Guides published jul 10, 2026
In this guide, you will learn two ways to stop burning Fable tokens on grunt work in Claude. Set Fable up as the planner and reviewer, then send browsing, coding, and research to Codex or a lower-cost Claude model so premium tokens are saved for the calls that need stronger judgment.
You will build a simple model-routing setup: Fable owns the plan, risky calls, and final review, while a cheaper worker handles the token-heavy loop. The result is a self-checking AI worker that knows when to escalate instead of pulling Fable into every step.
Save the delegation instruction as a Claude Code project instruction or reusable snippet. Then make a few worker templates for the work you repeat: a repo change, a research brief, a QA pass, or a content draft. Keep the same routing rule in each one: Fable plans, judges risk, and reviews the handoff; the worker does the bulk execution.
For a quick, low-risk task, skip the orchestrator setup and use the cheaper model directly. Bring Fable in when the work needs stronger judgment, not simply because it needs more tokens.
Open the Claude desktop app and make sure it is current before you start: go to Settings > Check for updates and install anything available. The current app has Home and Claude Code tabs. Then click your name and go to Settings > Usage.
This screen shows your weekly limit and your Fable limit. Check it before you let Fable run a large task, especially one with browsing, coding, file reading, or repeated retries. In our test, Fable consumed the remaining allowance quickly during a short session.
Pro tip: Keep Fable out of routine retries. Use it for the plan, risky calls, and final review, then let the worker handle the token-heavy loop.
If you have a ChatGPT plan, you can use Codex as Claude's workhorse. This is the official OpenAI Developers (Codex) plugin for Claude Code—not a third-party integration. Open a terminal in a new project folder, start a Claude Code session with claude, then follow the official OpenAI plugin setup instructions (https://developers.openai.com/learn/developers-codex-plugin) to add its marketplace and install the plugin:
/plugin marketplace add openai/openai-developers-for-claude
/plugin install openai-developers@openai-developersStart a new Claude Code session after the install. If you had the Claude desktop app open, restart it too. Then run the Codex setup check from Claude Code to confirm the integration is available. Claude can now send companion tasks to Codex instead of spending Fable tokens on routine coding, browsing, and website extraction.
Fable should be the orchestrator, not the person doing every mechanical step. Ask it to inspect the available Codex skills and route the work that needs execution to the right one. A short instruction is enough to set the behavior:
Use your Codex skills to save your own tokens.
Keep planning, complex decisions, and final review with Fable.
Delegate coding, browsing, file extraction, routine checks, and retries to Codex.
Escalate back to Fable only when the worker hits ambiguity, a high-risk decision, or a stuck loop.You can see the resulting work in Codex as Codex companion task threads. For coding work, add the Codex-first routing skill to Claude through Customize Skills > Add > Upload Skill. It tells Fable which coding tasks belong with Codex and which are worth keeping in the expensive judgment lane.
Ask the worker to return evidence, changed files, checks performed, and open questions. That gives Fable a small review package instead of a long stream of work-in-progress commentary.
You can use the same pattern without a separate Codex subscription. /advisor is a built-in Claude Code skill, but it works when you launch Claude Code in your terminal—not from the Claude desktop app interface shown earlier. Run claude in a terminal, use /model to select a lower-cost worker model, then use /advisor to set Fable as the advisor. Anthropic's advisor tool documentation (https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool) explains the same executor-plus-advisor pattern.
The worker continues with the task until it reaches a complex decision, an ambiguous failure, or a loop that is not making progress. At that point, it asks Fable for guidance and resumes the work. The advisor runs server-side and uses additional tokens, so use it for decision points rather than every turn.
The recorded setup used a lower-cost model as the main worker and Fable as the advisor. That is the useful split: the worker handles the volume, while Fable supplies the stronger judgment only when it can change the outcome.
Pro tip: Advisor mode is a lighter-weight routing option, not a guarantee of a fixed savings percentage. How much it saves depends on the task and how often the worker escalates.
Keep going
Members get all guides, courses, live sessions, and $1,000+ in partner perks.