinks, and less thinking keeps your Fable 5 tokens for the parts that matter.
Send token-hungry work to other models
In the same CLAUDE.md, give Fable 5 access to Sonnet, Opus and even Gemini (if you add an API key). Route heavy tasks such as computer use and reading images to them. Expect 50 to 60% of your work to be orchestrated by Fable 5 rather than done by it.
Connect Codex to Claude Code
Install the Codex plugin for Claude Code from the link below so Fable 5 can orchestrate Codex (GPT 5.5) directly. It uses your settings from ~/.codex/config.toml. Then turn on the review gate so Codex checks Claude's output before it is finalized.
Codex plugin for Claude Codehttps://github.com/openai/codex-plugin-cc
Turn on the Codex review gate/codex:setup --enable-review-gate
Split UI and back end between Fable 5 and Codex
Keep UI and UX with Fable 5 and hand back end coding to Codex. Use the plugin commands: review for read-only code checks, adversarial-review to pressure-test a design, rescue to hand off debugging or refactoring.
Read-only code review/codex:review
Skeptical design review/codex:adversarial-review
Hand off debugging or refactoring/codex:rescue
Paste the ready-made CLAUDE.md and adapt it
Paste the CLAUDE.md from the video (download it below) into your file. Then replace the sections that don't fit how you work, such as the cost scores or which model handles which kind of task.
Worauf Sie achten sollten
Fable 5 gets only half your weekly plan and burns about twice the tokens of Opus 4.8, so leaving it on every task will hit your rate limits fast.
Without routing rules, Fable 5 may pass tasks you want it to do (UI, architecture) to a lower model.
Setting effort above high makes Fable 5 spend precious tokens on thinking you don't need.
The model scores in the CLAUDE.md are defaults, not limits: if a cheaper model's output isn't good enough, let the smarter model redo it.
